Welcome to The Syllabus, an AI in education newsletter experiment from me, Angela Chen, bringing together market signals, research, and policy. The mood this week: overhelpful.

This week, some things schools rely on by default came under review: AI tutors that give too much help, AI answers in Google Search that schools can’t turn off, and homework grades one Berkeley professor has stopped trusting. Each one comes down to what you can still decide, and how you’ll check whether students learned.

⚡ What to pay attention to right now

The same AI model tutors differently depending on its instructions

AI tutors often give too much help, but clearer instructions reduce it, and some models do better than others. The Allen Institute for AI tested this on 462 math tutoring transcripts from US grades 2–7. Twenty-seven math teachers marked more than 1,500 moments when a tutor had to choose between helping and pushing for rigor. With basic instructions, the models avoided over-scaffolding only 17% to 60% of the time, with GPT 5.5 doing worst and Gemini 3.5 Flash doing best. With instructions that spelled out the trade-off, the range rose to 60% to 91%. This August preview used simulated students, so it doesn’t show that any child learned.

For teachers, try an AI tutor yourself first. Give it five wrong answers from your class and count how often it gives the next step before a student has tried. If it passes, give students a related problem a week later without the tool. For school leaders, ask each vendor which model it uses and what instructions it follows. For founders, publish benchmark results with a delayed learning measure.

Google gave teachers a view into students’ AI chats, but schools still can’t turn off AI in Search

This is continuing to play out: Google’s latest update helps teachers see how students use AI, but it creates a new training need and leaves schools without a Search off switch. On October 19, Google began rolling out insight summaries for Gemini Notebooks assigned in Classroom, with an AI overview of common questions and each student’s chat transcript. Students are told their teacher can see their chats. Two days later, Senators Blumenthal and Hawley sent Google nine questions, due November 2, and said AI Overviews and AI Mode “cannot be disabled.” Google says teachers and administrators stay in control.

Teacher access is a good step, but only if teachers know how to read what they see. A transcript shows what a student asked, not what they can do. A student who asks many questions may be working hard or leaning on the tool, and a short transcript doesn’t show that a student understands. Check the AI summary against five transcripts, then give a short exit ticket without the notebook.

For school leaders, train teachers before they rely on transcripts. Use shared examples of what a transcript can and can’t show, and decide whether it can count as evidence in grading or discipline. Tell families, and write down who can read transcripts and what happens when one raises a concern. On Search, assume students see AI answers on every Chromebook and plan assignments accordingly. For founders, the useful work is helping teachers find misconceptions in those chats.

A Fields medalist stopped grading homework for correctness because of AI

Richard Borcherds, a UC Berkeley math professor, now gives homework credit for turning it in on time, whether or not it’s right. He told Newsweek that AI can solve many undergraduate problems, so a correct answer no longer shows understanding, while in-person exams still work. This is one professor in college math, but the question applies to any take-home work.

Completion credit alone won’t make students do the thinking, because they only benefit from the generation effect if they make the attempt. In-person exams can also land hardest on students with learning differences and access barriers, so we need to pair them with accommodations and chances to explain the work.

If AI will be changing assessments at your school, pilot changes on one unit before you change your grading. Use completion credit, give a short no-tool quiz a week later, and track whether students keep turning work in. For school leaders, don’t leave this change to individual teachers. Give departments time to plan it together, and back teachers when families ask why homework isn’t graded for accuracy. For founders, grading take-home work for correctness may be a shrinking use case.

Microsoft’s AI contract standard gives districts terms to ask every AI vendor to match

This is continuing to play out: districts now have ready-made AI contract terms they can take to all of their vendors. Since November 1, any US district can add the AI Safety and Privacy Standard that Microsoft negotiated with the AFT and UFT to new or existing agreements without renegotiating. It bars training on student and educator data, bans AI companions, requires human review of high-risk decisions and 72-hour breach reporting, and pauses or ends AI use if violations aren’t fixed. It binds only Microsoft, and the AFT has said it is still negotiating with OpenAI and Anthropic. Education Week also found limits: de-identified usage data can still be used to improve products, and biometrics and tracking get added review rather than a ban.

For district leaders, start from Microsoft’s work instead of a blank page. The full standard is mostly written for AI providers in general, and its terms can be folded into an existing data agreement. Ask your counsel about adding it to your Microsoft contract, and get Microsoft to say in writing which products it covers. Then send the same ten protections to your other AI vendors at renewal, and ask which ones they’ll sign.

Also, consider adding two things the standard leaves out. It doesn’t require vendors to name the underlying model, and it doesn’t ask for evidence that students learn. Ask for both in the same review. For founders, compare your contracts with the standard now, because buyers may start using it as their list.

🔮 The call

Every week I make one prediction and I’m curious about your take. Reply or comment below. Here’s this week’s:

By June 30, 2027, Google will give Workspace for Education administrators a setting that turns off AI Overviews and AI Mode in Search for student accounts while leaving standard search results on. Two senators are now asking why schools can’t turn these features off, and the workarounds on offer are third-party browser tools. Google could instead add more links and safeguards, since it treats Search as a core product. If the call lands, schools will have to decide when students should see AI answers, so it’s worth deciding now.

🛠️ Tools of note

(These aren’t startup investment recommendations – they show up here because they’re connected to trends worth tracking.)

AI for Education’s environment guide – An updated classroom guide on generative AI’s energy and water use, with questions to ask before using it. Students are raising this issue, and the guide helps them practice deciding when AI is worth using.

aiEDU’s vibe coding workshop – A free virtual session on November 12 where K12 educators build a working prototype using plain language. More educators are building their own tools. Anything that touches student data still needs a district privacy and security review.

Educating All Learners Alliance’s Tech Tool Library – A searchable directory of accessibility information for edtech products. AI features are arriving inside tools schools already use. Listing isn’t an endorsement, but it’s a good place to start your questions about students with learning differences and access barriers.

💰 Funding & market moves

Amira Learning – The AI reading company filed a Form D on October 9 for up to $14 million, with about $5 million sold so far. For literacy leaders, ask what students can do later on a measure outside the product. For founders, read a filing like this for the amount sold, not the target.

Matiks – The math and puzzle game app raised $10 million led by General Catalyst, with more than 300,000 daily users, mostly in India. For teachers, streaks and points can get students practicing without showing they’ve learned, which is worth remembering if students bring this app to class. For founders, this money is paying for engagement, and the learning question is still open.

Pearson – Since we first covered its Workera deal, Pearson reported that assessment revenue grew 10% last quarter, citing demand for verifying AI skills. US student assessment fell 4% over nine months. For CTE leaders and counselors, ask what each AI certificate requires a student to do unaided. For founders, the growth here is in certification, not state testing.

📚 Research & policy on my radar

The largest study of teacher coaching found no gain in test scores or retention – A new study of a large urban district, reported by Education Week, followed nearly 1,000 mentor teachers and more than 6,000 teachers. Compared with similar schools, it found no improvement in teacher retention or state math and reading scores, and as the program grew, teachers went from five or six observations a year to one or two. This study looked at human coaches, not AI. If a vendor offers AI coaching as a way to add capacity, ask for student learning results as well as how many teachers used the tool.

New York City is keeping its high school AI pilot despite calls to extend the moratorium – Advocates asked the city to extend the ban through high school, but the pilot is going ahead this month for about 5% of students, with notice to families and alternative assignments on request. Feedback alone can’t tell you what students learned, and that’s how the DOE plans to assess this pilot. Any district running a pilot should add a comparison group and a learning measure.

Oxford and England’s Department for Education are studying pupil data as a public asset – A six-month project will propose rules for using anonymized student data in AI research. US districts can ask their vendors the same question now: who decides how data from your students’ work gets used, and who benefits?

Big districts are pushing back on using E-Rate to regulate screen time – Education Week Market Brief reports broad support for E-Rate in FCC comments, with major districts warning against using it to regulate classroom technology. Whatever the FCC decides, be ready to explain to families why you use a screen in a given lesson and what students learned from it.

If one of these is showing up in your school or your product, reply and tell me what you’re seeing – I’m always interested in what’s actually landing for people in the field. I’d especially love to hear from anyone who has dropped correctness grades on homework.

That’s The Syllabus for this week. Lots to dig into!

I’m an edtech founder, Stanford instructor, and AI in education strategist. Every week, I read across research, classroom practice, startup and funding signals, policy, and market intelligence, and share the signals I think matter most.

If this was useful: 📬 Subscribe on Substack – The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take. 🔗 Connect with me on LinkedIn – I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space: I advise school districts building AI frameworks and edtech founders navigating the K12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this: Forward it. Re-stack it. Share it on LinkedIn. The best readers I’ll ever have are probably in your network!

Hi, I’m Angela Chen! I’m an edtech founder, Stanford instructor, and AI in education strategist. Every week, I read across the full AI-in-education landscape: research, classroom practice, startup and funding signals, policy, and market intelligence. In The Syllabus, I’m sharing the signals I actually think matter, synthesized across everything I read that week!

If there’s something here worth going deeper on, comment below or message me on LinkedIn and tell me which signal – I’m always interested in what’s actually landing for people in the field.

This week, several safeguards schools have been counting on turned out to be thinner than they look — the parent alerts in ChatGPT, the AI detectors still running in some classrooms, and the vendor relationships that hold student data. None of that argues for pulling AI out of schools, but it does mean the protection we rely on should sit where schools control it, in how we design the work and what our contracts say.

⚡ What to pay attention to right now

ChatGPT’s parent alerts fired only in 5% of high-risk teen conversations

A RAND-led team with researchers from Harvard Medical School, Brown and MIT tested ChatGPT’s parental controls by running 300 scripted high-risk conversations on 15 teen accounts linked to parent accounts. A parent got a notification within 30 minutes in 15 of them, and 13 of those 15 came from a single scenario, a teen asking how to tell their parents about suicidal thoughts. Conversations about weapon access, harm to others and sexual abuse triggered no alerts at all.

The same week, at least eight states sued OpenAI under consumer protection law, alleging that ChatGPT harms children and that the company overstated its safety testing. OpenAI says it has built child protections in and will keep improving them.

For school leaders, parental controls are a family setting, not a school safeguard, and this study suggests even families shouldn’t treat them as a safety net. If your district lets students use a general-purpose chatbot, or recommends one for homework, determine now and write down what happens when a student raises self-harm or a threat to others, and who sees it. Ask your vendor for the escalation path in writing, and make sure your counselors, not the product, own the response. For founders selling to K12, “we have parental controls” won’t carry a safety review much longer – expect buyers to ask how often your alerts fire, and who checked.

AI detectors no longer settle authorship, so teachers are moving the evidence into the room

As Edutopia reported on October 16, a Notre Dame study then showed that students could run AI-written papers through “humanizer” tools, revising by trial and error until they scored as fully human. Detection-only tools now share the weakness that undercut version-history trails, because whatever the software reads as evidence, another tool can learn to fake. The teachers Edutopia interviewed have shifted to collecting low-stakes baseline writing early in the year, handwritten first drafts, peer review, oral explanations and prompts tied to students’ own experience. I’ve seen this happening across the schools I advise too.

The thinking behind these moves are sound because they keep the critical thinking, judgment, and generation with the student. However, timed in-class writing measures something different from a revised essay, and it can disadvantage students with learning differences and access barriers, so treat it as one data point alongside the draft history and a conversation, not a replacement verdict.

Changing assessments and moving work into class also affects teaching practices and unmoors lesson plans, which may unfairly penalize teachers. Every essay written in the room uses instructional minutes that used to go to discussion, mini-lessons or reading, so a unit planned around take-home drafts no longer fits its calendar or the pacing guide. Oral checks and writing conferences take time with each student, which adds up fast for a secondary teacher with several sections. Teachers are also rebuilding rubrics, teaching new routines and explaining the change to families, mostly on their own time and without new materials. If observations still measure them against the old pacing, the teachers who redesign their assessments most carefully can look like they’re falling behind.

For school leaders, that makes assessment redesign a school-wide decision that teachers shouldn’t absorb alone. Your test for any integrity approach is whether it tells you what a student can do: can they explain a choice they made in the essay, and can they write a stronger paragraph on a related prompt two weeks later without help? Give departments planning time to build shared models of in-class and oral assessments, adjust pacing guides to make room for writing in the room, and tell evaluators to expect it. For founders, the useful tools here lower the cost of this shift for teachers, like a quick way to run and record a short student conference, rather than adding another detection score.

Two writing and curriculum providers changed owners in the same week

IXL Learning acquired NoRedInk on October 19, adding a writing platform that says it has been used in 9 in 10 US districts since launch to a portfolio that already includes Teachers Pay Teachers and Rosetta Stone. Four days earlier, private equity-backed Activate Learning announced it had acquired Open Up Resources, the nonprofit that distributes openly licensed K12 math and EL Education ELA curricula. The combined company will operate as a public benefit corporation, but the announcement doesn’t say whether the Open Up materials will stay openly licensed and free. Both deals follow Carnegie Learning’s purchase of Ethiqly earlier this month.

For teachers and curriculum leaders, the product you chose is now part of a bigger company’s roadmap, and two details are worth checking before renewal. NoRedInk’s privacy policy, updated August 31, says some AI feedback reaches students directly without a teacher reviewing it first, and that AI output “may not always be accurate or appropriate.” Ask whether you can switch that route to teacher-approved, because feedback that skips the misconception, or gets it wrong, still lands on the student. The same policy offers a model clause: a buyer must honor the existing policy, and users get notice and a chance to stop using the product before any change. Check that your other vendor contracts say the same. If you rely on Open Up Resources materials, ask Activate in writing about the license terms for the next adoption cycle. For founders, both deals show where the exits are: buyers paying for classroom adoption that teachers earned by word of mouth.

Only 20 of more than 800 studies on AI in K12 can show whether a tool caused learning

Stanford’s SCALE Initiative reviewed more than 800 papers on AI in K12 and found only 20 high-quality causal studies of effects on students or educators. None of them tested a student-facing AI tool in a US K12 school — most were run in colleges or outside the US and most measured performance during use or soon after, so we know little about what knowledge students retained. The chapter came out in June and came back into view this month when Education Week cited it in coverage of a call for more research on AI in schools.

The thin evidence that does exist points to design. In a study of high school math students by Hamsa Bastani and colleagues, students using a tutoring-style tool that gave hints did better on practice problems but no better on a later closed-book exam, while students using a general chatbot did worse on that exam than students who had no AI at all. The authors conclude that “tools that guide students through reasoning rather than generating answers appear likelier to support durable learning,” while systems that hand over complete answers tend to reduce the cognitive work that produces learning.

On the teacher side, a randomized study found that ChatGPT cut lesson-preparation time while reviewers rated the materials comparable in quality, and the authors ask whether that saved time goes back into discussion, feedback and relationships with students.

For teachers, after students practice with an AI tool, give them a short check without it a week or two later. Scores while the tool is open show what students can do with help, and the later check shows what they learned. For district leaders, pilot new tools with a comparison group and track where any teacher time saved goes. For founders, a US K12 study that measures learning without the tool would be one of very few.

🔮 The call

Every week I make one prediction and I’m curious about your take. Reply or comment below. Here’s this week’s:

By June 30, 2027, at least one of Turnitin, GPTZero or Pangram will stop showing K12 teachers a standalone AI-likelihood score by default, and will lead with evidence of how the writing was produced instead. Humanizer tools now beat the detector the field trusted most, and the version-history trails sold as the fix can be faked too, so a percentage is getting harder for a teacher to defend in a meeting with a parent. Turnitin already hides scores below 20% to limit false positives, which shows the vendors know the number is fragile. The counter-case is that the score is what many schools pay for and what many teachers ask for, so vendors may add process features alongside it and keep the number front and center. If the call lands, schools will need to decide what counts as evidence in an integrity case, and that decision is easier to make in policy now than in the middle of a dispute.

Scoring the board. On September 28, I predicted that by June 30, 2027, at least one state education agency or top-25 US district would require a learning-outcome measure to approve AI tools, not only privacy, safety and usage terms. Verdict: RIGHT, but it doesn’t count as a forecast. Florida’s State Board of Education had already done it on September 16, twelve days before I made the call, by requiring districts to evaluate edtech by demonstrated learning outcomes and to show how that evidence drives purchasing and renewal. Other states now have a working model to copy, and districts outside Florida can ask vendors for the same evidence today.

This was a gap in my own research, and a reader caught it. That’s the reason I ask for your take every week. If you see something I’ve missed or gotten wrong, comment or reply, and I’ll correct the record here.

🛠️ Tools of note

(These aren’t startup investment recommendations – they show up here because they’re connected to trends worth tracking.)

Google’s classroom STEM simulations – Google is piloting a tool that lets teachers build interactive physics, chemistry, biology and math simulations for middle and high school, with a library of more than 30 teacher-vetted activities, through the Google for Education Pilot Program. This is a great move towards AI increasing learning friction. AI is starting to generate the simulation rather than the answer, so the student does the manipulating and predicting— and learning.

OpenAI’s interactive explanations – OpenAI’s education newsletter shows teachers how to prompt GPT-6 Chat for a working simulation inside the chat, with a sample-size lesson that asks students for a prediction before they roll and a reflection after. Two AI companies shipped the similar practical use ideas in the same week. Consider this a concerted move from AI chatbots to AI tooling for learning.

AI for Education’s “The Case for AI Literacy” – A free brief and family discussion guide giving school leaders the research, goals and three first steps for community-wide AI literacy. It’s built for the gap between banning AI and adopting it, and the family guide gives schools a way to bring parents into the conversation before an opt-out request does.

💰 Funding & market moves

Bloomy – The Y Combinator-backed K12 mastery-learning platform raised $3 million from Reach Capital, Silicon Schools Fund, Workshop Venture Partners and others. With Bloomy, students work one skill at a time with a tutor that asks questions instead of giving answers, and advance only after showing 90% mastery. The founder reports that pilot students mastered skills about twice as fast as the national average. However, take this with a grain of salt because they did not publish details on the sample or comparison. For teachers, a mastery gate is only as good as the check behind it – 90% on practice right after a lesson isn’t the same as remembering the skill a month later, so ask whether the tool brings mastered skills back for spaced review.

📚 Research & policy on my radar

An ETS breach exposed millions of California student records – Hackers used a compromised login to sit inside a system run by ETS, the state’s testing contractor, from May 22 to September 18, copying names, addresses, disability records, English learner status and more for current students and former students back to 2006. For districts, this is the data your state holds about your students through a vendor you didn’t pick. Families got notice weeks after the hack was found, so plan your own family communication now rather than waiting for the state. Ask your own vendors when they would tell you, and require multi-factor login in every contract.

Florida made binding AI rules for every public classroom from pre-K through college – A reader pointed me to this one, which I missed in September (see Scoring the board above). Florida’s State Board of Education adopted rules on September 16 that ban AI tools designed to simulate companionship or meet students’ emotional needs, bar vendors from selling student data or using it to train models, require notice to parents with a non-AI alternative, and direct districts to evaluate edtech by demonstrated learning outcomes rather than screen time. Districts must update their policies by July 1. For districts in other states, Florida’s list is a ready-made checklist for your next AI tool review, especially the requirement to show how evidence of learning drives renewal.

That’s The Syllabus for this week. Lots to dig into!

If this was useful: 📬 Subscribe on Substack – The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take. 🔗 Connect with me on LinkedIn – I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space: I advise school districts building AI frameworks and edtech founders navigating the K12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this: Forward it. Re-stack it. Share it on LinkedIn. The best readers I’ll ever have are probably in your network!

Hi, I’m Angela Chen! I’m an edtech founder, Stanford instructor, and AI in education strategist. Every week, I read across the full AI-in-education landscape: research, classroom practice, startup and funding signals, policy, and market intelligence. In The Syllabus, I’m sharing the signals I actually think matter, synthesized across everything I read that week!

If there’s something here worth going deeper on, comment below or message me on LinkedIn and tell me which signal – I’m always interested in what’s actually landing for people in the field.

This week, the signals pointed at feedback. A study of human teachers showed how rarely feedback helps students fix the thinking behind a mistake, while investors put money into AI that writes and grades feedback. A headline claim that AI tutors match human ones also turned out to rest on same-day results.

⚡ What to pay attention to right now

Most teacher feedback on wrong answers doesn’t help students fix the thinking behind them

In a USC study of nearly 1,000 math teachers, which didn’t involve AI, reported by Education Week, only about one in ten gave feedback aimed at the misconception behind a student’s error. Teachers asked to hold high expectations gave more specific comments. However, they also leaned on “you’re not trying hard enough,” and those comments were directed more often at girls.

High expectations help a student only when they come with a concrete next step. The study looked only at human teachers, but I’d hold AI feedback tools to the same bar, since a comment that skips the misconception doesn’t help a student whoever writes it. Pull ten recent comments, yours or a tool’s, and ask whether each one names the mistake in thinking and tells the student what to try next. Then check whether it worked with a similar problem a week later, without the feedback in front of them.

The claim that AI tutors match human tutors rests on same-day results

A Handshake-funded preprint, StudentBench, found that AI tutors matched expert human GRE tutors after a single session with 2,383 adults. Gains were measured only right after the session, and there was no group that practiced on their own, so we don’t know what students kept or how much came from the tutor rather than the practice.

Expect vendors to cite this in K12 pitches. When a product claims it “matches human tutors,” ask how long after the session learning was measured, compared to what, and with which students. Getting students to come back to a tool is only the first step. Retained learning and growth is what should be truly measured.

The money is going into AI feedback, and the evidence so far is the companies’ own

Carnegie Learning bought Ethiqly, an AI writing coach that will sit inside its grades 6–12 literacy curriculum, and UK startup sAInaptic raised £1 million for AI marking. Neither has published independent evidence that its feedback improves student work.

The big publishers are buying the feedback layer, so an AI coach may show up inside curriculum you already own. Review it as a new product, and borrow the bar from the teacher study above: does the tool’s feedback name the mistake in thinking and give the student a next step? The key question is who does the thinking: does the tool ask students questions about their draft, or rewrite it for them? For founders, a study of your feature’s effect on student work is what will set you apart from a tool a publisher can bolt on.

This is continuing to play out: screen-time limits are now colliding with digital state tests

More than half of the New York City Council asked the state to allow paper-and-pencil tests in grades 3 through 8, arguing that computer-based testing pushes up screen time. The state finished moving to digital tests only this spring, and the city now bans personal devices in its youngest grades.

The New York request is about screen time in general, not AI specifically. It still matters for AI, because screen-time caps set the number of minutes any AI tool can get in a school day. If your district caps screen time, count the device minutes your assessments require – state tests, benchmarks, screeners – before deciding what’s left for instruction. For founders, a minute cap means every product, AI tools included, competes for the same minutes, and the ones that keep their place will be the ones that can show what students learned.

🔮 The call

Every week I make one prediction and I’m curious about your take. Reply or comment below. Here’s this week’s:

By June 30, 2027, at least one company selling AI tutoring to K12 schools will publish a study comparing its AI tutor with human tutoring that measures what students learned at least one week after the sessions, rather than right after them. StudentBench showed how easy it is to claim an AI tutor “matches human tutors” on a same-day test, and I expect buyers to start asking what students still know a week later. The counter-case is that delayed measures cost more and can make a product look worse, so companies have every reason to stick with usage numbers and immediate gains. One to watch is Eedi, a UK math platform, which tested human-supervised AI tutoring against human tutors with Google DeepMind and has announced a follow-up study of longer-term learning gains. If the call comes true, schools will finally have a comparison worth using when they decide where tutoring dollars go.

💰 Funding & market moves

Kollegio – The New York-based AI college-guidance platform raised more than $5.3 million in pre-Series A funding from Reach Capital, JFFVentures, ECMC Group and others, announced October 7. It serves more than 300,000 students, works with more than 30 college partners, and says it wants to become “the navigation layer for all education and career decisions.” On the same day, OpenAI announced a free College Planner inside ChatGPT for Teens to help high school students track applications, deadlines and financial aid.

For founders, that timing is the lesson. When a frontier lab ships a free version of your core feature, a smoother chatbot won’t protect you. What a general-purpose tool can’t easily copy is the institutional side of the work – the college partners, the counselors who need to see what students are doing, and evidence of results. Kollegio is building that evidence by partnering with Learning Collider to measure whether the platform works, and the measure that will matter is whether students actually apply, get aid and enroll, not how many use the app. If your product sits in a category a lab could add as a feature, decide now which of those you own.

📚 Research & policy on my radar

Digital Promise is funding open tools for AI-driven formative assessment – Grants of up to $600,000, due November 20, will fund openly licensed tools that let teachers see and steer what AI is doing in their classrooms. For districts and research partners, this is a chance to build feedback tools in the open, with teachers deciding the next step.

AI improves fastest where answers can be checked – A secondary history teacher of 30 years, writing in Teaching in the Age of AI, asks what OpenAI’s recent math results mean for K12. OpenAI said on October 6 that an unreleased model had produced hundreds of new mathematical results, after earlier claims to have disproved a 1946 Erdős conjecture and solved a version of the Navier-Stokes problem, a claim mathematicians are still checking. He draws on Yuval Levin’s explanation in The New Atlantis: AI improves fastest where its answers can be checked, because code either runs or it doesn’t and a proof can be verified, so a model can test, fail and learn at enormous speed. On that view, AI progress may get narrower rather than more general, racing ahead on tasks with a verifiable answer and moving more slowly where quality is a matter of evidence, interpretation and judgment.

For teachers, this gives you a way to sort your assignments. Tasks with one checkable answer are where AI is strongest and improving fastest, so they’re the easiest for a student to hand off. That doesn’t make them less worth learning, because students still need that knowledge to judge what AI gives them. However, it does mean the practice should happen where you can see it, with the thinking shown. Tasks that ask students to weigh evidence, defend an interpretation or explain their reasoning keep more of the work with the student, and they’re where your feedback does the most.

If one of these is landing in your school or your product this week, reply and tell me what you’re seeing. I’d love to go deeper on whatever is most useful.

That’s The Syllabus for this week. Lots to dig into!

If this was useful: 📬 Subscribe on Substack – The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take. 🔗 Connect with me on LinkedIn – I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space: I advise school districts building AI frameworks and edtech founders navigating the K12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this: Forward it. Re-stack it. Share it on LinkedIn. The best readers I’ll ever have are probably in your network!

Hi, I’m Angela Chen! I’m an edtech founder, Stanford instructor, and AI in education strategist. Every week, I read across the full AI-in-education landscape: research, classroom practice, startup and funding signals, policy, and market intelligence. In The Syllabus, I’m sharing the signals I actually think matter, synthesized across everything I read that week!

If there’s something here worth going deeper on, comment below or message me on LinkedIn and tell me which signal – I’m always interested in what’s actually landing for people in the field.

This week, the strongest signals were about who protects the thinking students do with AI one tab away. Common Sense Media’s testers found that ChatGPT’s teen safeguards missed crisis signs and that students can switch off its study mode in a click, curriculum reviewers found AI features arriving with little evidence of their own, and the move to ban student AI spread from the largest districts to the institutions around them, with National History Day barring generative AI from student projects.

⚡ What to pay attention to right now

Independent testers rated ChatGPT’s Teen Mode an “unacceptable risk,” and its study mode is a toggle students can turn off

Common Sense Media ran more than 4,000 prompts on accounts registered to 13- to 17-year-olds, with responses reviewed by child psychiatrists and a pediatrician. The results prompted Common Sense to rate ChatGPT’s Teen Mode as an “unacceptable risk” for users under 18. ChatGPT missed more than one in four crisis referrals the testers judged warranted, while parent accounts received no alerts during testing of concerning prompts, including self-harm. Alerts appeared only after signs of trouble persisted over weeks, which may be too late.

ChatGPT also couldn’t reliably tell a user’s age, so a teen may never be routed into Teen Mode. For homework, study mode was easy to switch off, and even with it on, testers often saw a popup offering to show the answer. OpenAI disputes these findings and told Education Week the testing doesn’t reflect how its safeguards work in practice. It also pointed to technical errors that it believes blocked the parent notifications.

For school leaders, the problem is general-purpose chatbots, not one company. Think twice before giving students open access to any of them, including education editions like ChatGPT Edu and Claude for Education, which are built on the same chatbots and mostly add control over data and accounts. When Instruction Partners watched 20 student-facing AI products in use across 16 school systems, general-purpose chatbots posed the biggest risk to student thinking, and tools built for teaching gave the most consistent learning.

Don’t list a consumer product’s teen settings as a safeguard in your AI policy or your parent communication, since linking accounts didn’t produce timely alerts in this testing. When you evaluate a tool for students, ask whether a student can turn off its tutoring behavior, whether a parent or school would know when a conversation turns to self-harm, and what the tool was designed to do besides answer questions. For teachers, treat study mode as the student’s choice rather than a guarantee, and move the generating work – first attempts, worked steps, short explanations – into class where you can see it.

Only one of ten curriculum and edtech providers could point to third-party evidence that its AI feature improves student outcomes

EdReports, the nonprofit that reviews K12 instructional materials, published a qualitative scan of how 10 curriculum and edtech providers describe and support their AI features, including four of the ten largest publishers of purchased materials. In the full brief, only one company cited evidence from a third-party evaluator for the student outcomes of its AI tool. Companies more often pointed to efficacy research on their established programs, research that rarely covers the AI feature and often predates it, and in some cases that legacy data appears right on the marketing page for the new AI product. Only two named any independent evaluator for an AI feature at all, and several pointed to membership in responsible-AI alliances instead, which EdReports notes don’t evaluate individual products.

Classroom observation points the same way: of the 20 student-facing AI products Instruction Partners watched in use, only 7 had independent studies of student outcomes. A district can adopt a curriculum on the strength of its evidence and find an AI feature inside it a year later that no one reviewed. A product’s efficacy study covers the product that was tested, not the chatbot added to it this fall.

For district leaders and curriculum directors, ask for actual evidence on the AI feature itself before you adopt or renew:

• What testing has been done on it, by whom, and with which students?

• What has it been shown to improve, and by how much?

• For the districts that adopted it earlier, how has it moved student learning, measured on work students do without the tool?

If a vendor can only answer with research on the base program or with usage numbers, you have your answer. Then write the vendor’s obligation to tell you before adding or changing an AI feature into the contract. For founders, an efficacy study of the feature itself is becoming the cheapest way through procurement.

The move to ban student AI use has spread from the largest districts to the institutions around them, with National History Day barring it from student projects

This fall, New York City and Los Angeles, the two largest US districts, paused student generative AI – New York through eighth grade, Los Angeles on district devices for one year. Now the restrictions are reaching beyond districts. National History Day, a research competition for grades 6 through 12 with more than half a million participants a year, now bars generative AI for “research, analysis, design, or creation” of projects, effective October 1 for the 2027 contest. That covers uses it previously allowed, like brainstorming topics and finding sources, as well as revising writing and asking chatbots for feedback. Spell check, bibliography generators, and rule-based tools like search ranking are still allowed. Suspected violations will be handled much like plagiarism, and students who used AI for brainstorming or sources before October 1 won’t be penalized if they cite it. Education Week reports that NHD flagged at least two projects in the last two contest cycles that cited sources AI had invented.

The district pauses govern what happens on school devices and in school hours, while a national contest rule follows the student wherever they do the work. A student in a district that teaches guided AI use can now be working under a full ban on the history project, while a student in New York may be under two sets of rules at once. NHD’s reasoning is about the work itself – the contest exists to show what a student can do with primary sources, and judges question students about their process. Its rule also reaches the steps where students most often get stuck — finding a topic and locating sources — where an AI conversation can help without technically writing a single word of the project.

For social studies teachers running NHD this year, update your project timeline now and keep generative AI out of the project itself, then teach source evaluation with AI elsewhere in your course – for example, students write their own research questions first, then compare them with what a chatbot suggests and discuss what it missed. For school leaders, expect more contests, scholarships, and dual-enrollment programs to set their own AI rules, so your AI guidance should tell students to follow the strictest rule that applies to a given piece of work. If your district is weighing a pause of its own, it should answer the same questions an adoption would: what it protects, for whom, and how you’ll know.

🔮 The call

Every week I make one prediction and I’m curious about your take. Reply or comment below. Here’s this week’s:

By June 30, 2027, at least two more major national high school competitions or college-entry scholarships will require students to disclose or attest to how they used AI, the way the Regeneron Science Talent Search and Regeneron ISEF already do. National History Day went further than anyone else this fall by banning AI from the process itself. However, the more common move among prestigious programs is disclosure. The Regeneron Science Talent Search now asks students to cite AI help and keep a log of their prompts for uses like brainstorming and coding, while barring AI from writing the paper or supplying citations, and ISEF requires any AI used as a project resource to be cited. The Scholastic Art & Writing Awards rule out entries whose final product is AI-generated. Large college-entry scholarships like the Coca-Cola Scholars Program don't appear to have a published AI policy yet, and that's where I expect the next ones to come from. If this lands, keeping a record of how you used AI becomes a skill students need well before college, and it's one teachers can build into assignments now.

Which thread do you want me to pull next – what to ask before giving students an education chatbot, how to get real evidence on AI features, or what the spread of AI bans means for your students? Reply and tell me.

That’s The Syllabus for this week. Lots to dig into!

If this was useful: 📬 Subscribe on Substack – The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take. 🔗 Connect with me on LinkedIn – I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space: I advise school districts building AI frameworks and edtech founders navigating the K12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this: Forward it. Re-stack it. Share it on LinkedIn. The best readers I’ll ever have are probably in your network!

By now you have probably read the Wall Street Journal’s investigation, “Google Pushed AI Into Schools – and Even Students Say It’s Gone Too Far,” by Shalini Ramachandran, Georgia Wells and Molly Reinmann. It is built on more than 60 interviews with teachers, students, parents and clinicians, and it is hard to read if you have a child in school. I flagged it quickly in this week’s Syllabus. This is the follow-up on what to do about it:

First, schools deserve the choice to opt out, and Google should give it to them.

Second, we can’t wait for that choice to do the work. The AI in Search is the one AI tool nearly every student already uses, which makes it the best place to teach using AI with judgment and discernment.

If reading this sparks any thoughts, I’d love to connect!

The switch that doesn’t exist

Google lets schools turn off its Gemini products. Its own admin help pages show how to switch off the Gemini app and Gemini in Google Classroom for each group of students. The Gemini app page also notes that its setting “has no effect on other AI features.” Search is one of those other features.

The Journal reports that AI Overviews and the AI Mode chatbot “cannot be turned off in administrative controls Google gives partner school districts,” and Google’s admin help pages offer no setting for them. That is a real problem. In July, Common Sense Media rated Google’s AI Search an “unacceptable risk” for kids and advised students to stop using it until schools can turn it off. Google disputes that testing, saying it relied on “ambiguous and contrived queries.” Either way, schools can’t turn it off. So some of the largest districts are building workarounds instead. Los Angeles blocked AI Mode with its own tools. Fairfax County is building a plug-in to hide Google’s AI results on student devices.

Workarounds like these only reach so far. They cover the devices a school manages. Students also search on their phones, on family computers and anywhere else they can get online. And Search is where most of them meet AI first. A Common Sense Media survey of 1,200 children aged 9 to 17 found that about three in four had used AI summaries in search, more than any other form of AI.

A Fairfax student showed the Journal what that box can do. “The Google one was the biggest temptation, because it’s the easiest to access,” the 15-year-old said. Her grades in Geography and Algebra II fell to C’s while she leaned on it. They came back when she stopped.

Her story fits a wider pattern, with one caveat. The Journal cites a striking finding from the OECD’s PISA 2025 results: students who used AI daily for specific schoolwork tasks scored about 20 points lower in science than students who never used it – a gap equal to roughly a year of school. That is an association, not proof that AI caused the drop. The same data found that students who used AI to learn, with AI literacy training, scored slightly higher than non-users. The finding is a reason to teach students to use AI with judgment and discernment, so it boosts their learning.

Turning a lemon into lemonade

The lemon is the missing switch, and Google should fix it. A school that manages a student’s account should be able to decide what that account sees. Google already accepts this logic for Gemini. It should apply the same logic to Search for managed student accounts. Some schools will keep AI in Search on and teach with it. Others will turn it off for younger grades. Either decision belongs to the people who know the students.

A switch is only part of the answer, though. Like the workarounds, it would stop at school devices. It would not reach the phone in a student’s pocket or the laptop at home. A full ban has the same limit, and it also gives up the chance to teach.

One independent school I advise made this case in a thorough Q&A it sent families after the Journal’s story. Its students will use AI in college and at work. They build judgment by practicing with AI while a teacher is in the room, not by avoiding it until they leave. So the school does both. It limits where AI shows up at school, and it teaches students when AI helps learning and when it gets in the way.

That second half is the lemonade. Google made AI in Search unavoidable, but unavoidable also means shared. Students meet the same AI in the classroom and at the kitchen table. That gives schools and families one common tool to teach judgment with, instead of a dozen apps adults can’t track.

Deep learning still requires productive struggle. Our job is to make sure the box at the top of the page doesn’t take that struggle away, and that students learn to tell the difference. That job belongs to schools, students, parents and teachers together. Here is what each of us can do.

What teachers can do

Assume Google’s AI Search can do any take-home assignment. When Common Sense Media tested Google’s AI Search, AI Mode gave complete, submittable answers to every assignment it tried.

• Start with an AI Overview audit. Common Sense found that 43% of repeated history questions got materially different answers, and that 29% of citations came from forums and social posts. Give students one class period to search the same question twice, compare the answers and check where each one came from. Then ask them: if the answer changes every time, what did you learn by copying it? The goal is for students to watch the tool get it wrong themselves, so checking becomes their habit and not just your rule.

• Bring the hard part into class. Wharton professor Hamsa Bastani told the Journal that students reach for shortcuts “precisely at the moment the work starts to get mentally taxing.” Move that moment into the room. In English, students draft the thesis and outline in class and finish the essay at home. In math, they work the two hardest problems in class and do routine practice at home. You don’t need to redesign every assignment. Start with one per unit.

• Watch the gap between homework and tests, then talk. A Pennsylvania eighth-grader in the Journal’s story kept mostly B’s and failed his final. A widening gap between take-home grades and in-class results is an early warning. When you see it, have a private, restorative conversation, not an accusation. Ask what’s happening and what’s hard right now. Then agree on one change together, like starting the next assignment in class or checking in on a first draft.

What students can do

The question you type decides whether AI teaches you or does the learning for you. The Pennsylvania eighth-grader in the Journal’s story started by typing his algebra homework into the search bar. Then he leaned on Google’s AI for other classes too – to get the parts of a volcano and to write an essay about Edgar Allan Poe’s “The Raven.” He got the answers each time, and still failed the final.

However, the same tool could have taught him both of those things. Compare:

• “What are the parts of a volcano?” hands you a list to read. “I’m going to sketch a volcano from memory and label it. Then quiz me on what I missed” makes you pull the answer out of your own head first, then shows you the gaps. Teachers call this retrieval practice, and it is one of the most reliable ways to make learning stick.

• “Write an essay about The Raven” hands the thinking to AI. “I think the raven stands for grief that won’t leave. Push back on my reading using lines from the poem” starts with your idea and makes you defend it with evidence. The essay that comes out of that is yours.

I know this is easier said than done at 11 p.m. with three assignments due. That’s why the habits matter more than good intentions.

Start with an AI reflection. Take 30 minutes to an hour and write down your answers to a few questions – What do I want AI to do for me, and what do I want to keep doing myself? What counts as my own attempt in each class – a first draft, a worked problem, a list of ideas? When do I usually reach for AI, and why? What would good help from AI look like?

Turn your answers into a master prompt. Use what you wrote to tell AI how to work with you. Keep it in your phone’s notes and paste it at the start of any chat, or save it in the chatbot’s settings if it has a place for instructions. A starting point:

I’m a high school student. When I ask for help with schoolwork, don’t give me the final answer unless I ask you to check one I’ve already written. Ask me one question at a time to help me find the next step. If I’m wrong, tell me where, not what the right answer is. Facts, definitions and dates are fine to give me directly. In [class], my own attempt means [what you wrote]. Ask to see it before you help.

Then keep three habits:

• Try first, then check. Write your own attempt, even two lines, before you use AI. Before you hand in work, ask yourself: “Could I explain this without the AI?”

• Practice without it before a test. “When I was in a test and I couldn’t use AI, it was hard to think,” the Pennsylvania student said. Do one practice set with no AI, or ask AI to quiz you one question at a time, so the exam isn’t the first time you work alone.

• Use AI, then put it away. Ask AI to make a practice test or worksheet, print it, and do it on paper. Check your answers afterward. You don’t have to work alongside AI the whole time to get the benefit of it.

If you find yourself leaning on AI for advice about your life, talk with a trusted adult. Your advisor or counselor is a good place to start.

What parents can do

You don’t need to be an AI expert, but you do need to be present – more curious than watchful, and asking questions instead of setting traps. A teenager who expects to get caught learns to hide their AI use. A teenager who expects a real conversation is more likely to tell you what’s going on.

Your coaching follows what teachers and students are already working on:

• Do the AI reflections together. Sit with your teen while they answer the reflection questions in the students’ section, and answer them yourself. Say where AI helps you at work and where you choose not to use it. Then help them write their master prompt. You’ll learn more about how they use AI from that conversation than from anywhere else.

• Ask for a walk-through. Before an assignment is submitted, ask your child to talk you through the content and how they used AI.

• Notice the small hand-offs. A Georgia teacher told the Journal her students asked Gemini how many hours were left until 2 p.m., at 11:30 a.m. When your child asks AI something they could work out, ask them to try first.

• When a grade surprises you, ask before you assume. Strong homework followed by a weak test is the same early warning teachers watch for. Start with “What happened?” rather than “Were you using AI?” Then talk with the teacher, so you’re working on the same plan.

• Watch what AI is for. Notice if your child turns to AI for advice or company rather than schoolwork. If you see this, have an honest conversation.

That last point was the hardest-hitting part of the Journal’s story for me. It describes a high school junior who came to see Google’s AI as a confidante – one that, according to chat transcripts the Journal reviewed, described her mother’s screen-time limits as “coercive control.” Common Sense found that AI Overviews offered a hotline or medical referral in only 58% of searches that showed clear signs of crisis. If your child is in crisis, call or text 988.

What school leaders can do

School leaders make the policy calls and fund the training that make the rest possible. I work as an AI strategist with K-12 schools from California to Florida, and the same questions come up at every one: what’s allowed, who teaches it and which tools to trust.

• Write AI policy at the course level. A school-wide policy sets the reasons, but students need an answer for each class. Give every course a short role statement saying when AI may, must or must not be used, and why. It answers the question students ask first: “Am I allowed to use it for this?” I’m writing these course by course now for an independent school in the Bay Area.

• Train teachers before you ask them to teach AI, or to teach with it. A policy only works if teachers can implement it confidently. Pair teachers with one-on-one support and a short checklist for each assignment, and run the AI Overview audit as a faculty workshop first. Then measure teacher confidence as you roll it out, so you know whether the support is working.

• Choose AI that students open on purpose. AI built into search, documents and learning platforms shows up at moments of temptation. A separate tool that students open for a task gives them a moment to decide first. The Bay Area school I mentioned above keeps AI out of its learning platform and saves the Gemini chatbot for staff. List AI in Search in your approved-tools registry too, with guidance, because it’s the AI students meet most.

• Let learning science pick the tools. Hamsa Bastani, the Wharton professor quoted above, co-authored a study of nearly 1,000 high school math students in Turkey. Students who practiced with standard ChatGPT did 48% better during practice and 17% worse on the exam. Students who used a version that gave hints instead of answers held even. Google’s own research points the same way: a Google-funded trial in Sierra Leone found students gained more than a year’s math progress in eight weeks with Gemini’s Guided Learning mode, which guides students through questions instead of answering them – though the strongest students gained the most. Google has already built the kind of AI that helps students learn. Schools should be able to choose it. Before you sign your next contract, ask every vendor: does this tool give students answers, or does it encourage them to think?

Last, use your leverage. Los Angeles has already asked Google to turn off AI Overviews. A joint request from several districts and schools, timed to contract renewal, is harder to set aside.

What success looks like

The measure of success is not how much AI students use. It’s whether every graduate leaves with the judgment and confidence to use it well. Google should give schools the switch. Either way, the teaching is ours to do. Judgment, not tools, is what makes AI yours.

This piece follows this week’s Syllabus and draws on the Wall Street Journal’s reporting by Shalini Ramachandran, Georgia Wells and Molly Reinmann (September 30, 2026). If your school has found its own way through this, I’d value hearing what worked.

Hi, I’m Angela Chen! I’m an edtech founder, Stanford instructor, and AI in education strategist. Every week, I read across the full AI-in-education landscape: research, classroom practice, startup and funding signals, policy, and market intelligence. In The Syllabus, I’m sharing the signals I actually think matter, synthesized across everything I read that week!

If there’s something here worth going deeper on, comment below or message me on LinkedIn and tell me which signal – I’m always interested in what’s actually landing for people in the field.

A quick note: this issue is landing a day later than usual because I was away for a most interesting weekend wedding. Thanks for your patience.

This week, the strongest evidence was about what happens around an AI tool once it reaches students: the defaults schools can’t switch off, the permission it signals, the trust it strains, and the coursework students quietly stop doing. That is where most of our rollout decisions live, and it is where I’d direct your attention.

⚡ What to pay attention to right now

Schools can switch off Google’s Gemini tools, but not the AI answers in Google Search

A Wall Street Journal investigation published September 30, based on more than 60 interviews, found that every Gemini product Google sells to districts has an admin off switch, but the AI Overviews and AI Mode in Google Search do not – and Search is where students go first for homework, on the Chromebooks most districts already run. Google’s own researchers warned internally about cognitive and emotional dependence in children. The Journal also cites data showing students who use AI daily for school scored 20 points lower on science tests than students who never do, an association rather than proof of cause. Meanwhile, a Gallup and Walton Family Foundation survey of 2,101 teachers found teachers want control over classroom tech more than bans, and those with substantial oversight were less than half as likely to call devices a distraction (20% versus 52%).

Most district AI policies govern the tools a district buys. Now, students’ first AI is arriving inside tools the district already owns, so that policy now covers a shrinking share of what students use. Expect the next fight to be over defaults: districts asking Google whether Search AI can be disabled for managed student accounts, and boards asking why the district’s AI rules stop at the procurement list. Teachers should plan as if an answer engine sits one tab away from every assignment, which shifts more of the thinking work into class time where they can see it. For founders, an off switch a district controls may become a feature buyers will ask for by name.

A course-integrated AI tutor lowered final grades in a 2,379-student randomized trial

A University of Maryland team randomized access to a GPT-4o study assistant across 30 instructors’ courses in fall 2025. Access lowered final grades by 0.27 to 0.37 standard deviations and cut participation in instructor-designed activities like discussion posts and quizzes by 0.90 standard deviations, with grade losses for first-generation students more than twice as large. Only about 15% of students in the sections with access ever opened the tool, but grades dropped about as much for the students who never used it as for those who did. Nearly three-quarters of the requests students sent asked for explanations or solutions, and under 1% asked for feedback on their own work.

This is a working paper about undergraduates at one university, but its implications extend past higher ed. The authors’ likeliest explanation is the message the tool sent: an AI assistant built into the course page told students AI was welcome (students in those sections were nearly four times as likely to say their instructor actively encouraged it), and every student already had free access to ChatGPT and Gemini. The study can’t confirm that students leaned on those outside tools more, but participation in the course’s own discussions, quizzes, and assignments fell sharply, and grades fell with it.

So an AI tool rollout is a decision about classroom norms, and it deserves the discipline of an experiment – some sections piloting first, a clear statement of what the tool is for, and a check on whether students keep doing the practice and discussion you designed. The first-generation gap is the warning for equity-minded leaders, because the students with the least support at home are the ones it hurt most.

Teachers and students now suspect each other, and AI policies aren’t telling either side what to do

Harvard’s Center for Digital Thriving calls it “two-way suspicion” in An AI Policy Isn’t a Playbook. Of teachers who had faced an AI dilemma, 74% described something cheating-related, and some of the 31 teens interviewed said they deliberately turn in weaker work so they won’t get flagged. Microsoft’s six-country survey shows the mirror image: 42% of educators rank cheating and 41% of students rank being accused a top concern.

The cost lands on effort and relationships, which a policy audit can never capture. A student who diminishes the quality of their work to avoid suspicion is choosing not to learn, and a teacher acting as AI referee has less time to teach. Roughly four in five education leaders in Microsoft’s survey call their AI guidance clear, while about half of teachers and students rate it neutral or say it isn’t provided – leaders are judging the document, while teachers are judging the moment a hunch goes off.

What closes part of that gap is a set of AI academic integrity guidelines, which is the playbook Harvard’s report is asking for: what counts as evidence, how a student shows their process, and who the first conversation is with – co-written with students. It’s the work I do with the schools I advise, and it’s where I recommend you start. Labeling each assignment with how much AI use is allowed sets the line before the work starts instead of arguing over it afterward. Moving more of the thinking into class, through drafts, conferences, and short oral explanations, gives teachers evidence they’ve seen rather than a guess. Then ask students anonymously whether they’ve held back effort, which tells you whether it’s working.

A named funder and research partner are becoming what gets AI tools into districts

Third Space Learning is offering fully funded AI math tutoring to 3,000 students in grades 3 through 8 this school year, in a Gates-funded study with Stanford and Cornell that asks whether AI can match human high-dosage tutoring, whether students stay motivated, and whether access is equitable. Districts pay nothing and share aggregated data and teacher feedback in return.

For schools, this is the shape free AI pilots are increasingly taking: subsidized, attached to a named funder and/or a university study, and easy to say yes to. But before you sign, make sure you’ll get your own students’ outcomes back against a comparison group and know what happens when the grant ends.

The larger shift is what research partnerships are now for. When I was at the Stanford Accelerator for Learning, I saw a steady stream of inbound requests from companies wanting to partner on research, and I always suspected the partnership mattered to them for more than the eventual result. Jacob Kantor, who connects edtech companies with district leaders, wrote in Edtech Insiders that he has taken the Third Space story into more than 30 district meetings and come out with more than 30 pilot yeses. In his telling, the product pitch wasn’t the reason: when a superintendent hears “funded by Gates” and “studied by Stanford and Cornell,” saying yes to a no-cost pilot “becomes the safe choice.” In a crowded AI market where products are starting to look alike, a credible study can be what gets a company into a district in the first place, years before any RCT reports. That raises the stakes for universities lending their names, and it means founders should treat a research partner as distribution as much as validation.

Alpha School’s Timeback software could power low-cost private schools just as federal scholarship money arrives

Michael Horn and Thomas Arnett argue in the second of their two Alpha School pieces that Alpha’s $65,000 school isn’t disruptive, since it enters at the top of the market, but its Timeback platform could be. Timeback runs Alpha’s two-hour morning academic block on mastery progression and spaced practice, with adults working as guides and motivators rather than instructors, and Alpha is now offering it beyond its own campuses. Houston ISD is piloting Timeback at two elementary campuses this fall. That matters more than the school itself – staff are the biggest cost of running most schools, and while Alpha pays its guides well and keeps ratios low, which is part of why it costs $65,000, other operators could run the software with much leaner staffing. Once the platform travels, anyone can try the model, from a public district like Houston to a small private operator.

The money is arriving at the same time. On October 1, Treasury released rules for the new federal scholarship tax credit, which from 2027 gives donors a credit of up to $1,700 for gifts to scholarship organizations in states that opt in. Whether your state opts in will decide whether that new donor money reaches families near you, and it can pay private school tuition.

Put the two together and an AI-run microschool could become affordable for families who would never pay $65,000: the software lowers what a school costs to run, and tax-credit scholarships help cover what families pay. So competition for public schools may come less from a single well-known private school than from one platform licensed to many small operators, each funded by scholarships. For district leaders, the useful moves are to find out whether your state plans to opt in, and to decide which parts of the model you’d adopt yourselves, as Houston is doing. Teachers will hear parents ask about it, and the strongest answer names what the software does well, like paced practice, and what a teacher does that it can’t. For founders, the platform is the business to watch.

🔮 The call

Every week I make one prediction and I’m curious about your take. Reply or comment below. Here’s this week’s:

By June 30, 2027, at least one of OpenAI, Anthropic, Google, or Microsoft will publish results from a randomized controlled trial of its own student-facing education product that reports a learning outcome – a test score, grade, or assessed skill – rather than only usage or satisfaction.
The Maryland trial showed that a course-integrated tutor can lower grades, Microsoft's big education report this year was a perception survey, the Journal reports that Google's own researchers flagged risks to children internally, and the Third Space study shows how much districts already reward a named research partner. If a lab publishes an RCT, districts finally get a yardstick for comparing free tools. If none does, the evidence schools rely on will keep coming from independent researchers testing setups the labs don't control.

💰 Funding & market moves

Pearson – on September 29 agreed to acquire Workera, an AI-native skills assessment platform that combines AI with psychometrics to test employees through role-specific scenarios and simulations. Earlier in September it bought Internet Testing Systems, a Baltimore assessment technology company of about 230 people that runs remote testing and proctoring. Terms of both deals weren’t disclosed. Pearson also already proctors Anthropic’s Claude certification exams.

I’m writing a longer piece on this pattern for publication, so here’s the short version: over the past year, the largest AI companies made their teacher tools free, while publishers have been buying the tools that measure what happens in classrooms – McGraw Hill bought TeachFX, which analyzes classroom audio, two weeks after buying Teachally, and Pearson agreed to buy two assessment companies in September alone. Measurement tools hold far more sensitive data than a textbook, including recordings of students’ voices, and a vendor’s privacy promises may not travel with that data when the company is sold.

For district leaders, the quick win is in your contracts. At your next renewal, ask for a change-of-control clause that gives you written notice before any data moves and the right to delete it or end the agreement.

📚 Research & policy on my radar

A bipartisan Senate bill would pay states to keep phones out of schools – Senators Mark Kelly and Tom Cotton introduced the UNPLUGGED Schools Grant Act, which would fund lockers, pouches, and storage for states that bar personal devices, including smartwatches and tablets, during the school day, with exceptions for special education and medical needs. It has AFT and NEA support, which is rare. If it passes, the line between phone bans and AI bans will matter more, so districts should write their device policy and their AI policy as separate documents with separate reasons.

Microsoft’s enforceable AI privacy standard for schools takes effect November 1 – Microsoft, the American Federation of Teachers, and the United Federation of Teachers announced a legally enforceable National AI Safety and Privacy Standard in September. It bars using student or educator data to train AI models, bans AI companions and student monitoring, requires a human in the loop on AI-driven decisions, and lets districts pause or end a deployment if Microsoft breaches it. Negotiations with Anthropic and OpenAI to adopt the same terms are reportedly ongoing. For district leaders, this gives you contract language to point to in every AI renewal, including with Google after the Journal’s reporting. Ask each vendor whether they’ll sign terms like these, and treat a no as information.

OpenAI pulled its next model after it acted without permission in testing – The Wall Street Journal first reported on September 28 that OpenAI had canceled the planned October release of GPT-6.1 Astra, and OpenAI confirmed it the same day. Its head of safety systems said the model “didn’t quite meet the bar in terms of staying within scope and authorization.” A lab holding back a model for overstepping is the safety process working. It also hands you a question for any vendor offering agents to your school: what can the agent do without asking, and how will you know what it did?

What do you want me to go deeper on – what students can reach on a district Chromebook, the Maryland trial, the trust problem, or the Timeback-plus-scholarships question? Reply and tell me.

That’s The Syllabus for this week. Lots to dig into!

If this was useful: 📬 Subscribe on Substack – The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take. 🔗 Connect with me on LinkedIn – I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space: I advise school districts building AI frameworks and edtech founders navigating the K12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this: Forward it. Re-stack it. Share it on LinkedIn. The best readers I’ll ever have are probably in your network!

Hi, I’m Angela Chen! I’m an edtech founder, Stanford instructor, and AI in education strategist. Every week, I read across the full AI-in-education landscape: research, classroom practice, startup and funding signals, policy, and market intelligence. In The Syllabus, I’m sharing the signals I actually think matter, synthesized across everything I read that week!

If there’s something here worth going deeper on, comment below or message me on LinkedIn and tell me which signal – I’m always interested in what’s actually landing for people in the field.

This week, measurement moved to the center. We got the first clean experimental answer on what actually makes an AI tutor teach, we learned that the version history teachers have been treating as proof of authorship can be manufactured in an afternoon, and two separate market signals showed who really decides which tools reach your classroom.

⚡ What to pay attention to right now

Making an AI tutor more humanlike does not make it more educational

Three Boston University researchers ran a preregistered randomized trial that separated what an AI tutor teaches from how it talks to students. Students working with a tutor grounded in their actual course materials gained 6.63 points over matched peers who had access to consumer AI , and the gain showed up in written reasoning. Additionally, although access to conversational voice nearly doubled how much students interacted with the tutor — and cost 2.8 times as much to deliver — it didn’t improve student content mastery.

While the sample was 86 graduate MBA students, tool engagement and learning mastery became decoupled inside this experiment, with significant implications for K12. For teachers and instructional leaders, this rewrites the vendor conversation. Stop asking whether students come back to the tool and ask what the tutor makes students practice, whether it is grounded in your curriculum, and whether students can practice the skill or recall the information later without the tool. For founders, a conversational voice layer is a preference feature you are paying nearly 3x for until you can show it changes what a student can do.

Version history is no longer evidence of authorship

An instructional designer handed Claude a finished essay and asked it to retype the piece into a Google Doc over two hours, the way a person slowly works through a draft. When he returned, the document’s version history showed what looked like two hours of human drafting. Some instructors responded to generative AI by moving writing into Google Docs so they could watch the process unfold. Even Blackboard bought Cursive for keystroke and process replay earlier this month, but the process trail itself has now turned out to be forgeable.

What matters here is measurement rather than enforcement. Skill forms through productive struggle. The revision trail was only ever a proxy for whether that struggle happened, so losing it means the struggle now has to be measured directly. Underneath that sits the sharper risk of cognitive offloading. Some offloading is reasonable, in the way that most of us drive without understanding the engine, and our work ahead is deciding which skills fall into which category for which students in which subject — a judgment no single policy can settle at once.

For teachers and deans, your next move is to stop weighting artifacts of process and put weight on assessments where a student explains, defends, or extends the work in front of you. However, watch the equity cost as you shift your modes of assessment, because oral and in-person formats land hardest on students with learning differences and access barriers. For founders selling integrity tooling, a product whose output is a verdict on a finished document no longer has a defensible claim to make.

Governments are moving faster than capital, in opposite directions

QS classified the major education developments of the last two weeks and found government actors behind roughly 43% of them, nearly double the share led by companies, with venture capital barely registering. But governments didn’t all move in the same direction.

The UAE approved an AI curriculum for every public and private school and is training 22,000 teachers to use AI in teaching, assessment, and lesson planning, while Saudi Arabia’s communications ministry and Google are giving a million university students a year of free Gemini access, and Canada has launched a national AI literacy initiative with the Alberta Machine Intelligence Institute that provides free AI learning to students, educators, and workers. Over the same period New York City paused student-facing generative AI for nearly 600,000 students from 2-K through grade 8, and LAUSD restricted student access to generative AI on district-issued devices for the 2026-27 school year.

If you lead a district under a restriction, or are weighing one, your pause needs to answer the same questions an adoption would:

• what the policy permits and why

• how it reaches teachers, students and families families

• which tools get evaluated and against what learning criteria

• what students learn about AI while they cannot use it

• how each decision lands for students with learning differences and access barriers.

For founders, NYC’s review standards show what a large district will now ask a vendor to prove, and learning design and instructional impact are on that list.

The state narrowed your vendor list before your committee met

New EdWeek Market Brief survey data shows state-approved materials lists are now a significant driver of which curriculum products districts actually buy, according to the K12 officials they surveyed, who described the lists as playing a significant role in determining what their districts choose. For a district leader, that means a meaningful share of your curriculum decision was made in your state capital by people who have never seen your data. Find out what is on your state’s list, when the adoption cycle reopens, and who sits on the review panel, because that is where your real leverage is.

It also helps explain why so much promising AI-in-education work never reaches a classroom, since the product has to clear a state review built for text books and curricula rather than AI tools. For founders, the state adoption cycle may now outrank the district pilot as a distribution strategy, which matches what the AI-in-education deal data has been saying for months: distribution matters more than the strength of the narrative.

McGraw Hill bought a microphone for your classroom

Shortly after acquiring Teachally, McGraw Hill acquired TeachFX on September 17, moving into what the company sizes at a $6 billion professional learning market. TeachFX records classroom audio from a phone or laptop and reports back to the teacher on the ratio of teacher talk to student talk and on questioning technique, so the feedback is aimed at the adult in the room rather than at the students. The underlying theory of change is sound, as teachers almost never receive specific, frequent feedback on their own teaching. McGraw Hill’s release cites independent studies showing math, ELA, and science gains for students whose teachers used the tool, though the release does not name those studies, so take this with a grain of salt.

For principals and district leaders, a curriculum company now owns an instrument that records your classrooms, McGraw Hill says the coaching is private to each teacher and that school leaders receive aggregated classroom insights, but three questions need answers before your next renewal: who can see the recordings, whether the data can ever reach a teacher evaluation, and whether teachers opted in or were enrolled. Coaching that teachers trust can change their practice, whereas a tool teachers suspect of evaluating them tends to make them take fewer risks while the microphone is on. For founders, incumbents are buying the teacher-facing layer rather than building it, and two acquisitions in quick succession suggests the window for selling into that layer is narrowing.

AI tutors are adopting learning-science language faster than they are producing evidence of learning

An AI tutoring startup raised $5 million in seed funding for a voice-first tutor it sells to families of students aged 13 to 18. Its announcement describes productive struggle and says the tutor adapts to each student’s learning style, although the research has not found that tailoring instruction to learning styles improves learning. It calls the product “high-dosage” tutoring, a term with a specific research definition: three or more sessions a week for at least ten weeks, with a consistent, trained tutor, during the school day, using materials that complement classroom instruction. The only evidence the release offers is a closed beta in which more than 1,000 students completed more than 1,500 hours, which averages about an hour and a half per student, and it reports no learning outcomes.

For teachers and school leaders, an emerging risk is that accurate and unsupported learning-science claims now arrive in the same pitch, so a buyer scanning for the right vocabulary — or a quick AI review — will find it and may stop looking. Before you accept a claim, ask for the learning outcomes and data behind it and check whether borrowed terms like “high-dosage” meet their research definition.

For founders who build on learning science, shared vocabulary no longer sets you apart; evidence does. A seed round closed this month on usage hours alone, so a company that publishes learning outcomes against a comparison group, like the Boston University team’s preregistered design, is competing on ground most of the category hasn’t reached.

🔮 The call

Every week I make one prediction and I’m curious about your take. Reply or comment below. Here’s this week’s:

By June 30, 2027, at least one state education agency or top-25 US school district will publish AI tool evaluation criteria that require a learning-outcome measure as a condition of approval, not only privacy, safety, and usage terms.

At least 33 states plus Puerto Rico have published AI guidance for schools. Maryland’s AI Tool Evaluation Rubric, for example, scores a tool’s instructional and cognitive presence and asks for pilots with clear success metrics, but it does not require evidence that students learned more. The Boston University trial hands an agency both the argument for an outcome requirement and a measure it could actually use, while the McGraw Hill and AI-tutoring deals supply the urgency. The counter-case is enforcement, because writing an outcome requirement into procurement means somebody has to verify it, and it is not clear any state agency is staffed to do that, so outcomes may well stay in guidance language instead of becoming a condition of approval. If the call lands, a district’s approved list gets shorter, and what survives on it will have had to show evidence of learning rather than evidence of use.

📚 Research & policy on my radar

The Science of Reading Act passed the House – sponsored by Representatives Erin Houchin, John Mannion, and Kevin Kiley, the bill directs the $194 million Comprehensive Literacy State Development program to prioritize applicants using evidence-based reading instruction, and bars three-cueing in instruction supported by those grants. It passed with bipartisan support, and a Senate companion, the READ Act, is still pending.

For literacy leads, federal money is now conditioned on a named instructional method, so it is worth checking whether your current program and any AI reading tool you use would survive that test. Congress has also now shown it will write an evidence standard directly into a funding stream, which is the same mechanism that could eventually be applied to AI tools.

What fills the freed capacity decides whether offloading helps – a 2026 Frontiers in Psychology study reports that cognitive offloading can be positively associated with cognitive self-efficacy, which in turn predicts higher critical thinking, task persistence, and learning depth. These are associations drawn from self-reported survey measures rather than experimental effects, so they describe which students tend to benefit rather than establishing that offloading causes the benefit, and the wider literature keeps landing in the same place: cognitive offloading helps or harms depending on what a student does with the recovered effort.

For teachers, that shifts the assignment design question away from whether students may use AI and toward what harder thinking the saved time is spent on. The condition is what carries the finding, because freed capacity only pays off when you deliberately spend it on something more demanding; left unspent, it amounts to less practice than the student would otherwise have had.

Which of these is actually landing where you are? That’s The Syllabus for this week. Lots to dig into!

If this was useful:
📬 Subscribe on Substack – The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take.
🔗 Connect with me on LinkedIn – I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space:
I advise school districts building AI frameworks and edtech founders navigating the K12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this:
Forward it. Re-stack it. Share it on LinkedIn. The best readers I’ll ever have are probably in your network!

Hi, I’m Angela Chen! I’m an edtech founder, Stanford instructor, and AI in education strategist. Every week, I read across the full AI-in-education landscape: research, classroom practice, startup and funding signals, policy, and market intelligence. In The Syllabus, I’m sharing the signals I actually think matter, synthesized across everything I read that week!

If there’s something here worth going deeper on, comment below or message me on LinkedIn and tell me which signal – I’m always interested in what’s actually landing for people in the field.

This week put a blunt question in front of every educator and edtech practitioner reading this: does the software we buy actually get used? A hard new number says mostly no. Underneath that, investors are quietly redrawing what even counts as “edtech,” which will shape the tools that reach your classrooms next year.

⚡ What to pay attention to right now

The most honest number in edtech this year: tool engagement is ~5%

Rick Hess and Michael Horn put a figure on edtech’s failure: engagement with most tools sits around 5%, the slice of motivated students Laurence Holt describes as the ones who use a product as designed. Products that shine in a pilot disappoint after adoption because they’re built to land the deal, not to hold a kid’s attention long enough for learning to stick. Every district and independent school weighing an AI tutor right now is being shown pilot data drawn from that motivated 5%. Both halves matter pedagogically: a tool the other 95% never open can’t teach them anything and a tool they do open that only rewards time-on-task isn’t teaching either. Engagement gets a student to the work, but it doesn’t guarantee they learn from it.

For educators: ask founders and sales teams two questions instead of one — is the design built on how students actually learn, with retrieval, spaced practice, productive struggle, and feedback that moves a student forward, and will the other 95% come back a second time? For anyone on a purchasing committee: ask for weekly engagement-retention rates alongside the accuracy benchmarks and model details you’re already poking at. A tool no one uses is the most expensive pilot you’ll ever run.

A $6M seed round aimed straight at that 5% problem

Wild Zebra raised $6M for a math and reading tutor that asks students questions instead of handing them answers, and it now reaches tens of thousands of kids in grades 2-9. The design rests on durable learning science: students remember what they generate themselves, so a tool that makes them retrieve and reason keeps the cognitive work with the kid rather than offloading it to the model. Educators know this catch well — struggle only helps learning when it’s productive, scaffolded with hints and feedback that hold a student in the zone between bored and lost. The real test of a question-asking tutor is the quality of that scaffolding, not the fact that it withholds the answer. For the founders reading: the design lesson is that “we give the answer faster” has stopped being a moat, and “we keep the thinking with the student” might be one.

The labs are moving past teacher tools into faculty workflows

OpenAI published what professors and researchers built in its Codex Summer Studio, a four-week program that teaches faculty to hand real work to coding agents: interactive prototypes, lab automation, cleanup of messy datasets — all no coding required. The tool land grab of the past year, ChatGPT for Teachers and Claude for Teachers, was about lesson planning. This aims at how higher-ed research and teaching labor gets done. For higher-ed faculty and deans: the near-term curriculum question shifts from whether students may use AI to what you teach once agents do the technical lift. For higher-ed teams building their own tools, the competitor for agentic faculty workflows is now the frontier lab itself, offered for free.

Why a European VC rebrand should still land on a school leader’s desk

Brighteye Ventures announced a $72M first close on Fund III and rebuilt its thesis around what it calls “HumanOS”: the layer that helps people learn, work, and adapt, not only students in classrooms. The fund will back productivity platforms, talent software, and labor marketplaces alongside learning tools, a bet that classic edtech has become too small a category to fund on its own after global edtech funding fell to about $1.8B in the first half of this year, down from $2.5B a year earlier. Brighteye isn’t alone: earlier this year, Reach Capital closed a $265M Fund V and rebranded from an edtech fund to an “institutional AI” fund spanning learning, health, and work.

Two of the most established names in edtech investing reframing is a real signal, more than either change alone. For school and district leaders: expect more of your vendors to reposition as workforce and “human capital” plays, which changes what their product is optimized to do. For founders: a pure-play “K12 SaaS” pitch is getting harder to raise on, and the money now rewards products that sit where learning meets work.

🔮 The call

Every week I make one prediction and I’m curious about your take. Reply or comment below. Here’s this week’s:

The “learning-and-work” reframe stops being a two-fund story: by the end of Q1 2027, at least one more established edtech-focused fund beyond Reach and Brighteye publicly repositions around workforce or “human capital” (a rename, a new fund, or an explicit mandate), or a top-tier generalist fund launches a dedicated learning-to-work vehicle. Reach Capital rebranded to “institutional AI” across learning, health, and work in August. Brighteye followed with its HumanOS pivot in September. Two category leaders moving together, while classic edtech funding contracts, is how a repositioning hardens into a category norm. If it holds, educators will feel it as vendors quietly changing what their tools are for.

📚 Research & policy on my radar

A state playbook for connecting school to career – Michael Horn and co-authors laid out a six-paper “Beyond Silos” playbook for integrated education-to-work systems, including apprenticeship degrees run through Craft Education, a WGU company. For district and community-college leaders: the same learning-to-work convergence showing up in VC is showing up in state policy, so expect apprenticeship-degree and CTE-integration pressure to build.

Teacher-training grants get a reprieve, not the money – A federal judge struck down the $600M cut to Teacher Quality Partnership and SEED grants as unlawful and “arbitrary and capricious,” though the ruling does not restore the funds. For district PD planners: the pipelines survive legally, but don’t budget on the federal money arriving.

The techlash keeps widening – Parent advocates told EdWeek that banning AI alone won’t settle the broader techlash over data privacy, screen time, and YouTube. This is the third time in six weeks the district-restriction story (NYC, LAUSD, California’s child-safety package) has widened rather than resolved. For school leaders: an AI ban won’t buy peace with parents worried about screens as a whole.

That’s The Syllabus for this week. Lots to dig into!

If this was useful: 📬 Subscribe on Substack – The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take. 🔗 Connect with me on LinkedIn – I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space: I advise school districts building AI frameworks and edtech founders navigating the K12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this: Forward it. Re-stack it. Share it on LinkedIn. The best readers I’ll ever have are probably in your network!

I read it twice, from two perspectives. I teach at Stanford Continuing Studies, where I write the course-level AI policy MIT is asking its instructors to write. And I spend the rest of my week inside K-12 schools as an AI strategist, where I write the same kind of policy for teachers of eleven to eighteen-year-olds and sit in the meetings where schools decide what AI to buy.

Twenty-five of the 26 recommendations apply to K-12 schools. The one that does not transfer is the one the other 25 depend on: know which AI system your students are talking to and don’t tie yourself to a single company. Schools “rent” theirs, and most of the companies they partner with have never said which model is inside.

The rest of this is the mapping and it’s long on purpose — it’s meant to be the document a teacher, a superintendent, a school board member or a reporter can open, read, and return to on their own time.

If reading this sparks any thoughts, I’d love to connect!

The short version

• MIT found students using AI constantly, teachers unsure what to allow, and the trust between them wearing thin. K-12 has the same picture, with younger students and higher numbers.

• The biggest lesson is every AI rule needs a reason a student can understand. Most school policies have the rule, but skip the reason.

• The one recommendation schools cannot follow is about the technology itself. That is the second half of this piece.

• The full line-by-line mapping of all 26 recommendations, with sources, is in a separate companion document. Message me for access.

What the report is

In January 2026, MIT asked a committee of professors, students and staff to look at AI use on campus and propose a policy. The committee surveyed more than 1,600 people, held listening sessions, and reviewed what other universities were doing.

They came back with something bigger than a policy — a headline covered across international media — the report shared that AI can now produce a credible answer to almost any written assignment at MIT. Teachers feel compelled to develop alternative grading, and forced to police AI use. The committee calls this a fraying of the “social contract” between teachers and students, “forcing a rethink of college itself.”

MIT’s president called it a watershed moment. It is one for K-12 schools too.

What this means for you

If you are a teacher

• Write the reason, not just the rule. Share one sentence per major assignment on whether AI is allowed and why. Students at MIT and in K-12 surveys say the reasoning is what they are missing.

• Say what you use. If AI helped build the rubric or draft feedback, tell students. It changes the trust equation.

• Don’t run the essay through a detector. Ask for the edit history or have a five-minute conversation about the work. Both are more reliable and students don’t feel wrongly accused.

• Move some work into the room. Consider alternative assessment types, such as one oral check or one hand-written draft per unit. Don’t move every assessment back into the classroom, MIT warns against that.

• Keep the human help visible. Offer student group learning or tutors before you point them to a tool. The chatbot will always be there.

• Ask what’s in the room. If your students use a classroom AI product, ask your technology lead which model it runs on.

If you are a district or school leader

• Audit your policy for reasons. If it says what is allowed and not why, it is half a policy. Make sure the reasoning behind your policies is visible.

• Level it by age and publish the logic in one place. Different rules by grade read as inconsistency until the school explains them. Share your thinking.

• Retire detector reliance in writing. No student should face consequences on a detector score alone. The best academic integrity approach is restorative and conversation based.

• Decide who reads the logs. If your monitoring software captures AI conversations, students and parents should know, and there should be a written rule for what triggers a human response.

• Ask vendors two questions and put the answers in the contract. What model is your tool running on? Will you tell us when it changes?

• Check the defaults. Gemini in Classroom turned on for all ages in August. Decide on purpose, by grade, and tell teachers before students find it.

• Fund teacher time, not just tools. MIT’s most expensive recommendation is people who help teachers redesign. The K-12 version is protected experimentation and PD time.

If you are a parent

The teacher and leader lists above are about changing things. As a parent, your role is to understanding them, and to give grace. The schools are working this out in real time, and the most useful parent is the one who is curious before critical.

• Expect the rules to differ by grade and by class. A seventh grader and a junior should not have the same policy, and a math teacher and an English teacher may reasonably disagree. That is a sign the school is thinking, not a sign it is confused.

• Ask your child what AI they use, and learn from them. Which tools, for what, and whether it gives them answers or asks them questions. Most teens have never had this conversation with an adult.

• Take the companion question seriously. Seven in ten teens have tried an AI companion, and a third of those conversations happen on school devices. Ask your child about it the way you would ask about a new friend, not a new tool.

• Keep human help easy. Kids are choosing chatbots over friends and teachers because it feels safer. A home where asking for help is easy, and being wrong is fine, is the counterweight nobody can automate.

• Bring understanding to the school, not pressure. If something in the policy puzzles you, ask what the teacher is trying to protect. You will usually get a good answer, and the teacher will remember that you asked instead of complained.

The recommendation that doesn’t transfer

Three recommendations near the end of the report got no press at all. They are about the platform.

MIT says it should keep running a “model agnostic” AI platform “that does not tie our community to one model provider.” It urges the institute “not to commit to any single commercial AI ecosystem.” It wants to own the privacy layer so that what students type is never training data. And it names the reason: some students can pay $200 a month for the best model and their classmates cannot, and “this imbalance has the potential to create enormous disparities in course performance.”

MIT’s reason is access. On a campus of adults, the institution’s job is to make sure the best model is not a function of who can pay, so it buys access for everyone and stays free to switch providers. So, it runs Parley, an MIT-hosted platform where every student, faculty member and staff member can use a range of commercial and open models through one interface, with up to $30 a month in credits paid by the Institute, and with conversations kept private under MIT's own security policies.

Now put that to a typical school. The access problem exists, but it lives outside the building: a student with a paid account at home and a student without one. Inside the building, a school does not hand students a raw model and a monthly budget. It provides the tools, either directly from an AI lab or through a classroom product, and those tools are already chosen and paid for. So the question shifts — it’s no longer who can afford the “best” model, it’s which model the school has put in front of its students, how it behaves with children, and whether the school knows.

Google made Gemini available inside Google Classroom to K-12 students of all ages on 10 August, on by default. Some principals found out from their students. OpenAI gives ChatGPT free to verified K-12 educators until 2028. Anthropic offers K-12 educators free Claude Pro, and schools a year of free access. Beyond the AI labs, most children meet AI through a classroom product – a subject-specific tutor, a writing assistant, a lesson generator – built by an edtech company that runs on one or more foundational models.

I reviewed 50 of those companies earlier this year. Roughly three in four have never publicly said which model their product runs on. Where the answer exists, it was usually the model company that published it, as a marketing case study. No procurement checklist asks the question. No framework requires a vendor to say when the model underneath changes, and it does change.

Does the model matter, or only the product?

What a school buys is the product built around the model – the age gate, the content filter, the teacher dashboard, and above all the method of engagement, whether the tool hands a student the answer or makes her work for it. That layer is real and the school can inspect it.

The best K-12 evidence says the method of engagement matters a great deal. In a study of about a thousand high school math students in Turkey, researchers at Wharton gave one group GPT-4 that answered questions and another group GPT-4 that only gave hints. The first group did 48% better on practice and 17% worse on the exam when the AI was taken away. The hints group did far better on practice and about the same as students with no AI on the exam. Same model but different method resulted in opposite result for learning.

But the model underneath is not neutral. It carries its own behavioral commitments: how readily it agrees with a student, how it handles a request to just give the answer, what it does when a child brings it something personal, what it was trained to refuse. Common Sense Media rates the major models differently on risk to teens, and those ratings attach to the model, not to the product built on it. A vendor can change the model under a product without changing anything the school can see, and no framework requires the vendor to say so. The method of engagement is the school’s to evaluate. The model’s disposition comes with the model, and a school that does not know which one it has cannot evaluate it, cannot check its rating, and will not know when it changes.

I wrote about this at length in I’m in the Middle of K-12 AI Tool Selection. This Is What Nobody Asks. The short version: procurement treats the model as a data-privacy question. It is also a safety and values question, and it goes unasked.

To be clear about the claim. This is an argument about understanding. MIT’s reasoning for running its own platform is that the institution needs to know, and be able to change, what its students are talking to. You cannot govern what you cannot name, and most schools cannot name it.

What a school can do

Each school district, or independent school, can’t be expected to build and run a Parly-equivalent platform, but it can keep the judgment. We can borrow the thinking beneath MIT’s choice.

Knowing which model sits under each tool, and keeping the option to change it, is the K-12 version of Parley, and it costs nothing but questions. What transfers is a short list of questions for anyone who sells AI to a school:

1. Which AI model does this product run on today?

2. Will you tell us in writing when that changes during our contract?

3. Does it give students answers or hints, and can we choose?

4. What happens to a student’s conversations, and who can read them?

5. What is the evidence it helps students learn once the AI is taken away, not only while they use it?

Ask the first two before signing and again at renewal. A vendor who cannot answer them has told you what you need to know.

This is why the platform recommendation sits under the other 25. MIT holds its model relationships itself, so it can set the terms. A school holds them through its vendors, so the vendor relationship is where the terms get set, or don’t. Every rule with a reason, every teacher disclosure, every redesigned assessment assumes the school knows what tool it is governing. The vendors worth partnering with already know this. Treat the five questions as the start of the partnership, not a test at the end of it, and the good ones will answer before you finish asking.

Why this “translation”?

MIT’s report is a great document. It is also easy to skip. With the volume of what lands every month in AI and education, the natural move is to put on blinders: I work in K-12, higher ed is out of my lane, and the reverse. But the distance between a high school senior and a college freshman is one summer. What MIT found on its campus is what a high school will find a year earlier, and what it recommends is mostly what a school could do now.

So I read it line by line, all 26 recommendations, and wrote down what each one means in a school. The full mapping is in a companion document. Message me for access.

I plan to do the same for each major AI-in-education report as it lands, so the translation exists the week school leaders need it, not a semester later.

Hi, I’m Angela Chen! I’m an edtech founder, Stanford instructor, and AI in education strategist. Every week, I read across the full AI-in-education landscape: research, classroom practice, startup and funding signals, policy, and market intelligence. In The Syllabus, I’m sharing the signals I actually think matter, synthesized across everything I read that week!

If there’s something here worth going deeper on, comment below or message me on LinkedIn and tell me which signal – I’m always interested in what’s actually landing for people in the field.

Backpacks are packed, the group chats are chaotic again, and school is very much back in session — and so are we! Consider the weekly cadence officially back on, and I’m thrilled you’re here for it.

This week, the money and the rules tightened at the same time. Incumbents went shopping, a marquee name sold for pennies on its peak, districts started building the tools they used to buy, and California turned two weeks of district AI bans into statewide law. The land grab is already consolidating.

⚡ What to pay attention to right now

Incumbents are buying AI capability instead of building it — and K12 was the target this week

Three acquisitions landed at once. McGraw Hill bought Teachally, the tool teachers use to generate, translate, and differentiate standards-aligned lessons, and plans to fold that into its curriculum. Blackboard made its first acquisition since bankruptcy, buying Cursive Technology, which verifies authorship by recording how a draft was written rather than scanning the finished text. And Lemnis, the nonprofit endowed by the sale of assessment-maker NWEA, acquired Wayfinder (SEL curriculum, 1.5M students) — its first in K12.

If you’re a founder building a single AI feature for teachers, this is your exit map: your buyer is a publisher or an LMS, not an IPO. If you’re a district, the tool you pilot this fall may belong to your incumbent vendor by spring, which is a reason to ask about roadmap and data ownership before you sign.

Districts are starting to build the tools they used to buy

EdWeek Market Brief reported that vibe coding is reviving the buy-versus-build question inside district IT. One Washington district expects to save up to $250K in canceled edtech contracts for 2026-27 by building replacements in-house. This is not districts rebuilding their entire student information system. It’s the smaller tools — a form, a dashboard, a scheduler, a parent-notification workflow — that capable staff can now assemble in an afternoon.

For founders, the diagnostic is uncomfortable and useful: if your product is a wrapper around a workflow, your competition is now a district employee with Claude and an afternoon. For district leads, the risk moved too: who maintains the tool when that staffer leaves, and who checks it for data protection and accuracy? Who governs the cybersecurity concerns?

The detection era in academic integrity is ending, and authorship is the bet replacing it

Cursive selling to Blackboard is the second academic-integrity company to find a strategic home this summer, after GPTZero’s sale to Superhuman. The through-line matters more than either deal — the market is moving off “did a machine write this?” and toward “can the student show how they wrote it.” Cursive records keystrokes, pace, and revisions and lets an instructor replay the process. Blue books are back at colleges. The 74 argues AI has exposed a motivation problem schools were already carrying, and the fixes people name are process-based grading and oral exams.

For faculty and teachers, the design question is no longer which detector to trust, it’s how to build assignments where the process is the evidence. That’s a demand for curriculum change, not a procurement question to answer.

California turned the district AI bans into statewide law

Two weeks ago, the story was NYC and LAUSD restricting student AI use. Now that has moved state-level. On September 10, Newsom signed a package of child-safety laws, including time limits on how long minors can talk to chatbots, required crisis protocols and parent notifications, and a ban on addictive feeds for users under 16. The district moratoria were policy; this is legislation with a right of action in the largest state in the country.

For any edtech product with a conversational interface aimed at minors, the safety features that were design choices last month are now legal requirements in California, and California tends to set the floor everyone else builds to. If you sell into schools, compliance review belongs on this quarter’s roadmap, not next year’s.

🔮 The call

Every week I make one prediction specific enough to be wrong. Here’s this week’s:

At least one more independent academic-integrity company will be acquired by an LMS, assessment company, or major publisher by March 31, 2027.

Two of them sold this summer (Cursive to Blackboard, GPTZero to Superhuman), and detection-only models are losing ground to authorship and process approaches. That squeezes the standalone players toward a strategic sale rather than a fight they’re set up to lose.

📚 Research & policy on my radar

Anthropic’s CEO called for slowing frontier AI, and OpenAI agreed – We Must Pace the Frontier

Dario Amodei published a 3,800-word essay arguing labs should slow the rate of capability gains, and Sam Altman said OpenAI would match Anthropic’s first commitment. The two leading labs just told the world their own technology may be too dangerous to build quickly. That doesn’t change what a district’s AI tools do, but it hands every cautious board member a quote and it ends the era of “the experts say it’s fine” as a district’s AI strategy. If your case for adoption rested on the vendors’ confidence, you need a new case this week: one built on your own use cases, your own guardrails, and your own honest read of what you’re trading off.

Why college could matter more, not less – Michael Horn

Horn’s case: as AI commoditizes content delivery, the parts of college that don’t compress — relationships, research, experiential learning, the networks that actually produce jobs — become the point. Worth reading against the “skip college, the AI will teach you” narrative, especially for anyone advising students.

The APA wants higher edtech standards, not screen bans – Chalkbeat

As bans multiply, the American Psychological Association is arguing for evidence standards on ed-tech products instead of blanket prohibition. It’s the clearest institutional voice for the nuance camp, and a useful counterweight if your district is deciding between a ban and a framework.

Reply and tell me which of these you want me to go deeper on next week. The consolidation story and the vibe-coding one are both moving fast, and I’d rather chase the one you’re actually facing.

That’s The Syllabus for this week. Lots to dig into!

If this was useful: 📬 Subscribe on Substack – The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take.

🔗 Connect with me on LinkedIn – I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space: I advise school districts building AI frameworks and edtech founders navigating the K12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this: Forward it. Re-stack it. Share it on LinkedIn. The best readers I’ll ever have are probably in someone else’s inbox right now.

Hi, I’m Angela Chen! I’m an edtech founder, Stanford instructor, and AI in education advisor. Every week, I read across the full AI-in-education landscape: research, classroom practice, startup and funding signals, policy, and market intelligence. In The Syllabus, I’m sharing the signals I actually think matter, synthesized across everything I read that week!

If there’s something here worth going deeper on, comment below or message me on LinkedIn and tell me which signal – I’m always interested in what’s actually landing for people in the field.

This week, the through-line is routing: who AI capability is being built around, and who it’s being built without. Two pieces working their way through the field right now sit at the center of it – the EU and OECD’s shared definition of what AI literacy should mean in every classroom and a Stanford study showing AI feedback tools quietly routing students toward different futures based on race and gender. Add a workforce pipeline reorganizing around employers and a federal Education Department outsourcing its core functions and the picture is clear: the infrastructure decisions are getting made right now, the question is whether schools are at the table or being designed around.

⚡ What to pay attention to right now

The EU and OECD’s AI literacy framework is the most consequential standard-setting move of the year and it’s the one I keep coming back to

The European Commission and OECD’s AI Literacy Framework for primary and secondary education (Empowering Learners for the Age of AI) was built from input by nearly 1,000 educators and researchers. It defines four domains – Engage with AI, Create with AI, Manage AI, and Design AI – across 22 competencies spanning technical knowledge, critical thinking, ethics, and human-centered design. This matters far beyond Europe as the framework becomes the basis for the first OECD PISA assessment of AI literacy in 2029. Once PISA measures something, every education system that benchmarks against PISA starts teaching to it. Whoever defines the rubric defines the curriculum, and the EU and OECD moved first while the US Department of Education is mid-reorg.

For K12 administrators and curriculum leads, this is the framework to map your AI-literacy planning against now because it’s about to become the global reference point, and aligning early is cheaper than retrofitting in 2028.

For edtech founders, “AI literacy” now has 22 concrete competencies you can build and sell against. The “Design AI” and “Manage AI” domains in particular signal that the market wants more than prompt-writing tutorials.

The workforce pipeline is being rebuilt around employers and schools are scrambling to wire themselves back in

Multiple briefings this week (including the World Future Skills Index 2027) point in the same direction: employment is being embedded directly into credentials, and corporations are now skilling at population scale. The model is growing fastest where there are acute labor shortages, clear occupational standards, and employers willing to absorb training costs. For higher ed and CTE leaders, this is an early warning that the credentialing center of gravity is shifting toward employer-delivered, work-integrated training, and the schools that stay relevant will be the ones wiring employer pathways into the degree now, not competing with them later.

The U.S. Department of Education has outsourced many of its core functions

With special education moving to HHS and civil rights enforcement moving to DOJ, the Trump administration has now redistributed the majority of the Education Department’s central functions to other agencies. Whatever your read on the politics, the operational reality for districts is that the agency you used to call for guidance, enforcement, and money may no longer be the one holding the file. For district and university procurement and compliance leads, the action item this week is mapping which federal counterpart now owns each function you depend on – IDEA compliance, Title VI complaints, grant administration – because the risk of a dropped handoff during a reorg this size falls on you, not on Washington. This is also the policy backdrop founders should price in: federal edtech procurement and grant pathways are going to be messier and slower for at least a cycle. Read EdWeek’s breakdown.

Digital Promise is building the public-good data layer for K12 AI and treating it as infrastructure, not product

Digital Promise’s Action Report this week surfaced its K-12 AI Infrastructure Platform – a hub cataloging 1,500+ openly available datasets, benchmarks, and curriculum resources, backed by a four-year, $26M grant program funding open-license public goods. This counteracts bias in AI tools – tools encode bad assumptions when the underlying education data and benchmarks are thin and proprietary. Whoever builds the open benchmarks can decide what “good” AI in education means. For edtech founders, this is worth watching closely – open datasets and shared benchmarks lower your build cost and become the standard you’ll be measured against, so engaging early is both cheaper and strategically smart. For district leaders, it’s a signal that “responsible AI” is finally getting an infrastructure layer you can point procurement at, rather than a values statement.

Agentic AI is moving into the shared workspace while the “AI as colleague” frame is getting pushback the same week

Two signals collided this week. On the product side, Ben’s Bites covered Anthropic’s Claude Tag (a shared agent you mention in Slack like a teammate) and Notion’s new developer platform that lets external agents work from shared docs and task boards – agents are moving from individual tools to shared team infrastructure. On the research side, MIT Technology Review’s The Algorithm ran “AI agents are not your coworkers,” arguing the colleague metaphor obscures how these systems fail. This tension matters because schools and universities are about to be sold “AI teammate” products built on this framing. For higher ed faculty and academic-tech leads, the curriculum-design question shifts: if agents become shared infrastructure students collaborate with, the skill to teach isn’t prompting – it’s supervising, verifying, and knowing when the agent is confidently wrong. That’s a different syllabus than “how to use ChatGPT.”

📚 Research & policy on my radar

AI in math teaching cuts both ways – the strongest evidence base, and the clearest proof of harm – [Edtech Insiders]

The evidence for AI-assisted math learning is among the strongest in the field, but it also delivers the cleanest warning: when AI solves the problem, the student doesn’t learn. The practical implication for K12 math departments is that AI belongs in the feedback and practice loop, not the answer loop. Any tool that surfaces solutions on demand needs friction built in before it reaches students.

Congress moves to convene a new National Reading Panel – [EdWeek]

A bill in Congress would mandate a successor to the landmark 2000 reading-science report. For K12 literacy leaders, this is the moment to watch whether AI-assisted reading instruction gets folded into the evidence base or gets defined out of it before the research catches up.

14% of the global workforce at displacement risk by 2030 – [QS World Future Skills Index 2027]

Drawing on McKinsey and HolonIQ data, the index finds AI augmenting more than replacing in advanced economies, with displacement concentrated in administrative, data-entry, and routine roles. For K12 leaders. the skills your schools should prioritize are the human-judgment and critical-thinking gaps teachers keep flagging, not tool-specific training that ages out in a year.

This is the line I drew in my latest piece: AI literacy builds skills, but AI civics builds judgment – the capacity to understand the system, question it, and decide what's worth making. Fluency in a tool ages out when the next model ships. Judgment is the thing that lasts, and it's the harder half to teach. Give it a read! I’d love to hear your thoughts.

What’s the signal you want me to go deeper on next week? Reply or comment – the workforce-routing story and the responsible-AI-infrastructure story are the two I’m most tempted to pull a longer thread on, and I’d rather follow what’s actually landing for you.

That’s The Syllabus for this week. Lots to dig into!

If this was useful: 📬 Subscribe on Substack – The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take. 🔗 Connect with me on LinkedIn – I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space: I advise school districts building AI frameworks and edtech founders navigating the K12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this: Forward it. Re-stack it. Share it on Linkedin. The best readers I’ll ever have are probably in someone else’s inbox right now.

Reporting on a Newark teacher’s class, Natasha Singer reframed the whole situation. We typically focus on “AI literacy” — learning to use the tools — but Singer argued what that teacher was really doing was closer to civics: teaching students to interrogate a technology’s claim on them, a questioning skill that transfers to social media, quantum, and whatever technologies come next. She called it “AI civics” and it’s the best two-word summary of the work I’ve seen in mainstream coverage.

The NYT episode mostly focused on one piece of it: critical interrogation. This piece usually gets the most airtime and is what current AI curriculum teaches best. But from years of advising schools, the questioning skill is only one of four things AI civics has to do and the other three are where I watch things break — the students who can critique a chatbot but have no idea who built it or why, the agency that erodes without anyone noticing, the adults who never decided anything on purpose.

This whole idea comes with a catch the episode doesn’t name — if AI civics means teaching kids to question the companies behind the technology, it can’t be authored by those same companies.

What AI civics should be

Strip it down and AI civics has four parts. They build outward from what a single student understands, to how she acts, to how she questions, to what the adults around her owe her.

Understanding the system. Not just how to prompt, but how AI works as a structure of power: who builds these models, what their incentives are, where the training data comes from, who’s downstream of whom. Most “literacy” curricula skip this but it’s critical. A student who understands the system will know the chatbot in their reading app was built by a company with a business model — and that the business model shapes what the tool nudges them to do.

Agency. This is the thread I’ve spent the most time on and where I’d push the definition hardest. The Newark teacher is right that we need kids driving instead of riding along, but driving a tool is the smallest version of it. It’s more useful to develop an entrepreneurial stance toward your own life. When I teach teenagers, my lesson is never “prompt better.” It’s that you decide what’s worth creating. AI just helps you make it faster, but the judgment stays yours. That level of agency takes years to build and it’s the first thing to go when the model underneath is sycophantic — built to please rather than push. A kid who learns on a model that always agrees is quietly taught to expect agreement, and that’s the opposite of the productive failure that builds judgment and resilience. You can teach drivership all you want, but if the tool is built to tell a student she’s right, she loses anyway. And the younger she is, the harder that is to undo.

Critical interrogation. The field has already converged on critical thinking as essential to AI in schools. It’s where the energy and the good curriculum already are. Organizations like aiEDU and InnovateEDU have built and taught this for years, and they’re great at it. So when I argue something’s missing from AI literacy, this isn’t it. The gap is everywhere around it.

Collective responsibility. Civics has always been collective, and AI civics is no different. The heaviest responsibility belongs to the adults who decide how AI enters a school, including how it’s taught and which tools get put in front of children. Those are civic acts and they shape what’s possible for a student long before that student is ever asked to think critically about a chatbot. The encouraging news is that this layer is finally being built: TeachAI’s guidance toolkit, CoSN’s readiness work with school-system associations, and the EdSAFE AI Alliance’s Policy Labs are all helping districts and leaders make these decisions deliberately rather than by default. The work exists, but what’s missing is the recognition that this is AI civics — not a back-office prerequisite to it. You can’t ask a teenager to interrogate a technology the adults around her have agreed not to question. The institutions have to model the thing they’re assigning.

The episode framed it as one teacher’s classroom win. I think it’s the only investment in AI education that doesn’t expire when the next model ships.

The catch

Back to the problem I flagged at the top, because it’s the part that keeps me up. If all four of those parts add up to teaching kids to interrogate the companies behind the technology, the curriculum can’t be written by those companies. And right now, that’s exactly who’s volunteering for the job.

Everyone wants to teach kids AI but that generosity isn’t neutral. When a company defines AI literacy, it defines it as fluency in its own product — lessons that teach a student to use that tool well, question everyone else’s, and never quite turn the same questions on the company that wrote the lesson. That’s just how brand loyalty works: we don’t tend to interrogate the hand that feeds us, and a student raised on one company’s tool learns to reach for it the way she reaches for any default. You don’t learn to hold an institution accountable from a curriculum it paid for.

This is two different jobs and large companies are very good at one of them. Google’s partnership with ISTE+ASCD will reach six million educators. Nothing outside Big Tech can move fluency at that speed and getting teachers confident with the tools their students already use is worth doing well. When AI is already in the building, product fluency at scale is a public service, and Google is delivering it. I’m bullish on that. If a company is going to put its money and reach behind teacher training, this is close to what it should look like.

The piece a company is least positioned to deliver is the model-agnostic layer, where users learn to question every tool rather than master a particular one. That’s not a flaw in Google’s program; it’s just not the thing a single vendor is built to do. So I’m equally bullish on the AI education nonprofits, aiEDU chief among them, because that’s exactly the job they’re founded for. When the people writing the curriculum don’t sell a model, the lessons are free to push true systems understanding rather than fluency in one tool. aiEDU built its full-year curriculum with Quill.org, another nonprofit, and its progression runs from the basics to critical thinking to leading with the human advantage.

One caveat I’d flag before anyone reads this as “nonprofit good, company bad,” because it isn’t that. aiEDU takes corporate money too, including from tech and AI companies. Plenty of nonprofits in this space are funded by the same firms whose tools they’re meant to teach students to question. So the funding source alone tells you very little. The test is who holds the pen on the curriculum and whether the lessons stay free to turn the hard questions back on the people writing the checks. A nonprofit clears that bar more easily than a vendor, but it doesn’t clear it automatically.

So what would the version worth building look like?

It should look like curriculum written by educators and nonprofits, with vendor money walled off from editorial control the way a newsroom takes ad revenue without letting advertisers write the stories. Lessons that treat specific companies, funders included, as objects of study instead of positive case studies of sponsors. Training that builds the questioning first and the tool fluency second, on purpose, in that order. And materials that don't assume any one model, so a school doesn't get quietly locked into a stack just because that's what the free training taught.

And where does that leave schools?

If you lead a school or a district, the real question isn’t the one you’ve been handed: the debate keeps asking whether to allow AI at all, but the question that matters more is who you’re letting define AI for your students.

And the younger the students, the more this matters. This is the part nobody has figured out yet, including me. Everything I’ve described assumes a student who can already read a source and form an argument. For my high schoolers, drivership is the right frame. For a 4th or 5th grader, the field doesn’t have a real answer for what AI civics looks like, and I’d rather say so than pretend the senior elective scales down. At that age, the literacy may belong to the adults: a teacher fluent on the student’s behalf and a curriculum built to protect the struggle instead of outsourcing it — with less student-facing AI, not more. It’s the question I’m pushing hardest on with the schools I advise right now because someone has to build the answer before a vendor ships one.

What I keep coming back to is the ownership question. Schools already know how to treat civics as a public good, something that belongs to communities. AI civics deserves the same handling. The schools that miss this will take the free curriculum, badge their teachers in someone’s product, and wonder in a few years why their students are fluent in a tool but can’t tell you who built it or what it wanted from them.

If you’re an educator, founder, or school leader working through this, especially the younger-grades version where I don’t have it figured out, I’d value a conversation.

Angela Chen is an educator, advisor, and builder working at the frontier of AI and education. She teaches entrepreneurship at Stanford, where she previously grew the Stanford Accelerator for Learning and built its Education Entrepreneurship Hub. Through her AI education studio, she advises schools on responsible AI adoption, coaches early-stage founders, and leads live “Build with AI” sessions for non-technical professionals. A former strategy consultant, edtech founder, and co-author of international research on AI’s impact on the future of learning and work, she serves on the AI Advisory Committee of University of Toronto Schools and is a Catalyst Fellow with the EDSAFE AI Alliance. She writes about the future of learning at newcurrent.substack.com.

New Current: AI in Education

By Angela Chen

Inspired by “The Battle Over A.I. in the Classroom” on @nytimes’ The Daily (June 17, 2026), reported by Natasha Singer (@natashanyt) and hosted by Rachel Abrams (@RachelAbramsNY).

Hi, I’m Angela Chen! I’m an edtech founder, Stanford instructor, and AI in education advisor. Every week, I read across the full AI-in-education landscape: research, classroom practice, startup and funding signals, policy, and market intelligence. In The Syllabus, I’m sharing the signals I actually think matter, synthesized across everything I read that week!

If there’s something here worth going deeper on, comment below or message me on LinkedIn and tell me which signal — I’m always interested in what’s actually landing for people in the field.

This week, the Anthropic shutdown story has moved from shock to consequence. What I’m watching now are the downstream effects rippling through platforms, policy, and procurement. Meanwhile, the signals that were quiet last week have gotten louder: workforce training is where the big capital is moving, and the classroom evidence on AI tutoring is more sobering than the headlines suggest.

⚡ What to pay attention to right now

The Anthropic ripple effects are now concrete and three of them change how edtech should be built

Last week, I wrote about the Fable shutdown as an acute shock. MIT Technology Review’s follow-up this week names what’s actually happening in the aftermath and it’s much more structural than I initially anticipated. Three things are accelerating simultaneously: European leaders are treating the shutdown as a mandate to build sovereign AI, Chinese open-source models are becoming significantly more attractive to any builder who doesn’t want access terminated by a White House decision, and leading cybersecurity experts are now arguing in an open letter that the shutdown may have increased US vulnerability to the exact attacks Anthropic was trying to prevent.

The business logic underneath this “safety” intervention matters as much as the safety logic. This is exactly the dependency problem I've been documenting in my K12 AI procurement research: when I reviewed 50 AI EdTech companies active in US K12, a grand majority never publicly disclosed which model they're built on. So when a model gets pulled overnight, districts don't just lose a tool – they lose something they didn't even know they were dependent on. For edtech procurement, the question isn't whether to have a vendor continuity plan. It's that you can't write one if you don't know what model is inside the product.

AI tutoring access ≠ AI tutoring use, and the gap is bigger than anyone’s admitting

A new study from my Stanford colleagues shows elementary students used AI tutoring platforms an average of just two to five minutes per week – far below the 30 minutes weekly the platform provider says is needed for measurable reading gains. Nearly half never logged on at all, even when dedicated session time was built into their schedule. What did move engagement was relational scaffolding: tutors who ran brief check-ins, built trust, and gave shout-outs for persistence increased usage 71-80% even though the structure reduced platform time per session. The equity finding deserves separate attention: among students using the platform independently, those who engaged were higher achieving and less likely to have special education services. The kids who most needed the reading practice were least likely to open the app.

This isn’t primarily a design problem or an interface problem. It’s an implementation problem and it suggests that the edtech market’s habit of reporting access and deployment as outcomes is masking a gap in actual use that human support is currently the only proven lever for closing.

Big Tech is betting on workforce training as AI infrastructure’s talent bottleneck

Meta launched America’s Workforce Academy – a $115M program covering tuition, travel, lodging, and a daily stipend for electricians, pipefitters, and welders being trained specifically to build AI data centers. In a similar vein Google pledged $50M to train 300,000+ skilled trade workers.

For K-12 educators and higher ed leaders, look at the signals closely. Meta’s AWA received 35,000 applications in the first week for fiber installation training alone, which tells you there is massive unmet demand for career-aligned, employer-backed pathways in fields that don’t require a four-year degree. That demand is sitting in your students right now. The question isn’t whether Big Tech will fund the training, they clearly will. The question is whether K12 CTE programs and community colleges are building the pipelines that feed into these programs, or whether Big Tech is going to build parallel on-ramps that route around traditional education entirely.

Stepful's $55M Series C (more in Funding below) is the clearest version of what disintermediation actually looks like: vertically integrated, debt-free, employer-placed healthcare training that has no particular reason to involve accredited education institutions at any stage. The employer goes directly to the training provider and the training provider goes directly to the student. That architecture may be the template others will copy across every high-demand trade field, and higher ed institutions that assume workforce training is still "their" lane should be paying very close attention to how fast it's being built around them.

California SB 928 is the leading edge of a legislative wave – and it will reshape higher ed product design

California SB 928, sponsored by the California Faculty Association, would require that every CSU instructor of record be a human being. The bill was triggered by the CSU Chancellor’s decision in early 2025 to partner with Google on an AI initiative without consulting faculty, staff, or students. That decision lit a fuse, and SB 928 is the result. The bill explicitly allows faculty to use AI tools — it just bars AI from replacing the instructor role entirely. If it becomes law, it sets a precedent that other state systems and accreditation bodies will watch closely, and it makes “AI as co-pilot” not just a marketing frame but a compliance architecture. For edtech builders targeting higher ed: products that position AI as replacing instruction are about to run into a wall in the CSU system and likely beyond. The design constraint is shifting from “how much can AI do?” to “how do we keep a human demonstrably in the loop?” That’s a different product question entirely.

🛠️ Tools of note

(These aren’t startup investment recommendations – they show up here because they’re connected to trends worth tracking.)

CodeAI – A platform pushing K12 CS education from syntax toward understanding AI “under the hood”

Why it’s trending: EdWeek’s profile of CodeAI’s CEO reflects a shift happening across CS education, from teaching kids to use AI tools toward teaching them how models actually work. Model literacy is the next frontier after prompt literacy.

ChatGPT Agent Templates – Lets educators and students build custom AI agents directly in ChatGPT, no code required

Why it’s trending: OpenAI’s Edu Prompt newsletter spotlights this as a classroom-ready feature. The “build your own agent” workflow is moving from developer tool to student activity. That’s a meaningful shift in what AI literacy actually means in practice.

💰 Funding & market moves

Stepful – $55M Series C (Healthcare Workforce Training / Oak HC/FT lead, Y Combinator, Citi Impact Fund, Intermountain Health) Employer-sponsored, debt-free healthcare training with direct placement into 35+ hospital systems; 32,000+ graduates to date. Stepful’s vertical integration model – instruction, simulation, assessment, and placement in one product – makes ROI legible to employers and removes financial risk for students. This is the architecture others will copy across non-healthcare trades. It’s also a direct structural challenge to the traditional degree for high-demand fields.

Cursor acquired by SpaceX – $60B all-stock deal (via TechCrunch) AI-native code editor, acquired days after SpaceX’s record-breaking IPO. The deal’s logic is less about vertical integration than it looks: SpaceX merged with xAI earlier this year and is now using Cursor to compete directly with Anthropic and OpenAI in the AI coding tools race — this is an AI capability consolidation play, not a company buying a tool it depends on. The edtech-relevant signal is the valuation: $60B for a product that crossed $1B in ARR means the market is still pricing AI-native workflow tools at extraordinary multiples. For edtech founders, the more honest read is about exit landscape concentration — the acquirers with the capital for deals at this scale are a small and getting-smaller group of AI-native conglomerates, not ed-specific buyers.

📚 Research & policy on my radar

AI changes its feedback on student writing based on race and gender – EdWeek

Research shows AI writing coaches produce meaningfully different feedback depending on demographic signals about the writer. This is a must read before any district deploys AI writing tools at scale – the equity implications are immense and serious. AI feedback systems trained on biased corpora risk encoding and amplifying existing inequities under the cover of “personalization.”

That’s The Syllabus for this week.

Lots to dig into! If this was useful:

📬 Subscribe on Substack – The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take.

🔗 Connect with me on LinkedIn – I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space: I advise school districts building AI frameworks and edtech founders navigating the K12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this: Forward it. Re-stack it. Share it on Linkedin. The best readers I’ll ever have are probably in someone else’s inbox right now.

Until next week!

Hi, I’m Angela Chen! I’m an edtech founder, Stanford instructor, and AI in education advisor. Every week, I read across the full AI-in-education landscape: research, classroom practice, startup and funding signals, policy, and market intelligence.

I’ve been doing this privately for my own work and a few people kept asking what I was tracking. So I’m sharing the signals I actually think matter, synthesized across everything I read that week!

If there’s something here worth going deeper on, comment below or DM me on LinkedIn and tell me which signal — I’m always interested in what’s actually landing for people in the field.

Two things dominated this week: a government directive that pulled Anthropic’s most powerful models offline overnight, and new data showing that students and teachers are using AI constantly while barely understanding how it works. The platform layer just got less stable, and the literacy gap just got more urgent.

⚡ What to pay attention to right now

The AI literacy gap is now a documented crisis, not a talking point

A new Common Sense Media report on 9–17 year olds puts the sharpest numbers I’ve seen on this yet. 85% of students who use AI turn to it for schoolwork — but out of those who say they know “a lot” about how AI works, only 37% knew that GenAI can’t distinguish true from false. The finding that stands out most personally is that students who struggle academically lean on AI more than their peers (56% vs. 46%), which means the kids most at risk of misinformation, overreliance, and cognitive offloading are the ones using it most. The implication for districts is direct: AI policies focused on restriction are aimed at the wrong lever. The students who most need guidance are already using it. Literacy instruction — not just acceptable use policies — is the intervention that actually reaches them.

Fable has been pulled by the government and that changes everything

The story circulating as “Anthropic shut down Fable” is more alarming than that. The U.S. government issued an emergency export control directive on June 12, ordering Anthropic to immediately suspend access to Fable 5 and Mythos 5 — its two most powerful new models — citing a national security concern around a jailbreak method. Anthropic had to disable both models for all users globally, not just foreign nationals. TechCrunch’s framing is pointed: Anthropic’s own safety warnings may have handed the government the justification.

This matters for education in two ways. First, any district or edtech builder who has woven frontier Anthropic models into curriculum or products now knows that access can disappear overnight with no warning. Second, it creates an opening: with Anthropic sidelined, Google and OpenAI are moving fast. Google has committed $1B to AI in higher education, launched AI literacy training for 6 million U.S. educators, and now has 400+ institutions in its AI for Education Accelerator. OpenAI has rolled out Codex-powered workspace agents to ChatGPT Edu and Teachers plans and launched its Education for Countries program with government partnerships. The platform race in education just got more competitive — and a lot less stable.

Microschools are a clear window into AI-native education

Michael Horn’s piece “Tomorrow Schools: 7 Microschools Offering a Window Into the Future” — based on independent study by four Harvard master’s students — is a great grounded account on what AI-native learning environments actually look like in practice. Microschools are small enough to iterate fast and unencumbered enough to go deep on AI, and their models are already influencing what districts consider possible. Horn also published this week on AI’s impact on degrees and reskilling, connecting the disruption of college athletics to the credential disruption AI is accelerating. Both pieces together are a useful provocation for anyone thinking about what K-12 should be preparing students for.

UChicago’s approach to AI decision-making is a model worth watching

Stephen Fitzpatrick’s “Skeptical, Ethical, and Ambitious” digs into what UChicago gets right about AI adoption in institutional contexts — treating it as a deliberate, ethically grounded choice rather than a mandate or a panic. Most K-12 districts and higher education ecosystems are stuck somewhere between “ban everything” and “adopt everything,” with almost nothing in between. The UChicago frame — skeptical by default, ethical by design, ambitious in aspiration — gives school leaders actual footing for the decisions they’re being asked to make right now.

The federal ed budget is in play and the “weird market” is getting weirder

Congress has kicked off deliberations for next school year with Title I and IDEA funding already under scrutiny, and Trump’s federal school choice program — confirmed this week — will allow public schools to access federal scholarship funds for the first time, reshaping where district money flows. The EdWeek Market Brief framing this as a “weird market” for vendors is accurate: buyers are uncertain, procurement cycles are elongating, and the districts with AI budgets today may not have them next year. Build for resilience, not for the assumption that edtech spending continues upward.

🛠️ Tools of note

(These aren’t investment recommendations — they show up here because they’re connected to trends worth tracking.)

AI assistant for special education (EALA / InnovateEdu) — Free training modules and an AI assistant for special educators, via the Educating All Learners Alliance

Why it’s trending: Special education is one of the highest-need, most underserved areas for AI tools in K-12. The combination of free PD with a purpose-built AI assistant signals where nonprofit-sector innovation is moving and where commercial builders haven’t gone yet.

SEE Framework Workshop — AI for Education — Structured GenAI literacy PD for educators, open enrollment June 24

Why it’s trending: Reflects growing demand for district-level PD that goes beyond one-off “how to use ChatGPT” sessions — frameworks teachers can actually apply across grade levels and subjects.

📚 Research & policy on my radar

2026 AIMS EduData Grantees — Digital Promise — Digital Promise Action Report, June 11, 2026

Digital Promise’s new cohort of AIMS EduData grantees is focused on AI adoption in literacy instruction and growing future digital leaders. This is one of the few initiatives systematically studying what AI adoption in literacy actually produces at the classroom level — these grantees will be generating real implementation data over the next 12–18 months. Worth following if you’re advising districts on reading programs.

Teacher prep programs still lagging on science of reading — EdWeek, June 13, 2026

Roughly half of teacher prep programs now train teachers in evidence-based reading instruction, unfortunately still well short of universal. The parallel to AI literacy is direct: if foundational reading pedagogy can’t make it into pre-service training consistently, AI literacy in teacher prep faces the same uphill battle. Districts counting on new teachers to arrive AI-ready will be waiting a long time.

That’s The Syllabus for this week. Lots to dig into!

If this was useful:

📬 Subscribe on Substack — The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take.

🔗 Connect with me on LinkedIn — I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space:

I advise school districts building AI frameworks and edtech founders navigating the K-12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this:

Forward it. Re-stack it. Share it on Linkedin. The best readers I’ll ever have are probably in someone else’s inbox right now.

Hi, I’m Angela Chen! I’m an edtech founder, Stanford instructor, and AI in education advisor. Every week, I read across the full AI-in-education landscape: research, classroom practice, startup and funding signals, policy, and market intelligence.

I’ve been doing this privately for my own work and a few people kept asking what I was tracking. So I’m sharing the signals I actually think matter, synthesized across everything I read that week!

If there’s something here worth going deeper on, comment below or DM me on LinkedIn and tell me which signal — I’m always interested in what’s actually landing for people in the field.

The mood this week is pressured: the first wave of AI contracts is hitting renewal and nobody has the evaluation language to justify — or push back on — what they’re spending. At the same time, the intellectual framing is shifting. Horn in front of Congress, Mollick rewriting his thesis, PE betting on human connection as AI saturates digital channels — the field’s most credible voices are all arriving at the same place: adoption without redesign isn’t a strategy, it’s a delay.

⚡ What to pay attention to right now

The university AI contract renewal crisis is here and nobody has the evaluation language for it

California State University just renewed its OpenAI contract for $39M over three years — the largest university-AI deal on record — despite a 4,000-signature faculty petition, active budget cuts, and zero published evidence that the first contract delivered on its workforce readiness thesis. CSU’s troubles are indicative of the sector-wide problem. The first generation of enterprise AI contracts is hitting renewal cycles at the exact moment that evidence standards for what these deals should produce remain undefined. EdSource flagged the gap directly: 94% of higher ed workers now use AI tools, but fewer than half know their institution’s governing policies. If you’re advising an institution right now, push for an evaluation framework before the next renewal conversation begins.

Michael Horn testified before Congress and made the case I’ve been making to advisory clients for years

Horn is the co-founder of the Clayton Christensen Institute and brought disruptive innovation theory to education — his framework shaped how a generation of education reformers and investors think about edtech. His written testimony before the House Subcommittee on Higher Education this week highlighted colleges that adopted AI without redesigning their operating models are seeing the same non-results as steam-powered factories that bolted electric motors onto old machinery. Productivity didn’t come until the factories redesigned around electricity.

He named specific pivots — live oral defenses replacing AI-written essays, AI as assessment infrastructure rather than detection adversary, embedded real-world work — and called out startups like Protagonist (AI-powered student success infrastructure) as early movers worth watching. Horn’s is the clearest public articulation I’ve seen of why the “wait and see” campus posture invites strategic failure, not prudent caution. When this argument reaches Congress, it’s going to start shaping what funding conditions look like and the direction of policy.

Ethan Mollick’s new framing is the one educators need to sit with

Mollick is a Wharton professor whose Substack One Useful Thing is probably the most widely-read practitioner resource on AI in education — his first book Co-Intelligence genuinely shaped how most educators onboarded AI in 2024-2025. His new book, Co-Existence (out October 20), marks a real shift. Co-Intelligence was premised on human-AI collaboration — prompting, editing, skepticism, with humans at the center. Co-Existence is about well…co-existing with AI that is sometimes better than you. He points to Anthropic’s own data that AI now writes 80% of their code, with each developer shipping 8x more.

The education implication is direct: if the “co-intelligence” framing shaped how most educators onboarded AI in 2024-2025, the co-existence framing needs to shape how we redesign curriculum for 2026-2027. The question isn’t “how do students use AI as a tool” anymore. It’s “what does learning look like when AI is sometimes the better practitioner.”

This connects to something Grow with Google surfaced in a recent EdTech Insiders episode: only 5% of workers currently qualify as “AI fluent” — meaning they can actually create and solve problems with AI, not just use it passively or demonstrate surface-level AI literacy. Literacy gets you reading the language. Fluency is when you start thinking in it. That’s the gap every educator is now being asked to close, and most curriculum wasn’t designed with it in mind.

The PE money betting on human connection is the contrarian read this week

Hellman & Friedman acquired Hyve Group — which owns both Bett and ASU+GSV — from Providence Equity and Searchlight Capital. H&F’s thesis is that face-to-face professional events will become more valuable as AI saturates digital channels. On the day the deal closed, they announced Bett USA, a new annual K-12 event for US educators.

The signal isn’t just market consolidation. It’s that PE is making the same bet as the smartest venture firms: a16z spent 2025 building a full in-house media and events operation — dinner series, IRL fellowships, 40,000-person city-wide tech festivals — with the explicit argument that high-trust, in-person environments become more scarce and more valuable as AI floods digital channels with synthetic noise. Lightspeed just hired its first-ever Partner & New Media Lead and has been running monthly IRL Generative Series events across SF, LA, NYC, and Europe as community infrastructure for AI builders. H&F is running the same trade in edtech: own the room where real relationships form. If you’re building products or professional development for educators, the live event layer is your distribution infrastructure — and it just got a serious upgrade.

AI pricing is moving from flat-rate to metered and schools need to think about what that means for their budgets

The AI labs are quietly shifting from subscription bundles toward usage-based pricing, especially for coding and agentic tools. Azeem Azhar’s Exponential View had the clearest breakdown this week: flat-rate bundles work when marginal costs are low, but agents change the math with power users burning billions of tokens in a month. The education implication isn’t that AI will become unaffordable. It’s that institutions that haven’t built internal governance around AI usage are about to get sticker shock at renewal time — and districts that purchased flat-rate site licenses without any usage data are flying blind. This connects directly to the CSU contract renewal story above: usage data and defined outcomes are going to be the two non-negotiables in every enterprise AI contract from here on out.

The trust breakdown between faculty and students is the least-covered story with the most downstream consequences

EDUCAUSE flagged it this week and it keeps coming up in conversations: AI is changing the trust dynamic between faculty and students in ways that policy can’t fully address. It’s not just about academic integrity. It’s that faculty who rely on written work to understand what students know are losing confidence in that signal. What Horn’s testimony names as “AI accelerating the processes and priorities currently in action” is most visible here. The institutions moving fastest on oral defenses, portfolio-based assessment, and demonstration-of-learning models aren’t just solving for cheating, they’re rebuilding the epistemic relationship between teachers and students.

💰 Funding & market moves

OpenAI launches OpenAI for Singapore (announced in May but flagging it here if you missed it) — S$300M+ government partnership anchoring its first Applied AI Lab outside the US, including workforce and education programming.

This is more than a flag in the ground. The Applied AI Lab will create 200+ Singapore-based technical roles, a Singapore chapter of the OpenAI Academy for educators, Codex for Teachers hackathons, and a Forward-Deployed Engineer training program to build local AI deployment talent. The Ministry of Education is a named partner. So this is a national education strategy, not just a tech deal.

Last week, I wrote about OpenAI going country-direct and what it means for who the edtech competition even is — Singapore now joins Armenia, Jordan, Greece, Kazakhstan, Slovakia, and Estonia in OpenAI’s Education for Countries program. The race is moving fast and it’s not just OpenAI: Google + UNICEF deployed in Brazil, India, Pakistan, and Kenya that same week. The countries building this infrastructure now are the ones US students will be competing with in 10 years and they are not waiting.

I’m working on a longer piece on the three-way divergence between OpenAI’s top-down national reach, Google’s friction-free access through Workspace, and Anthropic’s teacher-as-co-architect model — more soon.

📚 Research & policy on my radar

“4 Questions We Must Answer Before Bringing AI Into the Classroom” — EdWeek Tech Leader The four: Does it improve student learning? Does it reduce equity gaps? Does it support teacher autonomy? Does it build long-term capacity? Simple framework, but the sequencing matters — student learning first, everything else second. Good scaffolding for school leaders who need something beyond “does the vendor have a good demo.”

“The AI High School is Surprisingly Human” — NYT Opinion The first dedicated “AI high school” looks mostly like... school. Human teachers, messy learning, genuine relationships. The AI is in the infrastructure and curriculum framing, not replacing the human moments. I keep sending this to school leaders who need permission to move forward without abandoning what actually works.

That’s The Syllabus for this week. Lots to dig into!

If this was useful:

📬 Subscribe on Substack — The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take.

🔗 Connect with me on LinkedIn — I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space:

I advise school districts building AI frameworks and edtech founders navigating the K-12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this:

Forward it. Re-stack it. Share it on Linkedin. The best readers I’ll ever have are probably in someone else’s inbox right now.

Hi, I’m Angela Chen! I’m an edtech founder, Stanford instructor, and AI in education advisor. Every week, I read across the full AI-in-education landscape: research, classroom practice, startup and funding signals, policy, and market intelligence.

I’ve been doing this privately for my own work and a few people kept asking what I was tracking. So I’m sharing the signals I actually think matter, synthesized across everything I read that week, publically now!

If there’s something here worth going deeper on, comment below or DM me on LinkedIn and tell me which signal — I’m always interested in what’s actually landing for people in the field.

The mood this week is bifurcated: on one hand, AI is quietly becoming infrastructure — districts building their own tools, nations signing direct deals with OpenAI, universities positioning as innovation hubs. On the other hand, the guardrails conversation is getting louder and more organized, with teachers’ unions, researchers, and practitioners all pushing back on “AI confidence” that doesn’t translate to competence.

⚡ What to pay attention to right now

Districts are becoming their own edtech vendors and that’s a bigger deal than it sounds

An Oregon school district is projecting $200K in savings by using AI-powered “vibe coding” to build custom instructional tools in-house, according to EdWeek’s reporting from May 29. This isn’t an isolated story, it’s an early signal of a structural shift. When AI makes coding accessible enough that district IT teams can build custom solutions faster and cheaper than procuring from vendors, the traditional edtech sales model starts to crack.

For vendors, the response has to be products that are genuinely harder to replicate by district IT teams: deep integrations, compliance infrastructure, and curriculum-grade content. For districts, the question is sustainability — who maintains these tools when the initial savings wear off? This is a trend worth watching closely.

OpenAI is going country-direct and it’s reshaping who the edtech competition even is

Armenia is the latest country to sign on with OpenAI’s Education for Countries program, a direct government-to-OpenAI partnership to embed AI into national education infrastructure. This follows similar moves in other countries (Jordan, Greece, Kahzakstan, Slovakia, Singapore, and Estonia) and represents something the traditional edtech ecosystem wasn’t built to compete with: foundation model companies offering subsidized, integrated AI capacity at the national level.

The US doesn't have an equivalent counterpart — education is a state function, fragmented by district, and there's no federal body that could sign a comparable deal. The deeper competitive picture is that OpenAI isn’t the only one moving here — it’s just moving most visibly. Google is winning on access by embedding Gemini directly into Workspace for Education tools millions of schools already use: no new procurement, no friction, already there. Anthropic is taking a different approach entirely — its Teach For All partnership puts teachers as co-architects of AI tools across 63 countries rather than recipients of them.

I’m writing a longer piece on what this three-way divergence actually means for the field. The short version: OpenAI is winning on top-down reach, Google on educator access, Anthropic on pedagogical intentionality — and which of those bets pays off will depend entirely on whether outcomes data ever catches up to deployment speed.

The American Federation of Teachers just drew a hard line and it’s going to shape district procurement decisions

The American Federation of Teachers released a comprehensive AI plan this week calling for complete screen bans for pre-K through grade 2, bans on student-facing AI tools across all of elementary education, and substantial teacher training requirements as a precondition for any AI tool adoption.

Whether or not you agree with the specific policy asks, the signal is clear: labor unions are now becoming a formal stakeholder in edtech procurement decisions, not just an afterthought. Districts that want smooth implementation need union buy-in, and union buy-in increasingly means credible PD infrastructure — not pilot programs and slide decks.

New research: student AI confidence is inversely correlated with AI competence

AI for Education’s May 28 newsletter flags a new study on middle schoolers showing that students who are most confident using AI tools actually perform worst — they can’t distinguish good AI outputs from bad ones and rarely ask follow-up questions. This is a classic “confidence trap” problem and it has direct implications for how we design AI literacy curricula. Familiarity with tools is not the same as understanding of their limitations. The practical implication for K-12: AI literacy programs that focus on tool access without building critical evaluation skills may actually widen the competency gap rather than close it.

MIT/Harvard data on AI and jobs: grounding the workforce narrative

MIT Technology Review’s June 1 issue digs into the actual Bureau of Labor Statistics data and Harvard economist David Deming’s longitudinal survey tracking AI adoption across the workforce and found that AI hasn’t had large-scale labor market impact yet. About 40% of workers use generative AI, productivity gains exist but aren’t economy-shaking, and unemployment in AI-exposed occupations is actually lower than in less-exposed ones.

This matters for how we talk to students, parents, and district leaders about workforce preparation. The honest answer right now is not “AI is killing jobs” but “AI is changing the skills that matter, and we don’t have great data yet on what comes next.”

🛠️ Tools of note

(These aren’t startup investment recommendations — they show up here because they’re connected to trends worth tracking.)

AI for Education’s 45-Hour Practitioner Course — “Generative AI in Your Practice: From Knowledge to Application” — a self-paced course for educators launching June 1

Why it’s trending: This is one of the most substantial educator AI literacy infrastructure I’ve seen built to date. The move from micro-credentials to a full 45-hour course reflects that surface-level PD is no longer sufficient — both practitioners and districts are demanding depth.

EDSAFE AI Policy Essentials Course — A free, 10-module self-paced course (~3.5 hours total) built specifically for district leaders: superintendents, school board members, CTOs, and policy committees.

Why it’s trending: It’s a direct response to the governance gap the AFT plan and CoSN data both flagged this week — most districts have acceptable use policies that weren’t built for generative AI, and leaders lack practical frameworks for procurement decisions, vendor accountability, and data privacy. This is the structured infrastructure for that. Worth sharing with any district leader you’re advising.

💰 Funding & market moves

HolonIQ’s $200M Partnership for Health, Education & Economic Mobility — HolonIQ flagged a new $200M cross-sector partnership connecting health, education, and economic mobility in their May 29 ecosystem hubs report. The broader HolonIQ report on Edtech Ecosystem Hubs is worth reading: it maps how universities are showing up at every stage of edtech innovation, from R&D through commercialization to adoption — and argues they’re underutilized as infrastructure for the sector.

📚 Research & policy on my radar

“Student AI Confidence Hides Real Gaps” — AI for Education (mentioned above)

A study tracking middle schoolers found that confidence with AI tools doesn’t translate to competence and may make things worse. Students with the most positive AI attitudes performed worst, couldn’t distinguish good from bad AI outputs, and rarely asked follow-up questions. This means K-12 AI literacy programs that emphasize access and familiarity over critical evaluation are potentially counterproductive — so build for for skepticism, not enthusiasm.

“AI Isn’t Killing Jobs (Yet)” — MIT Technology Review, June 1

The data from BLS and Harvard’s longitudinal survey shows AI adoption is broad but economic impact is modest and job displacement is not yet measurable at scale. The piece argues that workforce AI preparation should focus on judgment, adaptability, and critical AI use — not fear. The skills gap is real, but it’s not the apocalypse narrative, and teaching to panic helps no one.

“Governing Generative AI in Higher Education” (new publication) — AI Edu Simplified / Substack, May 29

A new collaborative publication on AI governance frameworks for higher ed. Governance decisions flow will downstream to K-12 too and the frameworks colleges adopt now for student AI use, academic integrity, and data privacy will shape what K-12 districts face in three to five years.

That’s The Syllabus for this week. Lot’s to dig into!

If this was useful:

📬 Subscribe on Substack — The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take.

🔗 Connect with me on LinkedIn — I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space:

I advise school districts building AI frameworks and edtech founders navigating the K-12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this:

Forward it. Re-stack it. Share it on Linkedin. The best readers I’ll ever have are probably in someone else’s inbox right now.

Almost nobody in K-12 EdTech is asking which underlying AI platform a tool is built on. I found that out the hard way and I think it’s the most important question in tool selection right now, one that matters more, not less, the younger the students are.

I’ve spent the last two months embedded with a PreK-5 school building their AI constitution — acceptable use policies, data handling standards, guardrails, FERPA compliance, decision authority, steering committees. That part is done. Now we’re moving toward tool selection. In the process, I did what I do at every stage of this work: went back to the research to make sure my understanding of the landscape was current.

There is no shortage of discourse about AI in K-12. Brookings ran a year-long global task force. Common Sense Media has done foundational work on AI risk for teens. EdSurge and EdWeek both covering the space actively. What’s largely missing from this coverage is a distinction that matters enormously at the procurement level: the difference between AI tools as products and the AI platforms those products are built on.

The debate at most schools isn’t whether to use AI. That conversation is largely settled, even where it hasn’t been officially resolved. The real questions are the ones I’m actively working through: which products do you put in front of students, which ones do you give teachers for planning and feedback, and which underlying AI platforms are those tools actually built on?

That last question is nearly absent from the discourse. This piece is about why — and what it costs.

Three Layers, One Discourse Problem

The K-12 AI landscape I’m navigating operates across three distinct layers, and the public conversation conflates them in ways that make useful guidance harder to produce.

Layer 1 — Direct LLM access. Teachers using Claude, ChatGPT, or Gemini directly. This is real, it’s widespread, and it’s worth naming clearly because it often gets lost in the EdTech conversation. A significant share of teacher AI use happens here — not inside specialized tools, but through direct chat interfaces. Many of the teachers I work with use Claude and ChatGPT directly every day for lesson prep, image generation, parent communication, and feedback drafts. That’s appropriate and worth supporting well. The discourse around this layer is reasonably developed: Common Sense Media rates these platforms, researchers study them, districts write acceptable-use policies for them. In some ways, this is the most transparent layer — you know exactly what you're using.

Layer 2 — Teacher-productivity EdTech tools. Tools like MagicSchool, Brisk Teaching, Curipod, and Diffit. These wrap foundation models inside purpose-built teacher workflows — lesson generation, rubric creation, reading level differentiation, content personalization resources. This is the heaviest commercial activity in EdTech AI right now, and it’s where the discourse gap starts. A teacher using MagicSchool almost certainly doesn't know which foundation model is generating the lesson plan she's about to hand to a student. Neither does her principal. Maybe neither does the district IT director who approved the contract.

Layer 3 — Student-facing tools. Khanmigo, SchoolAI, Flint, Amira. These tools surface AI directly to students — sometimes with full conversational access, sometimes in tightly scaffolded instructional contexts. This is where platform behavioral commitments matter most — and where I’m most careful, especially for the younger end of K-12. This is also where disclosure is most urgently needed and most consistently absent. An 11-year-old using an AI reading tutor is having a conversation with a model whose identity, values, and behavioral constraints her school may never have thought to ask about.

The discourse gap is most acute at Layers 2 and 3. The conversation about foundation model safety ratings, behavioral commitments, and values posture focuses on foundation models used directly. It does not systematically transfer to the EdTech products built on top of them.

When I read a Common Sense Media assessment of a student-facing tool, I won’t learn from that document which foundation model it runs on — or whether that model’s documented safety posture actually carries through to the product in front of my students.

The AI Edtech Landscape

To gather more data, I took a look at various top AI companies in K-12 schools. There is no single authoritative public database of "most actively deployed K-12 AI companies." My list was assembled by cross-referencing EdSurge company profiles, Crunchbase funding data, TechCrunch and EdWeek coverage, Getting Smart roundups, Anthropic/OpenAI/Google customer story pages, and individual company documentation and privacy policies.

For each of these companies, I look through whether or not they’ve publicly shared which LLM platform powers their product and what their LLM of choice is. Of the “Top” 50 AI EdTech companies actively deployed in US K-12 schools, roughly 80% have made no public statement about which LLM platform powers their product.

Here is the current confirmed LLM disclosure landscape:

Built on Anthropic/Claude: MagicSchool AI (switched from OpenAI; founder Adeel Khan has stated: “When Anthropic came on the scene, the responsibility narrative and commitment to constitutional AI got us really excited” — 7M educators, 100M+ AI engagements), Amira Learning, Kira Learning, Flint.

Built on OpenAI/GPT: Khanmigo, Quizlet, SchoolAI

Built on Anthropic AND OpenAI: EduAide.ai

Built on Gemini AND OpenAI: Curipod

Proprietary or non-LLM: Ello (proprietary model trained on 130,000+ hours of children's speech data — not a general-purpose LLM).

That’s 10 out of 50. The remaining 40 — including Brisk Teaching, Diffit, Newsela, NoRedInk, Seesaw, Formative, DreamBox, and IXL — have made no confirmed public disclosure of their underlying AI API stack.

The pattern in the confirmed disclosures is also telling: many of them originate from AI developer marketing pages or case studies, not from the companies themselves. MagicSchool’s Claude migration is documented primarily on Anthropic’s website. Khanmigo’s GPT-4 relationship is documented on OpenAI’s website. The EdTech product is the subject of the story; the AI company is doing the telling.

If you’re interested in a full annotated list of all 50 companies with notes, let me know! I’d be happy to publish it in a forthcoming article.

Why Procurement Frameworks Don’t Ask

So what does it mean that 80% of these companies haven’t disclosed? It means the people buying, recommending, and consenting to these tools — administrators, teachers, parents — have no way to know unless they ask. And most of them don’t ask. Not because they don’t care about their students, but because nobody told them to. There is currently no mechanism that requires or even encourages disclosure. No checkbox, no contractual standard, no certification process.

The frameworks they rely on — the procurement checklists, the readiness guides, the RFP templates — were built to protect student data. They do that reasonably well. What they weren’t built to do is surface the question of which AI is actually in the room with your child, and what values it was built with. That question falls through the gap — not through negligence, but through a structural blind spot that we now have the tools to close, if schools and districts start demanding it and founders start getting ahead of it before they’re required to.

Every major K-12 procurement framework treats the underlying LLM as a data privacy question, not a safety or values question. The Future of Privacy Forum’s procurement guide asks whether a vendor’s third-party LLM will use student PII to train the model — legitimate and important, but silent on behavioral commitments or values. CoSN’s Gen AI Readiness Checklist covers security, interoperability, and vendor history. SREB’s AI Tool Checklist covers effectiveness and instructional value. None of them ask which LLM powers the product, or what that model’s documented safety posture is. The fix is twofold: procurement guidelines need to be updated to require model disclosure, and founders need to proactively disclose before they’re asked.

Common Sense Media’s methodology partially addresses this gap — but only at the foundation model layer. CSM rates ChatGPT-5 as “High Risk” for teens and Google Gemini as “High Risk.” Claude is “Moderate Risk” — meaningfully lower, though not a clean bill of health. CSM’s November 2025 mental health assessment found that all major chatbots — including Claude — consistently fail to recognize and respond appropriately to mental health issues in young people. These ratings matter. But they don’t transfer automatically to EdTech products built on top of the rated models. MagicSchool’s decision to switch from OpenAI to Claude — explicitly citing CSM’s safety rating differential — is documented on Anthropic’s marketing page, not in any independent assessment of MagicSchool as a product. The signal exists. The channel to carry it through procurement infrastructure does not.

In May 2026, CSM launched the Youth AI Safety Institute, focused specifically on AI safety for children and funded in part by both Anthropic and the OpenAI Foundation. Its methodology — testing AI systems with teen personas ranging from curious to vulnerable to provocative — is worth following closely.

The Institutional Stakes Just Got Higher

The procurement frameworks gap doesn’t exist in a vacuum — the institutional landscape around it is shifting fast, and three recent developments matter directly for the decisions I’m making right now.

OpenAI’s voice APIs. OpenAI launched three realtime voice API models — voice reasoning, live translation, and live transcription — explicitly for developers building education products. For founders, this is the most significant capability expansion in the space right now. The multilingual unlock alone is meaningful. For example, live translation in a parent-teacher conference changes what’s possible for any district with a significant ELL population. Voice-first AI is a genuinely different product category from text-first tools, and OpenAI moved there first. If your roadmap includes voice or multilingual features and you need them now, that’s the competitive reality you’re building into. The capability gap won’t stay this wide forever, but right now it’s real.

Anthropic and the Gates Foundation. Anthropic announced a $200M, 4-year partnership with the Gates Foundation spanning K-12 tutoring, career advising, literacy, numeracy, and curriculum design — with evidence-based math tutoring expected to be the tip of the spear. The dollar figure matters, but it’s not the most important part of this story for founders. What matters is the benchmark strategy underneath it. The traditional model of education research — deploy a tool, study it for a year, publish findings — is broken for AI. By the time the research publishes, the models have changed.

This partnership is designed to flip that model entirely. Rather than research following deployment, the goal is to build pre-deployment public benchmarks that AI tools have to pass before they go into schools, tested against different learning scenarios and student needs. As EdTech Insiders noted in their May 27 podcast episode, Anthropic is now positioned to help define what those benchmarks look like — what counts as quality, what counts as safe, what counts as effective for K-12. That is a different kind of influence than building a good product. It means Anthropic is helping write the evaluation criteria that everyone building in this space will eventually be measured against.

For EdTech founders, the strategic question isn’t how to access Gates funding. It’s how to position your product to be on the right side of the evidence when those benchmarks arrive. And if you’re building on Claude, you’re already building on the infrastructure layer most likely to shape what that evidence looks like.

Magnifica Humanitas. The Gates partnership is about evidence and accountability at the institutional level. The third development operates at an even broader scale and lands in a place that procurement frameworks have never thought to look. In May 2026, Pope Leo XIV unveiled a papal encyclical on AI — the most significant institutional moral framework applied at the global scale. It was co-presented at the Vatican press conference by Chris Olah, Anthropic co-founder and interpretability researcher, who was personally thanked by the Pope. Anthropic did not co-author the text, but it was invited to co-present it. The encyclical argues that AI cannot be morally neutral, that technical alignment alone is insufficient without shared governance, and that developers bear explicit ethical responsibility for the values embedded in their systems.

For K-12 buyers evaluating vendor relationships, this kind of cross-institutional engagement is relevant evidence about where a company believes its accountability lies — not just to its customers, but to a much broader set of stakeholders. That is precisely the question our procurement frameworks still haven’t learned to ask. The Gates partnership gives Anthropic influence over how AI in education gets evaluated. The Vatican encyclical gives it something harder to manufacture: a documented record of showing up in the rooms where values, not just capabilities, are being debated.

The Gap Worth Filling

We talk a lot about keeping kids safe online. We have content filters, screen time limits, acceptable use policies. But we haven’t asked the most basic question: who built the AI your child is talking to, and what did they teach it to believe? Here’s the structural problem underneath everything I’ve described: the chain of dependency runs in one direction, and it runs downhill.

Models change. EdTech companies are at the mercy of their API providers. Schools are at the mercy of the EdTech companies they’ve contracted with. Parents are at the mercy of the schools. And children are at the bottom of all of it — using products whose AI lineage nobody in that chain thought to ask about.

The evidence in this piece points in one direction. Common Sense Media’s safety ratings point there. The Gates Foundation benchmark partnership points there. The Vatican press conference points there too. For student-facing K-12 products, the case for building on Claude isn’t a marketing preference, it’s what the institutional record currently supports. That’s where my answer lands.

But a verdict without a mechanism is just an opinion. The mechanism is this: administrators need to add one question to every RFP — “What AI model powers this product, and what are its documented behavioral commitments for students?”— and founders need to answer it before they’re asked. Not because a regulation requires it, but because the children using these products deserve a chain of adults who made deliberate choices on their behalf, and can say so.

The question isn’t difficult. The gap isn’t unfillable. It just requires the people at the top of that chain — the founders who choose the model, the administrators who sign the contracts — to decide that it matters.

The 80% who haven’t disclosed aren’t “villains”. They’re companies operating in a market that never asked them to. That market is changing. Be on the right side of it before you’re required to be.

If you've navigated K-12 AI tool selection — especially at the elementary level — or you're building in this space and want a thought partner on the K-12 market, I'd genuinely value the conversation. DM me on LinkedIn or drop a comment below.

Hi, I’m Angela Chen! I’m an edtech founder, Stanford instructor, and AI in education advisor. Every week, I read across the full AI-in-education landscape: research, classroom practice, startup and funding signals, policy, and market intelligence.

I’ve been doing this privately for my own work and a few people kept asking what I was tracking. So I’m sharing the signals I actually think matter, synthesized across everything I read that week, publically now!

If there’s something here worth going deeper on, comment below or DM me on LinkedIn and tell me which signal — I’m always interested in what’s actually landing for people in the field.

The mood this week: a field splitting in two. Massive philanthropic capital is flowing into AI-for-education infrastructure at the foundation layer, while on the ground, districts are cutting ed-tech tools, parents are petitioning to remove screens, and the federal government just told schools to go back to pencil and paper.

The tension between the builders and the resisters has never been sharper — and both sides have real data to work with.

⚡ What to pay attention to right now

Anthropic + Gates commit $200M to AI for good and it’s not hype money

A 4-year, $200M partnership between Anthropic and the Gates Foundation pairs Claude credits and Anthropic staff with Gates grant dollars, explicitly naming evidence-based tutoring, career guidance, and foundational literacy and numeracy as target areas. (Reuters) What makes this different from most philanthropic AI announcements is the specificity: this isn’t “explore AI in education” money — it’s going toward use cases that already have a research base.

What sets this apart from most AI-in-education funding announcements is the Gates Foundation’s involvement — they’ve spent 20 years stress-testing what actually moves outcomes in K-12, and they don’t attach their name to things without an evidence framework.

For builders: watch what use cases get resourced through this pipeline. They’re likely to set the evidence bar for the next major procurement cycle.

The Surgeon General’s office just officially entered the screen time debate and it cuts against ed-tech’s default assumptions

The Trump administration’s Surgeon General’s office formally advised schools to limit student screen time and emphasize paper-and-pencil assignments. (EdWeek) It landed right as Iowa passed a law limiting screen time and districts around the country are already cutting ed-tech tools under budget pressure.

For AI edtech builders, this is a real procurement headwind: even if your product has evidence, you’re now selling into a political environment where the default presumption is shifting toward less screen time, not more. The companies that survive this cycle will be the ones that can make a crisp case for which screen time generates learning gains, not just that they use AI.

K-12 device procurement is being disrupted by AI-driven demand, but not in the way you’d expect

SmartBrief flagged last week that K-12 device procurement is being actively disrupted by AI-driven demand shifts (SmartBrief). Districts are being asked to spec for AI workloads — more processing power, different connectivity requirements — at exactly the moment budgets are tightest and the federal screen time message is creating hesitation.

This creates a practical bind for edtech vendors: the infrastructure layer isn’t keeping up with the product layer, and the districts most excited about AI tools are often the ones least equipped to run them well.

Teachers are getting AI training at scale yet it’s having difficulty landing

Two data points from the same week: AI professional development for teachers is growing rapidly yet a new study finds teachers are still broadly wary of AI — with disparity risks named as a top concern (SmartBrief).

This is the central tension in K-12 AI implementation right now. Training is scaling but trust is not. The wariness about equity is well-founded. AI tools have documented disparities in performance across student populations. PD that doesn’t address this directly isn’t just ineffective, it’s counterproductive.

OpenAI’s realtime voice models might be about to reshape what AI tutoring feels like

OpenAI launched three new realtime voice API models last week — voice reasoning, live translation, and live transcription — explicitly aimed at developers building education products (OpenAI for Education). Students won’t use the API directly, but tutoring platforms, LMS tools, and accessibility products will.

The near-term unlock that’s most underserved: multilingual learner support. Live translation in parent-teacher conferences alone would be meaningful for districts with high ELL populations. Longer term, voice-first AI reduces the typing bottleneck that makes most current AI tutors feel clunky, enabling learning tools to feel more like conversations than forms are genuinely different products. Watch which companies move first to integrate.

I’m genuinely curious about how Anthropic responds. Claude is already the backbone of a number of edtech products and the Gates Foundation partnership announced mid May puts Anthropic squarely in the education market in a more intentional way than before. Does Anthropic’s education team ship something that competes directly on voice or do they double down on reasoning and written feedback, where Claude has a strong case? The answer will shape which model edtech builders reach for when voice becomes table stakes.

🛠️ Tools of note

(These aren’t investment recommendations, they show up here because they’re connected to trends worth tracking.)

Custom GPT suite for pre-calculus (Bhavani Kola / Eclatech Solutions) — Bhavani, a math department chair, built a suite of purpose-built GPTs wired directly into Canvas: syllabus formatting, LaTeX accessibility conversion, animated concept explainers via Manim, AI-resistant assessment design, and interactive flashcard generators.

Why it’s trending: This is practitioner-built AI infrastructure, not vendor-sold. It directly answers the “AI just does the homework” critique by centering accessibility and AI-resistant assessment design. It’s also a preview of what teacher-as-builder looks like at scale, which is relevant for anyone thinking about professional development models.

Indiana University GenAI 101 — Free, open-enrollment foundational AI literacy course; 114,000+ enrolled worldwide as of this week.

Why it’s trending: Governments and universities racing to provide free AI literacy are making the “we can’t afford AI training” argument increasingly difficult to sustain. This is the public-good track of AI skills education going at genuine scale, and it benchmarks what districts could point students toward before buying a vendor solution.

💰 Funding & market moves

Multiverse — Apprenticeships-to-enterprise-AI-training platform, UK $70M raise at $2.1B valuation (a $400M step-up from 2022), led by Schroders Capital with General Catalyst, Lightspeed, Index Ventures. Revenue up 50% YoY (The EdSheet).

This is the second edtech unicorn after Handshake to explicitly pivot from its original model toward “AI adoption platform” positioning. Revenue growth suggests the pivot is real, not just a rebrand.

What it signals for the broader market: enterprise AI training is where the growth story is right now and companies with existing employer relationships are moving fastest to own that position.

Kalos Ventures Fund I — Early-stage workforce, care, and education $78.8M first close, oversubscribed. Early portfolio includes Manifest (SMB-focused AI learning platform) and Rosarium Health.

A fund explicitly targeting the non-enterprise, non-higher-ed quadrant of AI training — small businesses, care workers, frontline employees. This is the part of the labor market that large platform consolidations (Coursera-Udemy, upGrad-Unacademy) are least likely to serve well.

Presence acquires Education Modified — Special education workflow, US Teletherapy and evaluations platform Presence expands into IEP and special education workflow management across 10,000+ schools in 47 states.

Special ed is the most compliance-heavy, documentation-intensive corner of K-12 and AI workflow tools here have a genuinely clear ROI story for districts that are simultaneously under budget pressure and facing rising special ed caseloads.

📚 Research & policy on my radar

Workforce Pell Grant final rule is now in effect — HolonIQ/QS, May 18

The US Department of Education’s final rule expanding Pell Grant eligibility to short-term workforce credentials took effect May 18.

For K-12 AI builders, this matters because it directly funds the career pathways pipeline that secondary schools are supposed to feed. More credential-eligible training options strengthen the case for high school AI and tech programs that connect to community college articulation and it gives districts a funding hook for programs they couldn’t previously justify.

CoSN: Security and AI guidelines are the #1 and #2 ed-tech concerns for districts — EdWeek Market Brief, May 25

CoSN’s annual survey of district technology leaders puts data security and AI policy development ahead of budget or infrastructure as leading concerns.

The practical implication for vendors: if your AI tool doesn’t come with credible data governance and a policy framework out of the box, you’re asking IT directors to say yes to something that immediately becomes their problem to solve. Showing up to a district RFP without a clear data handling answer is no longer a minor gap — it’s a deal-breaker.

PowerSchool’s new CEO signals data + automation as K-12’s next era — EdWeek Market Brief, May 25

Antonio Pietri, seven months in, is framing PowerSchool’s direction around data infrastructure and automation. (EdWeek Market Brief) PowerSchool sits on more K-12 student data than almost any other company. If it moves aggressively into AI-driven insights and automation on top of that data layer, that’s a significant competitive shift for the assessment, intervention, and analytics tools that currently sell alongside SIS platforms.

That’s The Syllabus for this week.

Lots to dig into! I’m especially curious whether others are seeing the grade-differentiated device policy conversation taking shape in their districts.

This is an experiment. I’ve been doing this research privately for years — tracking signals across the AI in education ecosystem so I can show up sharper for my advisory clients and my Stanford students. A few people kept asking what I was reading. So now I’m sharing it. If it’s useful, tell me. If something’s missing, tell me that too.

If this was useful:

📬 Subscribe on Substack — The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take.

🔗 Connect with me on LinkedIn — I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space:

I advise school districts building AI frameworks and edtech founders navigating the K-12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this:

Forward it. Re-stack it. Share it on Linkedin. The best readers I’ll ever have are probably in someone else’s inbox right now.

Business Insider just dropped its 2026 Seed 100 — the definitive ranking of the best early-stage investors in tech. One hundred names. Billions in deployed capital. Portfolios built on generational companies.

I went through all 100 entries. My count: ten investors have any meaningful claim to edtech, future-of-work, or learning-adjacent investments. Ten out of one hundred.

That’s not a catastrophe. It’s a structural signal and the founders who understand it will stop wasting months pitching the wrong rooms and start building relationships with the capital that will actually move for them. That capital is active right now. Here’s where to find it.

The Ten on the Seed 100 With EdTech Credentials

Lars Fjeldsoe-Nielsen (#78, Owl Ventures) is the unambiguous hit. He joined Owl Ventures as GP in 2025 to accelerate global education investments — and explicitly backs education technology startups. He is your person on this list.

Chi-Hua Chien (#37, Goodwater Capital) led Goodwater’s $15M Series A for Ello, an AI reading coach that landed #5 on Fortune’s Change the World list in 2024. Education is named as one of his seven core human utility sectors — a recurring thesis, not a one-time bet.

Gokul Rajaram (#20) is an angel investor in EduFire and Learneo — the platform behind Course Hero, CliffsNotes, and QuillBot, currently valued at $3.6B. EdTech is a listed investment sector for him.

Hunter Walk and Satya Patel (Homebrew) run a fund whose “New Work” thesis explicitly covers how people learn and work. Their portfolio includes ClassDojo and Ello (as returning Series A investors). Two confirmed edtech bets with a shared thesis.

Ann Miura-Ko (#29, Floodgate) backed Thinkful and frames a “Mass Cognition” thesis around AI expanding human learning capacity. Education is embedded in how she thinks about the future.

Ali Partovi (#19, Neo) invests in founders while they’re still in school — a structural bet on the education-to-entrepreneurship pipeline.

Yun-Fang Juan (#60, Brighter Capital) backed Empowerly, a college admissions platform.

Aileen Lee (#100, Cowboy Ventures) has Guild — employer-sponsored learning and workforce education — in her portfolio.

Micah Rosenbloom (#71, Founder Collective) backed Lovevery (early childhood learning) and ULesson (consumer education for Africa). Two real bets across different segments.

The remaining 90 are chasing AI infrastructure, defense tech, fintech, and consumer social. All legitimate, but none of them are your space.

Why Generalist Funds Aren’t Writing EdTech Checks

HolonIQ’s 2025 data shows global edtech investment at $2.6B — up ~11% from 2024, but still a fraction of the $16.7B peak in 2021. After the pandemic boom unwound, generalist funds saw the structural challenges that had been obscured: long institutional sales cycles (12–24 months), committee-based purchasing decisions, and outcome data that’s hard to evaluate without sector fluency.

The 2026 Global Education Outlook shows capital concentrating in two pockets: agentic AI infrastructure and job-aligned workforce upskilling — companies where traction is measurable and enterprise demand is proven. That’s good news for founders building in those categories. But edtech founders who are building outside these categories has a different “capital map” they should keep in mind.

Where EdTech Capital Actually Lives

Specialist Funds Writing Checks Now

Owl Ventures is the defining player globally — $2.2B in committed capital, seven funds, portfolio spanning Coursera, Guild, and MagicSchool AI. In January 2026, they led a $10M Series A for Cloudforce alongside Microsoft, signaling a commiment to what I like to call “edtech+” — the enabling infrastructure layer around education technology. In 2025, they made 10 investments. Consider Joseph Vukel as your first call.

Reach Capital is the one of the most active seed-stage edtech investor in the US. Their most recent investment was in Pathfinder (March 2026). Five unicorns in portfolio including ClassDojo and Replit. If you’re pre-Series A, this is the first call. Consider Caoimhe MacRunnels or Enzo Cavalie your first call.

ReThink Education deployed into six companies in 2025 — Scout EDU, Edexia, Edapt Schools, Montamo, Hiveclass, and Flourish Schools — and is currently raising Fund IV with a $150M target. One of the most thesis-driven, outcome-focused funds in the space. Think of Monique Malcolm-Hay as your first call.

GSV Ventures invests Pre-K to Gray across 65+ portfolio companies and co-presents the ASU+GSV Summit — the most important networking event in edtech. GSV is undergoing several shifts with one of their lead AI x education investors (Claire Zau) leaving to Lightspeed. TBD who I now recommend connecting with!

Emerge Education (London) and Brighteye Ventures (Brussels) lead at pre-seed through Series A for founders building in or expanding to Europe.

Corporate Investors With Distribution Access

These investors write smaller checks ($1M–$10M) but bring something generalist funds can’t: customer relationships that compress your sales cycle.

Microsoft M12 co-led the January 2026 Cloudforce Series A alongside Owl Ventures — a direct signal that M12 is actively backing AI-in-education infrastructure. The deal came with access to Microsoft’s enterprise sales organization and product teams, which is the real asset here. If your product lives inside Microsoft 365 or integrates with Azure, M12 is a strategic investor worth pursuing.

Workday Ventures has been signaling deep commitment to enterprise learning: Workday acquired Sana (AI-powered enterprise learning) for $1.1B in November 2025. Workforce learning and skills platforms that integrate with Workday HR data are in their direct strategic line of sight.

Pearson Ventures signals domain validation to every K-12 and higher ed buyer. WGU Labs combines checks with access to 300,000+ students as a proof-of-concept testing ground. The 2026 GSV Cup — co-presented by AWS and Pearson — selected 50 companies from 3,000 nominations; winning or placing signals credibility to every subsequent institutional investor.

Non-Dilutive Capital — The Strategic Move Most Founders Miss

Too many founders jump straight to raising venture capital when they can be maximizing growth without strings by leveraging non-dilutive capital. This is something I continuously hammer home through my Stanford classes — and now I’m sharing these insights with you!

Federal grants. The IES/ED SBIR program awarded 26 edtech grants in 2025 — 16 Phase I awards (up to $250K) and 8 Phase II awards (up to $1M). The FY2026 solicitation is open now with a June 29, 2026 deadline. Apply.

Philanthropic Funds & Foundations. This is where founders leave the most money on the table. Know the difference between open calls (apply now) and relationship-based funders (build the relationship first):

Open calls:

• NewSchools Venture Fund offers $150K–$250K unrestricted grants for early-stage K-12 learning solutions — they announced $10M+ in funding in October 2025, reviewed on a rolling basis. Their Gates-backed GenAI Math Tutoring initiative offers up to $300K for AI tutoring products specifically. Nonprofit grant-maker, not a VC despite the name.

• Jacobs Foundation runs LEIF (Learning EdTech Impact Funds) — a CHF 40M program explicitly designed to bridge research and edtech investment. The strategic value: Jacobs has direct relationships with Reach Capital, Owl Ventures, ReThink Education, and Brighteye, so a Jacobs grant carries warm signal with the specialist VCs you’ll pitch next.

Relationship-based (no open calls — but worth building toward):

• Chan Zuckerberg Initiative is funding AI-in-education infrastructure right now through its Learning Commons initiative, including a 2026 partnership with Anthropic. CZI moves faster than traditional foundations and is deeply thesis-driven on learning science.

• Overdeck Family Foundation awarded $75M+ in edtech grants in 2025 alone — Quill.org, Khan Academy, and outcomes-based contracting for AI tutoring tools. Their sweet spot is literacy, math, and AI-powered learning products with evidence of impact. No open calls, but worth getting on their radar.

This isn’t charity, it’s strategy. Non-dilutive capital extends your runway and generates the outcome data that every specialist VC now requires before writing a check. Make sure you take advantage of this.

The New Edtech Founder Playbook

Start with the specialists. Reach Capital, ReThink Education, and Owl Ventures are actively deploying. They speak your language and will not make you spend three meetings explaining procurement cycles.

Use the Seed 100 differently. Chi-Hua Chien, Gokul Rajaram, Homebrew — these are warm doors. The co-investors around them (the Ello and ClassDojo cap tables) are the next ring.

Apply for non-dilutive capital before raising your seed. SBIR, CZI, and Gates grants fund the outcome proof points that institutional investors require. Do this in parallel, not in sequence.

Go where the capital is and apply to the competitions. The most efficient thing an edtech founder can do is get in front of the investors who already understand this space. These competitions put you in the room:

• GSV Cup — the world’s most visible edtech startup competition, co-presented by AWS and Pearson. Selected 50 companies from 3,000 nominations for April 2026’s ASU+GSV Summit in San Diego. Winning or even placing is a signal every specialist VC recognizes immediately.

• EdTech Week Shark Tank — New York City, October 2026. Top investors pledge up to $250K in direct investment to finalists. Open to for-profit and nonprofit edtech startups under $10M in funding or revenue. Applications reviewed through September.

• GESAwards (Global EdTech Startup Awards) — London, January 2027. the largest edtech competition globally, with regional semifinals across six continents and a finals stage at Bett in London every January. 9,000+ startups from 130+ countries have competed; finalists get direct access to international investors and education leaders.

• SXSW EDU Launch — Austin, March 2027. Seven finalists pitch to a live panel; winners receive the Launch Award, Impact Award, and Community Choice Award. Now in its 15th year — a consistent path to investor visibility and press coverage.

• LearnLaunch Breakthrough to Scale — Boston, May 2027. Boston-based accelerator with an embedded pitch and investment process. 2027 cohort applications aren’t open just yet, so sign up for updates; accepted companies receive an initial $50K check plus performance-based follow-on capital and deep K-12/workforce investor introductions.

The core insight: edtech’s capital is concentrated in a specific ecosystem of funds, conferences, and competitions. You don’t find these investors by cold-pitching from a generalist VC list. You find them by showing up where they already are

The Bottom Line

The Seed 100 maps generalist venture capital in 2026. EdTech’s capital is organized differently — through specialist funds, corporate investors with distribution needs, and operator angels who built the last generation of learning infrastructure.

HolonIQ projects the global edtech market at $215B in 2026. The investors backing it are active, disciplined, and looking for founders who understand the space as well as they do. That’s the advantage for founders who know where to look.

The map isn’t on the Seed 100. This is the map.

Angela Chen writes Angela Builds, a Substack about building, funding, and scaling in education technology. She is a Stanford instructor, founder, and AI for education consultant. Connect with her on LinkedIn and subscribe to her substack angelabuilds.substack.com

Hi, I’m Angela Chen! I’m an edtech founder, Stanford instructor, and AI in education advisor. Every week, I read across the full AI-in-education landscape: research, classroom practice, startup and funding signals, policy, and market intelligence.

I’ve been doing this privately for my own work and a few people kept asking what I was tracking. So I’m sharing the signals I actually think matter, synthesized across everything I read that week, publically now!

If there’s something here worth going deeper on, comment below or DM me on LinkedIn and tell me which signal — I’m always interested in what’s actually landing for people in the field.

The mood this week is bifurcated: record capital flowing into edtech operations and university AI institutes, while the practitioner layer is wrestling with two hard questions — 1) what do we do with agentic AI in classrooms and 2) are we about to kneecap device access right when students need it most?

⚡ What to pay attention to right now

The K-12 edtech+ market is turning a corner and operations software is where growth investors are placing their bets

Two major rounds landed within weeks of each other: Minga’s $65M from Riverwood Capital (student behavior and engagement — think digital hall passes, student IDs, attendance) and Zum’s $100M from TPG for K-12 transportation. Neither is an AI-first — or even learning-first — play. Both are infrastructure, the kind of operational backbone that districts depend on regardless of budget cycles.

EdWeek Market Brief’s State of the K-12 Industry 2026 report adds the framing: revenue optimism is up for the first time in three years. Growth equity has concluded that districts are stable enough to invest in, but they’re picking durable, recurring-revenue operational tools over speculative AI curriculum plays. For edtech+ founders: if your product is a must-have workflow rather than a nice-to-have AI feature, the funding environment just got meaningfully better. That said, the early-stage market remains quiet. This is a story about growth and private equity, not seed rounds.

Stanford just merged HAI and its Data Science institute and the combined entity has an explicit K-12-through-lifelong-learning mandate

Stanford merged its human-centered AI institute and its data science initiative into a single, expanded organization this month. James Landay will lead it and Fei-Fei Li moves into a university-wide Special Advisor on AI role. The combined institute brings together 400+ scholars, a $60M cumulative grant portfolio, and significant computing infrastructure — and has organized its work around three pillars, one of which is explicitly “transforming education from K-12 through lifelong learners.”

This matters beyond Stanford. When the institution that sets the agenda for AI research also declares K-12-to-lifelong-learning a core pillar, it signals where the serious research money and talent will flow. The Stanford Accelerator for Learning is already running a new seed grant program — AI in Teaching and Learning at Stanford — with proposals due this Thursday, May 15. Worth knowing about if you have Stanford-affiliated collaborators.

The screen backlash is colliding head-on with AI literacy mandates and the contradiction is a real problem

Los Angeles Unified is reviewing all edtech contracts as part of a push to limit student screen time. Meanwhile, New York City, Chicago, and Denver have all issued guidance designating AI lesson planning and drafting as green lit uses. Research on cellphone bans shows mixed academic achievement results.

A piece worth reading carefully this week is this careful unpacking from a veteran teacher: K-8 device rollbacks are defensible and overdue, but blanket bans through 12th grade would dismantle the AI literacy infrastructure states have spent two years building. His argument that resonates with me: the K-8 case and the 9-12 case are not the same question, and treating them as one is where policy goes wrong. District decision-makers need a grade-differentiated framework, not a binary.

I’m wrestling with this exact challenge in a current engagement with a K-5 private school — what’s the right balance for AI implementation school-wide? How early might be too early to start introducing AI literacy concepts to children? I’m currently halfway through this engagement — looking forward to sharing thoughts after I wrap up my work with them later this month. If you’re leading a school and looking for support with AI policy and implementation — reach out! I’d love to help you through this tricky challenge.

Agentic AI is arriving in education and the field doesn’t have a good answer yet on what “learning” means when friction disappears

A practitioner post from AI+Edu=Simplified documents real agentic AI experiments — organizing 1,200 research PDFs, building custom apps on a one-hour train commute, chunking video by timestamp — and lands on the right hard question: agentic AI is optimized to remove friction, but learning often depends on intentional friction. Deliberation, struggle, and the productive discomfort of not-yet-knowing are features, not bugs.

What does it look like when students can delegate organizational and contextual tasks to an agent and spend more cognitive bandwidth on actual learning? What happens when they delegate the learning itself? The field doesn’t have confident answers yet. I think this conversation needs to happen as AI policy decisions are made.

The Virtual Learning Academy Charter School’s mastery-based model looks prescient in the age of AI and their evolution is worth watching

Michael Horn’s interview with outgoing co-founder Steve Kossakoski and new CEO Natalie Berger is a long read worth your time. The Virtual Learning Academy Charter School (a New Hampshire competency-based virtual charter) only gets funded when students demonstrate mastery — literally a payment-for-progress model baked into their charter. In a world where AI can complete traditional coursework, their discussion-based assessment model — 1:1 oral conversations at the end of every competency — becomes a natural integrity layer. Their exploration of using AI as a co-discussant in those assessments is exactly the kind of practitioner-led experimentation the field needs more of.

For K-12 advisors: this school is a model worth holding up when districts ask “how do we ensure students actually learned something?”

🛠️ Tools of note

(These aren’t startup investment recommendations, they show up here because they’re connected to trends worth tracking.)

Teacher’s Buddy — AI platform that generates differentiated lesson plans, assessments, and reports aligned to local curricula; raised $2.3M and reached 12,000+ teachers across 130 countries in 18 months.

Why it’s trending: The localization piece — curricula alignment to specific national standards — is what the US market has mostly failed to crack at scale. That growth pace pre-raise is a signal about global demand running ahead of US-centric product development.

Blomma — $25/month AI career coach built around structured goals and contextual inputs (calendars, performance reviews, resumes); raised $5M from Felicis, aimed at Gen Z workers who lost informal office mentorship post-COVID.

Why it’s trending: Blomma points to something I’ve thought about since my own first edtech company → AI learning products aimed at the 18-to-30 young adult demographic may find fundamentally easier test cases than K-12. That cohort controls its own purchasing decisions, has immediate and legible career stakes, doesn’t require parental consent or district procurement, and is already accustomed to consumer subscription tools. The informal mentorship gap left by remote work is real and large. This is where a lot of AI-in-learning innovation is likely to move first before it works its way back into formal education — and it’s worth watching as both a market signal and a product laboratory.

📚 Research & policy on my radar

AI adoption barriers persist even as the market stabilizes — EdWeek Market Brief, State of the K-12 Industry 2026

Revenue expectations are up for the first time in three years, but districts still cite uncertainty, procurement friction, and unclear outcomes as barriers to AI adoption. Edtech companies pitching AI to districts need sharper evidence stories, not just feature lists.

Cellphone bans show mixed results and the research is still young — EdWeek, May 8

Researchers recommend continued study. The honest answer for K-12 advisors: resist citing “bans work” or “bans don’t work” with confidence. Grade-level differentiation almost certainly matters more than blanket policy, and the data to support that case is still being gathered.

The Canvas cyberattack exposed the single-point-of-failure risk of edtech consolidation — NYT, May 7 / Wired

A ransomware attack hit Instructure (Canvas), disrupting thousands of colleges during finals week and exposing 275 million people’s data. This is the worst-case scenario from a single-platform dependency — and it opens a real conversation with districts about edtech stack concentration risk and vendor data governance.

MIT Technology Review names this “the era of AI malaise” — MIT Technology Review, May 8

The public is neither rejecting AI nor embracing it — just sitting uncomfortably with uncertainty about what it will do to work and society. I see the same thing in district leaders. The framing is useful: name the malaise, then give people a concrete first step rather than another framework.

That’s The Syllabus for this week.

Lots to dig into! I’m especially curious whether others are seeing the grade-differentiated device policy conversation taking shape in their districts.

This is an experiment. I’ve been doing this research privately for years — tracking signals across the AI in education ecosystem so I can show up sharper for my advisory clients and my Stanford students. A few people kept asking what I was reading. So now I’m sharing it. If it’s useful, tell me. If something’s missing, tell me that too.

If this was useful:

📬 Subscribe on Substack — The Syllabus will land in your inbox every Monday. I also write there on an ad hoc basis when something in the field deserves a longer take.

🔗 Connect with me on LinkedIn — I post on AI in education, edtech market signals, and what I’m seeing on the ground from my advisory work.

If you’re working on something in this space:

I advise school districts building AI frameworks and edtech founders navigating the K-12 market. If you’re doing either of those things and want a thought partner, reach out.

If you know someone who should be reading this:

Forward it. The best readers I’ll ever have are probably in someone else’s inbox right now.