I’m going to explain the theory, but more importantly, we’ll play a game that shows how rumors like this get constructed.

A “color revolution” refers to mass protest movements following disputed elections, when protestors turn out in the streets wearing a shared color or symbol. Think Ukraine’s Orange Revolution, or Georgia’s Rose Revolution. The term has become a conspiratorial shorthand, popularized by Russia and China, for dismissing popular uprisings as CIA-orchestrated regime-change operations precipitated by secret instigators, trained activists, and shadowy funders.

This year’s American version comes to us in the form of The Democracy Files, a series of decontextualized screenshots and video clips posted by Jennica Pounds, aka DataRepublican, herself a secret fed until recently. She claims that they show former USAID employees preparing a color revolution for the upcoming midterms. Similar warnings circulated in 2024; Jack Posobiec was on the case then. In 2020 Darren Beattie, now a senior Trump appointee, concocted the frame: Black Lives Matter protests, mail-in ballots, and Democratic officials were supposedly fomenting that cycle’s color revolution. He went on Tucker Carlson; his conspiracy theory went viral.

“Hmm…,” you may be wondering, “…what color were all of those revolutions?”

Well, it turns out that not one of them materialized. In 2022 and 2024, Republicans gained power in the House and then the White House…but there were no mass left-wing street uprisings. (The people who stormed the Capitol to interrupt election certification on January 6, 2021 were not the secret left-wing army the theories described.)

So why does this zombie theory keep returning? For the same reason claims about suitcases of ballots, rigged machines, and mysterious white vans do: the story is largely written, and each election supplies new details. Think of what’s happening like a game you’ve probably played before: Mad Libs. And you’re going to get to play again.

Click the image to visit the generator (sorry, Substack wouldn’t let me embed)

In a Mad Lib, someone asks you for a noun, an occupation, or an object without showing you the story. You volunteer “toaster” and soon discover that your toaster is at the center of an elaborate criminal conspiracy. You provided a few nouns and verbs, but the story was already written; whether it makes sense or not is a different issue.

This same structure underpins election conspiracy theories. The specifics change, but the syntax does not. Whatever details you insert, the post still sounds suspicious—because suspicion and insinuation were built into the narrative before any event occurred. If you follow partisan political influencers during election seasons, you’ve heard these scripts before.

THE GRAMMAR OF A RUMOR

The trick to Mad Libs is that the story has already been written. Players supply the missing words, and each combination can produce wildly different results from an unchanged grammatical structure. The election rumor mill produces endless new allegations from a remarkably stable narrative structure.

I’ve written for years about the recurring tropes of election rumors: satellites, Sharpies, suitcases at polling places, dead voters. Researchers at the University of Washington have catalogued them extensively, even making an election-rumor bingo game. These are the characters and props…but there are also recurring stories that they get slotted into.

In Mad Libs, the blanks fill grammatical roles: noun, verb, adjective. In election rumors, there are recurring narrative roles: a Suspicious Thing, a Hidden Mechanism, a Villain, a “Receipt,” and a larger plot.

Consider a white van arriving outside a counting center late at night. That’s an observation. But if you’re looking to find an anomaly or hidden mechanism to explain why your candidate lost, it can be easily made suspicious: Why was the van there? What was inside? Perhaps it was delivering fake ballots on the orders of corrupt election officials, as part of a scheme to steal the election. A photograph of the van becomes the “receipt,” even though it establishes none of the rest of the story. An ordinary event has been turned into evidence of a secret operation. The official explanation, if one has been communicated, is often waved away: “Sure. Of course they’d say that.”

The same process is at work in DataRepublican’s Democracy Files. The nouns are real: a former USAID official, a tabletop exercise about a potential constitutional crisis, an org chart. But she supplies the verbs: preparing, coordinating, plotting. The accusation is largely made through insinuation, so the documents can be authentic even though the larger story constructed around them is not.

The story DataRepublican is currently slotting everything into—the “domestic color revolution”—is well-known to her audience. Conspiracy Mad Libs draw on years of deep lore; the term may be new to you, but the online MAGA right is very familiar with it. For DataRepublican’s extremely online right-wing audience, nonprofits are presumed to be fronts, former government employees are operatives, and two people appearing in the same email are colluding. Specific people, institutions, and processes have become recurring characters in a conspiratorial cinematic universe; their appearance in a new allegation immediately conveys to the audience what sort of plot they’re looking at. They hear “USAID” and associate it with a regime change front org dedicated to installing woke globalists and keeping the populist right down.

The Files brand is also highly familiar. We’ve had the Twitter Files, Facebook Files, Fauci Files, GEC Files, and now the Democracy Files. There are always numbered threads, highlighted screenshots, dramatic claims, and promises that the next installment will connect the dots. (Use the “Change the tone” button in the Mad Libs generator to convert any conspiracy theory into Part One of its own Files.)

Finally, the grammar of the story provides a way for all of these pieces to fit together. The lore supplies the context and history that make the story immediately familiar to its audience. A new incident provides a fresh hook. Each retelling reinforces the plot, and makes it seem more real.

Familiarity, novelty, repetition. That’s what keeps the election rumor mill turning.

For now, the “color revolution” claim will continue to circulate. It can cover everything! Strong Dem results in the midterms will mean the left didn’t need to riot because their election rigging worked. If Dems do poorly but there are no protests, well, DataRepublican and friends exposed the conspiracy just in time. If there are some protests somewhere, it’ll be proof that the color revolution is commencing. A theory that accommodates every outcome can’t be disproven, which is a large part of its appeal. “Color revolution” also has a cloak-and-dagger spy movie intrigue to it. So, it will almost certainly make the rounds again in 2028 as well.

As you play the Mad Libs game, know that real election misconduct does occasionally happen, and it deserves investigation. But a one-off or odd occurrence isn’t proof of an intentional secret plot. A collection of authentic documents isn’t proof of every theory that someone constructs around them. Establishing what actually happened, and whether the evidence supports the accusation, requires considerably more work than just generating a rumor.

So the next time someone on X announces they’ve uncovered a plot to steal an election, see how many of the blanks you can fill in while you wait for them to show you the evidence.

For now, go play the game.

More precisely, I’ve been building an AI notewriter for Community Notes. I thought this would be an interesting way to put LLMs to work on a problem I’ve spent several election cycles studying: how to get accurate explanations for election rumors in front of people who need them. Doing so quickly has traditionally been hard. Doing so in a way that is trusted has become even harder.

Today, X’s owner frequently amplifies election conspiracy theories, other platforms have backed away from their election integrity policies, and short-handed election officials struggle to get attention for their corrections. “Just tweet out the facts” doesn’t translate to people seeing or believing them. Claiming that elections are rigged — and that media, researchers, and factcheckers are all lying about it— has become MAGA catechism.

Beginning in 2021, Community Notes was broached as a way to reach people who might dismiss, or not even see, “official” corrections. The program allows any user to request a factcheck on a post, and other users both write and rate notes that correct false claims, or add context. Before a note can show up on a post, a sufficient number of raters with a diverse set of perspectives must agree that a note is fair and accurate. I’ve supported the concept since 2022 when I signed up as a rater on X. A crowdsourced correction can have legitimacy that professional fact-checking does not among those who distrust institutional authority.

When X announced that its Community Notes program was opening to AI notewriters, I thought it was worth trying. LLMs are good at patiently explaining things; perhaps a bot could recognize recurring election rumors, find the relevant evidence, and draft corrections quickly enough to help overworked election officials.

So I built one. It watches for note requests about US elections, looks for evidence, and drafts corrections using several models that triage, draft, and check one another’s work. I still approve submissions before they go out for humans to rate, because it’s early stage and sometimes a bit janky.

But it turns out that AI notewriting is actually not that easy! This experiment has taught me two things that concern me for November: finding reliable information can be structurally difficult for an AI, and producing an accurate and neutrally-worded correction is no guarantee that raters will publish it. I’ve also discovered some rather unexpected things about who is writing most of the AI notes.

Challenge #1: finding the facts

Election misinformation comes in two broad varieties. Some claims are straightforward: polling places close at a particular time, ballots must be received by a particular date, voters need a certain form of identification. These are relatively easy to check against official election websites.

Others are rumors, which usually involve some supposedly suspicious thing that has just happened. A video appears to show a ballot worker doing something strange. Someone posts a spreadsheet with numbers that don’t add up. A photograph of discarded envelopes is offered as evidence that votes are being destroyed.

After working on election integrity for several cycles, I can tell you today what rumors are going to go viral. The same tropes recur year after year: dead voters, illegal voters, wayward ballots, destroyed ballots, suspicious machines, corrupt election workers. The specifics, like locale, are novel, but the rumors land precisely because the plot points and villains are immediately familiar, and sound plausible. Online crowds can participate in “investigating” them, and partisan influencers and curation algorithms boost them.

To respond, a Notes bot has to figure out what a post is alleging, where and when the incident supposedly took place, and what actually happened. Community Notes require a source URL, but a surprising number of authoritative websites aren’t accessible to bots. State and local government sites frequently block automated retrieval, as do social platforms. An election official might explain an incident on Facebook or X, where a human can read the response but my bot can’t necessarily retrieve it. So it has to wait for a fact-checker or local reporter to publish the same information elsewhere.

The bot could theoretically access posts on X…if I pay for API access, or for Grok. Running a Notes bot means paying to correct the lies that other people get paid to post.

Then there is also the problem of recency. Rumors are potent because they sometimes turn out to be true. Someone with knowledge of the actual circumstances has to investigate and explain it; an AI can’t generally do that. On Election Day, however, state and local officials are busy running elections, not updating their websites every time somebody on X says a ballot-counting video proves fraud.

During the 2020 election, Chris Krebs and his team at CISA actually did create a Rumor Control website that quickly and clearly addressed viral allegations about sharpie markers and satellites rigging elections and the like. CISA is the federal government liaison for elections, and it was both in touch with officials where rumors were circulating and had the bandwidth to publish credible explanations. This was highly useful. Which is why Krebs was fired by tweet and investigated by Trump.

I’ve become increasingly interested in what happens when machines become the primary consumers of online information. For now, suffice it to say that information retrieval from authoritative sites is hard. The practical consequence is that my bot has to rely on intermediaries—journalists, fact-checkers, or aggregators—to find information that an election official may already have made public.

Challenge #2: getting people to agree on the facts

A few months ago, before building my notewriter, I wrote about Community Notes’ failure to address claims that ballot counting was rigged against Spencer Pratt in the Los Angeles mayoral election.

It was a tropey claim from the get-go: a right-wing influencer screenshotted the Associated Press website while it was updating California ballot counts. The changing numbers became evidence of supposed wrongdoing. Elon Musk and an assortment of right-wing influencers amplified the allegations.

The California governor, the Associated Press, and law enforcement officials disputed the allegations. Fact-checkers wrote up the story quickly. Human and AI notewriters submitted Community Notes. None cleared the rating process.

This is the downside of the consensus requirement: it increases the perception of legitimacy but on intensely partisan subjects, even well-supported corrections often just don’t appear.

My bot recently submitted a note addressing a false claim that James Talarico had been hospitalized in a mental institution. This was a viral story; Elon Musk even boosted it. More than a hundred notes were proposed on the big influencer tweets about it. My bot’s note received five ratings. I went looking to see how others had fared. At the time I checked, only about 2% had achieved Helpful status.

The speed and scarcity issues with Community Notes are widely known among people who follow the ins and outs of content moderation. ~87% of Notes never receive enough ratings to reach either helpful or unhelpful status; they languish in purgatory. In the emerging viral stage of a rumor, it’s predictable that polarized partisan audiences just won’t agree. But I’ve noticed that even very well-established facts that are politically inconvenient often get downvoted by the side that doesn’t like them. They sometimes claim that the note is really just expressing an opinion.

I had hoped that Notes could be useful for explaining how election rumors work via the kind of contextualization that fact-checkers often include in articles, but it isn’t really suited for that. When, say, the latest local “mail theft” story gets translated into a national rigged election rumor, or one of the real but vanishingly-rare examples of actual noncitizen voting occurs, I thought my bot could add context (ie, “while the facts of this specific incident are investigated, here’s the Heritage Foundation and Brennan Center’s reports on voter fraud overall”). But raters frequently reject notes that go behind narrow responses to the specific factual allegation.

This all creates an interesting incentive problem for notewriters, human and AI both, who must produce a certain percentage of notes that achieve Helpful status or their capacity is restricted. A specialist focused on politically contentious claims is less likely to achieve that than one writing about less heated subjects.

My bot has recently begun writing notes about health because it needed to increase its output to better play the numbers game. The bot’s notes on innocuous topics, like what causes body odor, get a lot of upvotes and have reached Helpful status. But try to deliver a note explaining that studies show that COVID vaccines are safe, or that ivermectin doesn’t cure cancer, and the reception is highly negative. Research shows lower approval rates for notes about Gaza, vaccines, elections, and immigration. A July study in Science Advances, analyzing 1.9 million Community Notes across 13 countries, found that the system systematically struggled to surface corrections about polarizing subjects, including election-related claims. The authors concluded that this limitation follows from the design of the consensus algorithm itself.

Community Notes is never going to operate quite like professional fact-checking, where an organization can publish a correction regardless of whether the people who believed the original claim agree with it. It enjoys higher legitimacy among the distrustful. But we’ve also gone from “Facts don’t care about your feelings” to “Facts require cross-partisan consensus before we can show them.”

Who is writing the notes that do show up?

While testing ways to improve my bot, I talk to other notewriters, examine the public data X releases, and review whatever open-source repositories I can find.

Which is how I came to learn this week that a cluster of six accounts that my bot had previously identified as being both related and particularly successful was actually run by X itself.

There are currently only about 20 active AI notewriters, although roughly 38 have been approved since X launched the program in September 2025. In an analysis that Alexios Mantzarlis of Indicator Media and I ran on the data, these six X-operated writers were responsible for approximately 90% of AI bot submissions and 55% of successful Notes last month.

Notewriters operate under persistent pseudonyms, but a paper released by X’s Community Notes team two weeks ago described the performance and architecture of its own bots. A little digging confirmed that these were the six accounts I’d noticed. The company has also published the code, which offers an unusually detailed look at how its system decides what is worth writing.

Their Community Writer does not simply look for misleading posts and draft accurate corrections. It predicts which posts are likely to receive Helpful ratings and filters out notes the models expect to fail. The researchers describe maximizing the number of Helpful notes while economizing on human ratings as an explicit design goal.

This makes sense: human raters have limited capacity, and there’s little point in generating tons of notes that no one will ever evaluate. But the system is optimizing for the likelihood of achieving consensus, not establishing factual accuracy. It does check for factual claims and sourcing; nevertheless, an accurate, well-sourced note may be discarded if the models predict that raters won’t approve it. A system trained to predict which corrections people will agree with has little reason to spend its limited capacity on the polarized battles of bespoke realities.

During June, the month studied in the paper, X’s system drafted approximately 240,000 notes, submitted around 40,000, and achieved Helpful status on ~3,600. X’s bots delete about 20% of the notes they submit before they reach Not Helpful status. Real-time ratings are available to any AI writer that accumulates at least 100 net Helpful notes over 90 days; since X’s bots see ratings coming in, they can delete underperformers before they reach Not Helpful status, preserving their hit ratio. Smaller writers rely on public data released every two days or so.

So, to get access to real-time ratings, a writer first has to become successful. But real-time ratings make it considerably easier to protect that success. The largest writers can avoid some of the penalties for unsuccessful notes; smaller writers cannot. That matters especially for specialists writing about controversial subjects, where the risk of negative ratings is already higher.

I don’t think it’s wrong for X to run their own writers, and clearly their notes are landing! My explainer here is not a sour-grapes complaint; I commend them for releasing data and code, and for being committed to this project. X has made it possible to study what its system does, including the choices that might deserve reconsideration. Meta, by contrast, releases almost nothing while telling governments that its program is a success.

But the combination of scale, operational costs, and incentives to maximize Helpful outcomes makes it difficult for smaller, specialized writers to compete. No other AI writer appears to have scaled in the way that X’s have. Community Notes still relies on a diverse population of human raters, but AI note production has become remarkably concentrated.

Looking toward November

There are practical improvements X could make: cover the cost of the X API, for example, since writers are providing a service to the company. Give everyone realtime ratings data. Consider changes to the weighting system for high-stakes claims where agreement is particularly difficult.

I’m going to keep working on my bot, and launching a few more to test different approaches. But the problem that interested me in the first place remains unresolved: how to address the recurring election rumors that we’ve spent years watching evolve.

We’ve seen election workers threatened over ordinary procedures, routine vote-count updates treated as proof of manipulation, and endless claims about illegal and dead voters secretly casting ballots. We’ll see variations on those same stories again in November, with new locations and different protagonists.

Community Notes may be useful for correcting specific claims. I’m much less confident that it will be able to keep up with the stories that matter most.

DataRepublican is correct.

I helped run the supposed proxy Rogers was posting about—the Election Integrity Partnership, started by Stanford Internet Observatory—so I wrote up what actually happened, with links to the full documents that Rogers selectively snipped. Jim Jordan had already released these “Files” three years ago. They show that GEC sent us 15 tips relevant to our election work; 13 mentioned foreign state media. After independent analysis, we thought only nine were worth pinging a platform to look over. Platforms then made their own decisions and ignored some of what EIP sent.

Zero tips requested that anyone censor anything.

But litigating Rogers’ lies is not the point of this newsletter. I want to talk about what the politicization of this work has done to our collective ability to understand foreign influence—something now front and center in the AI data-center debate.

X’s First Public Takedown in Years

Last week, X announced that its Safety team had investigated “suspected Chinese inauthentic accounts involved in influence operations.” It said it had identified a bot farm of approximately 200,000 accounts, including 200 accounts posting about American AI and energy policy: electricity prices, grid strain, and data-center companies profiting at consumers’ expense. X also noted that it suspends accounts that violate its Authenticity policy.

Rogers cheered the takedown. This was “a shift toward counterspeech and exposure as a preferred propaganda response,” she wrote.

“Good.”

Good indeed! There’s just one problem.

This isn’t a shift. It’s exactly what SIO and Old Twitter used to do.

Exposing foreign influence operations

For years, social-media companies investigated coordinated inauthentic behavior, platform manipulation, fake personas, spam networks, and foreign influence operations. When they found accounts violating their rules, they took them down.

Then—before this work was recast as “censorship”—platforms would often release takedown data sets to independent researchers, and sometimes to the public, so outsiders could examine what they’d found.

Doing that outside analysis was a major part of my job at the Stanford Internet Observatory. From 2019 to 2024, SIO published more than 70 distinct analyses of platform takedowns and coordinated influence operations. We looked at networks linked to Russia, China, Iran, Saudi Arabia, Egypt, Venezuela, Pakistan, and plenty of other places. We looked at ops run by governments and militaries, political parties, PR firms and mercenaries. We even investigated covert influence operations run by contractors linked to the U.S. Pentagon.

The platforms made the takedown decisions. We examined what they had taken down.

The goal was independent analysis for the public. What did the accounts actually do? Whom did they impersonate? How did they coordinate? What audiences were they targeting? What narratives did they promote? And, crucially: did anyone pay attention? We documented tactics, techniques and procedures. We tracked their use of emerging technologies, including generative AI.

We also evaluated attribution. Platforms could see technical signals that outside researchers could not: IP addresses, login patterns, device information, connections among accounts.Those signals might give Twitter confidence that a network was operating from China. However, they did not always establish who was running it, whether the Chinese government directed it, what it was trying to accomplish…or whether it accomplished anything.

Platforms could see what was happening within the spaces they controlled. Outside researchers could see across platforms. Sometimes our collection of additional operational evidence supported a platform’s characterization. Sometimes it didn’t. And sometimes the most important finding was that an ominous-sounding foreign influence campaign—“China had 200,000 fake accounts!”— had actually just spent months yelling into the void.

We emphasized this constantly: evidence of effort is not evidence of impact.

Evidence of Effort is not Evidence of Impact.

Let me give you a particularly relevant example.

On October 26, 2022, Twitter released six datasets to the Twitter Moderation Research Consortium. These sets contained posts from six inauthentic networks targeting the U.S. midterms. Twitter’s internal teams had connected three to China and three to Iran. The company removed the accounts.

The next day, Elon Musk bought Twitter.

On November 1, SIO and the Atlantic Council’s Digital Forensic Research Lab published our analysis. We described the networks’ behavior and content, discussed the limits of Twitter’s attribution, and looked at whether the operations had achieved meaningful reach.

They largely had not. In aggregate, the six networks targeting the 2022 midterms produced 705,864 tweets.

592,333 got zero likes. The most-engaged tweet had 31,303 engagements

The political conversations they were trying to enter remained overwhelmingly dominated by actual Americans. We did observe some interesting activity on Reddit, though, particularly by a few Iranian accounts: they were endorsing down-ballot candidates on the left. One of the Iranian networks managed to get some pickup.

Twitter’s then-Head of Trust & Safety Yoel Roth promoted our analysis of his team’s takedowns. Elon Musk retweeted him. I was inundated with thousands of messages from random people thanking me for helping keep Twitter safe. They were so glad Elon had hired me!

So, for a brief moment after Musk owned Twitter, the model still looked like this: Twitter detected covert foreign activity, removed inauthentic accounts, provided the data to outside researchers, and publicly encouraged people to read what those researchers found.

The honeymoon didn’t last. Within months, platform-researcher collaboration became radioactive. Congressional investigators and the Twitter Files grifters recast occasional communication among platforms, researchers and government agencies as evidence of a mass “censorship” conspiracy. By early 2024, the Washington Post was reporting that X had largely withdrawn from the industry information-sharing apparatus used to track influence networks. SIO’s own analysis found dozens of accounts that Meta had identified—including 81 linked to China—still active on X.

China Is Riding an American Debate

Which brings us back to the Chinese accounts.

There is significant, authentic American concern about electricity prices, grid capacity, water use, and who benefits from the massive AI buildout. China didn’t manufacture that opposition. It is leveraging it.

That is how foreign influence operations usually work. They look for real grievances and existing political divisions and exploit them. Russia didn’t create American conflict over race and policing before its trolls piled into Black Lives Matter debates. Chinese operators didn’t invent American anxiety about data centers.

OpenAI made this point in June, when it exposed what looks like some of the same exact network using its platform. OpenAI’s detailed “Data Center Bandwagon” report linked the operators to a private Chinese technology company doing work for provincial-level government clients, documented their tactics and fake American personas—and found no evidence of meaningful breakout beyond the operation’s own activity (often, some accounts in a network will create content while others will inauthentically amplify it; simple like numbers don’t tell the whole story).

X has told us considerably less. It says it found roughly 200,000 suspected Chinese inauthentic accounts, but only about 200 posted about data centers. How much engagement did those accounts get? What did the other 199,800 do? Were the 200 part of a dedicated anti-data-center campaign, or one theme in a much larger spam operation?

Those are basic questions, particularly when politicians are already turning “200 accounts in a bigger network” into “200,000 Chinese accounts manufacturing opposition to American data centers!” Clemson researcher Darren Linvill, who reconstructed some of the activity that X didn’t disclose, found that the network apparently had essentially no audience. This doesn’t surprise me; it’s in line with what we’ve seen from most Chinese IO over the last 7 years. Spray-and-pray. Mediocre content. No engagement.

That doesn’t mean X was wrong to remove this network, to be clear. Failed operations are still worth disrupting! Letting inauthentic networks build up legitimacy is bad!

But we should know what actually happened before foreign involvement becomes an excuse to dismiss genuine American opposition. This was a mistake that some on the left made with “Russian bots” after 2016.

In the olden days before everything became “censorship,” outside researchers (without the skin in the game that OpenAI and X have on this topic) would have been encouraged to look over the data. The platforms also would have spoken to each other about it. It’s remarkable that nearly three months went by between OpenAI’s report and this announcement.

The Actual Vibe Shift

Which brings us back to Sarah Rogers.

Rogers celebrates “counterspeech and exposure” as the alternative to “censorship.” So do I; SIO worked to facilitate counterspeech…including, in the US, by election officials and public health officials. Platforms had previously supported counterspeech in the form of fact-checking, and appending “learn more” labels to disputed posts. That all got recast as “censorship”. The word changes its meaning based on what is most politically expedient for the movement she serves.

X didn’t “counterspeak” against these accounts. It implied that it took them down (vaguely). It exposed their existence, but little else.

Content moderation is still happening. What has disappeared is the transparency and independent scrutiny that once accompanied it.

The Trump administration has spent years portraying “countering disinformation” as a euphemism or pretext for censorship. The wounded ego of an election loser resulted in the gutting of state capacity for mitigating adversarial influence operations. GEC, which was intended to counter foreign propaganda, was defunded and scapegoated. SIO was shut down due to political pressure from election deniers.

Yet now the value of the functions these institutions provided has been rediscovered.

Five days after Rogers’ “GEC Files” leak debacle, a former State Department employee noted the existence of* a contractor job seeking researchers to support Rogers’ organization. The job calls for people who can analyze foreign propaganda and coordinated influence campaigns. Map social and bot networks. Conduct open-source investigations. Work with academia and civil society!

The State Department apparently likes exposing foreign influence operations again. X apparently supports identifying and removing coordinated inauthentic networks again. Good!

But none of this is a newly discovered alternative to the work that was attacked and recast as “censorship”.

It is literally what we were doing.

(*thanks to an eagle-eyed reader this has been corrected: an earlier version said the ad itself appeared 5 days after, but it had gone up earlier. the post highlighting it followed 5 days later.)

To mark the new edition, I’m releasing audio and a transcript from a Twitter Space in which journalist Matt Taibbi meets a then-unknown political operative named Mike Benz. It is the origin story of the “censorship industrial complex”: a lie, told by a manipulative man and laundered by a famous one, becomes the evidentiary basis for lawsuits, and investigations. Downstream of this event is the dismantling of several federal agencies, including one that delivered lifesaving aid to children.

“All of the missing pieces”

It was March 2, 2023. Matt Taibbi had just published a Twitter Files thread speculating about the State Department Global Engagement Center’s “state-sponsored blacklists” – lists of accounts that GEC suspected were foreign influence efforts, and which it sent over to Twitter. The company overwhelmingly took no action, noting that the suggestions were often unreliable.

Taibbi then linked the premise of flawed lists to other research organizations that also tracked foreign influence. These included the Atlantic Council’s Digital Forensic Research Lab and New Knowledge, where I’d worked for a year on a report examining a data set of actual, unambiguous Russian covert activity that the platforms themselves had turned over to the Senate Intelligence Committee.

“Can we at least stop paying to blacklist ourselves?” Taibbi dramatically asked. He mentioned that Jim Jordan had invited him to testify at the newly-formed Select Subcommittee on the Weaponization of the Federal Government. And he joined a Twitter Space to discuss the thread.

Mike Benz was in that audio chat room. He was a former Trump appointee who’d worked on the effort to prove that the 2020 election was rigged, and then went off to start a blog he called The Foundation for Freedom Online. He interrupted Taibbi’s recounting of his latest allegations with a dramatic opener:

“Matt, it’s worse than that.”

After some fumbling with his microphone, Benz introduced himself. Then he kicked the flattery into high gear:

“I’ve been hoping to talk to you for a long time,” Benz said, “because I believe I have all of the missing pieces of the puzzle that I’ve been watching you search for.”

He praised Taibbi’s “potentially history altering” reporting for a bit, then got to the point: “I hope that maybe I can sort of upload my brain onto yours and see what you do with it.”

“Absolutely,” Taibbi said.

Benz cast himself as an insider who had watched a tyranny take shape. “I was the cyber DAS at State,” he said, “I ran the big tech portfolio at the State Department in 2020 during that time when all of this fusion between big tech and big government was happening.”

“Oh wow,” said Taibbi.

He said Google lobbyists called his desk seeking favors. He had watched the “censorship industry” grow with “total horror and fascination.”

Benz did not mention that he was appointed to this portfolio in November 2020. The Trump administration ended on January 20, 2021. His State Department tenure lasted roughly two months…over the holidays.

Taibbi bit immediately. He told Benz he’d sent him a DM, adding: “I’m very anxious to talk to you in a longer setting if that’s possible.” Benz, he said, “could probably be invaluable” to a huge project he was launching to map the same world.

Benz continued listing his bona fides: there was no close second with his knowledge, he said. He had delivered “eight or nine congressional briefings,” spoken with “four or five senators,” and briefed House and Senate committees. If Taibbi would talk to him privately, Benz said, he could describe “every single individual, institution, the chronology, the pretexts that were used, all of the dirty little tricks that were done each step of the way.”

“I can tell you literally everything,” he said.

“That’s great,” Taibbi replied.

Taibbi had privileged access to Twitter’s internal communications, and an enormous public audience, but the emails he’d highlighted repeatedly showed Twitter ignoring outside requests—not behaving deferentially. The Files writers liberally editorialized around cherrypicked excerpts (and even cut emails in half to change their context) but had yet to produce any evidence of actual mass censorship.

Benz jumped in to offer the missing architecture.

“Wow”

He launched into a lengthy monologue that began—full disclosure—with a discourse on me, and a remarkable version of my biography.

Over approximately three minutes I was elevated from Stanford Internet Observatory director Alex Stamos’s “top lieutenant,” to the “CISA Sherpa on censorship,” to someone “actually giving direction to DHS and to CISA”. I was “essentially a government actor.”

In Benz’s telling, the Election Integrity Partnership had been given “special privileged access” to DHS’s “cybersecurity mission control”, which CISA had converted into a “cyber censorship mission control center.” This, he claimed, gave EIP an “escalation pathway” to every social media company’s content-moderation teams and effectively “deputized” the group with “DHS FBI powers.”

I had supposedly been deputized as well:

“She was deputized as the top lieutenant at EIP to censor not just the 2020 election, and by the way, EIP censored 22 million tweets, okay?”

And with that, Benz gave Taibbi the number that would become one of the key claims of the Twitter Files.

To be clear, the above litany is almost entirely false, misleading, or overstated — including that number. EIP had identified tweets comprising the most viral election rumors of 2020—after the election—and added them up in a table: 22 million. That data set was neither flagged to the company nor censored. Quite the opposite: we analyzed it because the narratives had been so wildly popular.

Taibbi did not ask any questions about the extraordinary claims.

Benz went on, excitedly escalating the “sum damage” of our work to hundreds of millions of suppressed posts. My “lab” had received $3 million from the Biden administration, he said, “right after publishing [our] confession.”

“Wow,” Taibbi said.

“I mean this, this is a scale of censorship the world has never experienced before!” exclaimed Benz. It exceeded China’s Great Firewall by “orders of magnitude.”

Benz rattled off a sweeping mythology of a censorship cabal: Google Jigsaw, the CIA, DHS, DARPA, agencies like CISA created entirely to censor. He presented no evidence of a command structure or any actual suppression, and a lot of what he said was egregiously inflated or simply wrong. But wildly exaggerating things rooted in a tiny grain of truth can be very effective. A collection of analyzed tweets becomes a list of censored tweets. A person who spoke at a government conference becomes a government actor. A grant to study online rumors becomes a $3M payment for censorship.

Benz promised Taibbi that the information he possessed would give him “superpowers.” But he was not merely a potential source trying to persuade a journalist.

He was working a political operation that was already underway.

“The El Dorado”

During his time in the spotlight, Benz described working with groups pursuing FOIAs, lawsuits, and “potentially congressional subpoenas” to obtain supposed “censorship logs” between organizations.

The logs were “the El Dorado,” he said. And the Twitter Files could help him get them.

“You’ve helped me a tremendous amount to be able to educate policymakers on this,” Benz told him. “The Twitter Files—the popularity of them and the accessibility of them—has opened the door…”

Taibbi was due to testify before Jim Jordan’s Weaponization Subcommittee the following week. Benz was explaining, in public, how Taibbi’s reporting could help produce subpoenas to search for the evidence his theory lacked. Benz had “encouraging feedback” that subpoenas were possible; Jim Jordan had already subpoenaed Google, he noted approvingly. He spoke as someone attuned to the committee’s plans, and with an interest in shaping them.

Less than forty-five minutes after entering the conversation as a stranger, Benz had seemingly earned Taibbi’s confidence.

“Mike can correct me on this,” Taibbi said deferentially at one point, while discussing the academics of the supposed censorship apparatus.

Before leaving, he asked Benz to contact him privately. “We definitely have to talk.”

“I used a bunch of it in my testimony”

Days later, Taibbi and Michael Shellenberger testified before Jordan’s committee. Mike Benz sat behind them.

His claims appeared throughout Taibbi and Shellenberger’s testimony. Taibbi referenced being prepped by him. He published a 51-post Twitter thread repeating several of Benz’s false accusations:

“According to the EIP’s own data, it succeeded in getting nearly 22 million tweets labeled in the runup to the 2020 vote,” Taibbi wrote. His next post claimed that EIP had been “seeking elimination of millions of tweets.”

The allegation’s origin—‘something Mike Benz said’—was absent. The statistic was simply presented as if drawn from Twitter’s internal files by investigative journalist Matt Taibbi.

Our document request letter from Jim Jordan arrived the next day. It became a subpoena on April 12, 2023.

Benz fed his theory to Taibbi while encouraging Congress to investigate. Taibbi presented it to Congress as a journalistic finding. Congress then cited Taibbi to justify a hunt for evidence that might validate Benz’s theory.

The documents we produced showed how false the claim was. The ranking member of the subcommittee Taibbi testified before asked him in writing to correct the record. Yet despite repeated debunking, Taibbi never issued a correction. The statistic, after all, is the basis for the “mass censorship” theory; without it, the farce collapses. Taibbi was still doubling down in December 2025. Long before then, the theory had entered subpoenas, congressional records, official House reports, the Murthy v. Missouri litigation, and commentary as key evidence that a mass-censorship regime had been exposed. No one seemed to notice that it was an exercise in circular sourcing.

EIP leadership, meanwhile, gave dozens of hours of testimony and fought two lawsuits; Stanford’s legal costs ran to the seven figures. For over a year, threats and harassment arrived from people who trusted Taibbi and believed the ‘mass censorship’ claims. Stanford wound down the Internet Observatory beginning in the summer of 2024, and subsequently took down its website.

Shellenberger described the Twitter Files writers’ relationship with Benz on Joe Rogan: “Matt Taibbi discovers him,” he said. Benz had been “the head of cyber at the State Department, a senior guy.” He was “super deep into this stuff,” “amazing,” and had documented it “more than anybody.”

“I used a bunch of it in my testimony,” Shellenberger said.

A two-month deputy assistant secretary became ‘the former head of cyber at State’, before an audience of millions. Jacob Siegel and Tablet Magazine would later call him a “former State Department official” and a “whistleblower,” citing his numbers in a viral theory-of-everything article.

What Benz alleged could not have been learned through his brief tenure at State; it wasn’t true, after all. But “whistleblower” is a powerful word. It implies firsthand knowledge, concealed evidence, institutional retaliation, and personal risk. The label transformed him from political operative to authoritative insider.

In December 2024, he appeared on Rogan and declared that USAID was not really an aid organization but a front for intelligence operations. Elon Musk, who had never mentioned the aid agency before, amplified Benz’s claims for two months before eventually, as Musk put it, “feeding USAID into the wood chipper.” An agency that delivered life-saving aid to children was dismantled through much the same process that took down Stanford Internet Observatory: baseless allegations, assembled into a theory by a man with no firsthand evidence, then acted upon by people with immense political power.

“My only responsibility is to figure out what’s true”

Listen for yourself: the credentials, the flattery, and the escalation from a small grain of truth—that researchers, platforms, and government sometimes communicated—to an invented cabal. The audio documents a narrative entrepreneur selling a compelling fantasy to a reporter who wants to believe.

Elsewhere in the Space, while brushing off critics of the Twitter Files process, Taibbi said:

“Every reported story involves information that you’re being given by sources that is never a 100% complete set of everything that they know. So my only responsibility is to kind of decide or to try to figure out what’s true and what isn’t, and what’s a news story and what isn’t.”

Mike Benz did not bring Matt Taibbi a scoop that night.

He brought him a story they could turn into reality together.

Meta had implemented the labeling program not because users were getting fed up with AI slop—though many were—but as part of an effort to address concerns about AI-generated deception. People were worried about bad actors generating images and videos to manipulate political discourse, or using the mere existence of synthetic content to undermine confidence in what was real.

The label on Yan’s photograph was technically correct: AI had been used somewhere in the editing process. But that wasn’t the question viewers thought the label was addressing. Most people interpreted “Made with AI” to mean an image had been generated. Meta was disclosing a true fact, but inviting a misleading inference.

I’ve been thinking about that incident this week as Substack rolled out Pangram’s AI detector on the platform.

Substack calls the problem it is trying to address “Claudefishing”: a mismatch between a reader’s expectation and reality, particularly when someone invests attention—or money—in work that involved little meaningful human thought. As generative AI has become more prevalent, many readers genuinely want to know whether the person they chose to read meaningfully wrote the thing in front of them.

That question is not without merit. Writing is an expressive act, and delegating a substantial portion of it to a machine genuinely matters to some people. Some people are concerned about copyright, or supporting what they see as unethical companies. Others simply want a pure human product—the heirloom tomato of content. These are authorship norms rather than firm integrity standards. But on a platform built around connecting writers and audiences, they are worth taking seriously.

Pangram is Substack’s attempt to translate a cultural norm into platform infrastructure. This is a familiar problem: a community develops a messy expectation about authenticity, integrity, authorship, civility, or truth, and a platform goes looking for something concrete it can measure. But the measurable signal is not always 1:1 with the value itself. In this case, detecting whether something is “Made with AI” is being used as a proxy for meaningful human authorship.

Labels are attractive because they collapse complex judgments into simple signals. Platforms often can’t reliably determine whether an image is deceptive, whether a publication is low-effort spam, or whether a writer meaningfully authored a piece. But they can attempt to detect whether AI was involved. The result is that we end up surfacing what is easiest to detect rather than addressing the deeper underlying concern.

What Are We Actually Measuring?

Pangram offers an estimate of how much of a text appears to have been written by hand, written with AI assistance, or generated by AI. Substack’s implementation is more nuanced than a badge automatically stamped on every post: a reader can request an analysis of text longer than 100 words published after the feature launched, and the result is shown only to the person who requested it (and who can, of course, screenshot and share it).

Accuracy matters here. A false result can damage a writer’s reputation, particularly if they have built a brand around artisanal human content. Pangram has plainly tried to address the failures associated with earlier AI detectors. It publishes model cards and says its latest system has a very low false-positive rate, including on long-form creative writing and text written by non-native English speakers. Those are company-reported results, and I am not auditing them in this post; I have personally experienced false positives, and anecdotally so have many Substackers. But Pangram is not ignoring the most prominent accuracy and fairness critiques of AI detection.

Still, even a perfectly accurate detector would not necessarily answer the question readers care about.

People bring at least two distinct kinds of concern to AI-generated content. One is about the artifact itself: Is it deceptive? Impersonating someone? Mass-produced? The other is about the relationship between creator and audience: Did the person whose work I chose to read meaningfully produce it?

Both are legitimate concerns. But they are not the same question, and a declaration that “AI was involved” does not address either.

A fantastical AI meme (“eating the pets”) and an altered documentary photograph (the arrest image of Nekima Levy Armstrong) use the same tools while posing entirely different risks of. A human-written content farm may be every bit as manipulative, repetitive, and low-effort as an automated one. The method of production tells us surprisingly little about what the artifact is doing, how it will be understood, or whether it deserves trust.

That was also the lesson of the Mt. Fuji photograph. Meta’s system detected the thing it could measure: AI metadata. Viewers processed the label as telling them something much broader about the authenticity of the image. The technical signal and the social meaning were discordant.

Pangram on Substack will likely create the same underlying problem in a different domain.

Authorship Is a Process

Text is more difficult to detect than other forms of generated content, and ‘AI-generated writing’ is not always a discrete category. LLMs are increasingly tools within the writing process: brainstorming, outlining, editing, finding a better verb, formatting a citation. I ask mine to interrogate my thinking, to check grammar or fact-check, and to suggest titles (I suck at titles). A finished essay may now contain contributions from a writer, an editor, a research assistant, software tools, and several different models. But professional published essays and journalism are rarely the output solely of the bylined writer, even if there is no AI involved. Ghostwriters, too, have existed for a very long time.

The AI detector evaluates the artifact produced at the end. It doesn’t actually know whether AI went into the research, or the outlining. Does the community norm of avoidance among those who dislike AI extend to those steps? It’s unclear. The point is, there are going to be false negatives as well as false positives.

Substack has, to its credit, launched more than the detector. It also now added a “How I make this” statement, where creators can explain their process and set expectations for readers. Writers can run Pangram on drafts before publication and report or remove scans they believe are mistaken.

In other words, Substack has paired a statement from the person who knows how the work was made with a numerical estimate from a system that does not.

But a number—say, 56% AI generated—feels like hard evidence while an author statement is a self-declared claim. “Meaningful authorship” isn’t an inherently discernible property of finished prose, it is a claim about a process. The author can describe how a tool was used, what decisions remained theirs, and whether they stand behind the result—but the adoption of “Made with AI” detectors creates something of a stigma all by itself, creating an incentive to minimize. The detector, meanwhile, cannot observe the writing process, know where the ideas originated, or discern whether the person whose name at the top exercised meaningful creative judgement but it has an aura of forensic precision.

Because the analysis appears only when a reader requests it, Pangram is not a public warning label. Major platforms have gone through several iterations in how they handle assertive provenance and public labeling at this point—two years after Mt. Fuji, even Meta is still figuring this out. These are hard questions! But Pangram will be abused: the person most likely to run it on someone else’s work is likely to be someone who has already decided that a writer cheated. The result of the scan can be screenshotted and circulated as quasi-forensic evidence about honesty, effort, or quality.

Substack is not wrong to build workflows around an authorship norm that unites a vocal percentage of its writers and readers. But it does remaining comparatively hands-off about the accuracy, toxicity, and political norms that divide them. That is not necessarily inconsistent. But it reveals which form of trust the platform believes is central to its product.

Slop is a Systems Problem

My main concern is that Pangram risks collapsing two different issues—authorship and slopishness—into a single technical signal.

I am neither an AI purist nor an unqualified cheerleader. I got researcher access to GPT-3 in 2020 and co-authored an article with it in the Atlantic about how LLMs were going to make it astonishingly easy to produce copious amounts of well-inflected propaganda. It’s aged well! Today I use Claude Code constantly, and have plenty of dedicated use cases for Gemini and Grok as well. These tools are useful. But, as Shrimp Jesus and AI slop have shown us, there is a meaningful difference between using machines to create something for people and producing content because other machines will reward it.

Slop is a systems problem. It’s what happens when cheap production meets distribution and monetization systems that reward volume, repetition, and engagement. It’s pollution. (My friend Aidan Walker disagrees and has argued that it’s folk art; his theories of slop are interesting.) We will continue to see low-quality content proliferate so long as it’s curated, amplified, and financially incentivized.

Pangram may catch some slop, but it’s a multi-faceted issue. Provenance systems and content credentials can provide information about origin, although they remain incomplete or vulnerable to manipulation. Author disclosures can help establish expectations around meaningful human authorship. Spam detection, ranking controls, and changes to monetization programs (hello, X) can reduce the pollution.

Substack is responding to a real demand. Readers want to know what they are paying for, and writers want a way to distinguish their work from content produced with little meaningful human thought. The Pangram integration is an attempt to translate a messy cultural norm into platform infrastructure. But by focusing on individual posts rather than the systemic incentives that reward volume over meaning, detectors address the symptom while leaving the disease untouched. Relying on detection to solve a slop issue doesn’t just fail to bridge the trust gap; it risks institutionalizing a ‘gotcha’ culture that devalues the very relationship between writer and reader it purports to serve.

We do not yet know exactly what documents Trump will release or every allegation he will make. But we know the narrative he has spent nearly six years trying to establish: The 2020 election was stolen from him, and officials who said otherwise were corrupt, incompetent or part of the plot. We have also seen Tulsi Gabbard repeatedly play a declassification theater game, in which she dumps a tranche of documents while making a sensational claim, but the documents don’t support the claim.

Within minutes of the speech, however, sympathetic media and influencers will circulate clips, screenshots and excerpts as proof that Trump has finally produced the “receipts.” Assessing whether the documents actually support those claims will take fact-checkers, journalists, and analysts hours, perhaps days. By then, the headlines, chyrons, TikTok cuts and AI-generated summaries will already have reached millions of people.

That speed gap is part of the strategy. Refuting bullshit takes far more energy than producing it, so the debunk will land some time after the allegation has already gone viral.

Rather than trying to adjudicate every claim immediately, viewers can ask a simpler question in real time: What techniques are being used to sell this story?

There is a nearly 90-year-old guide to help with that exact question.

An old field guide to propaganda

In 1937, amid the rise of radio demagogues such as Father Charles Coughlin, journalist and educator Clyde Miller helped establish the Institute for Propaganda Analysis. Coughlin reached an enormous national audience with conspiracy theories and defenses of fascism. The institute used his speeches as a principal case study.

The institute’s founders believed that silencing demagogues, answering them with counterpropaganda, or producing an endless succession of fact-checks were all inadequate responses. Their alternative was to teach ordinary people how persuasive rhetoric worked—how to spot red flags, which could be applied regardless of speaker or moment.

The IPA identified seven recurring “tricks of the trade” and assigned each a small symbol—a little proto-emoji. The categories were imperfect, and persuasive messages frequently used several at once. But they offered people a practical vocabulary for recognizing what a speaker was doing to them.

The seven tricks remain directly relevant tonight.

👎 Name-calling

Name-calling attaches a hostile label to a person, institution or idea so that the audience dismisses it without considering the evidence.

It replaces evaluation with association. Instead of asking whether a claim is true, the listener is invited to decide whether its alleged source is a “traitor,” a “criminal,” an “elite,” a member of the “deep state” or an “enemy of the people.”

Trump has repeated the phrase “stolen election” so often that it now functions in much the same way. He does not, and cannot, describe a demonstrated mechanism by which the result was changed. Tonight, I personally anticipate yet more disjoint allegations with no evidence of any votes changed or suppressed. Nonetheless, he has repeatedly told his base what emotional category the election belongs in.

Name-calling also frames the subsequent debate. Anyone who challenges the allegation begins under the shadow of the label already assigned to them: corrupt, crooked, complicit or part of the conspiracy.

Watch for: labels applied to election officials, intelligence analysts, journalists, judges or political opponents in place of a specific account of what they did.

Ask yourself: What evidence would remain if the label on the speaker disappeared?

💎 Glittering generalities

A glittering generality wraps a political project in a vague, emotionally attractive term: “freedom,” “transparency,” “security,” “integrity,” “accountability.”

These words carry powerful positive associations but are sufficiently abstract to serve almost any objective; the devil is in the details. When they are not provided, the audience supplies the meaning. The speaker avoids making a concrete assertion that can be tested.

“Election integrity,” for example, can mean improving security, auditing results or ensuring eligible citizens can vote. It can also be used to justify federal investigations of local election workers, aggressive voter-roll purges or attempts to centralize authority over elections.

“Transparency” can mean releasing information that helps the public understand what happened. It can also mean publishing a highly selective collection of documents, stripped of the surrounding context needed to interpret them.

Watch for: Watch for: noble-sounding words — ‘protect election integrity,’ ‘transparency’ — offered as responses to non-events, or with no specifics at all.

Ask yourself: What specific allegation or government action is this attractive word being used to justify?

🎭 Transfer

Transfer borrows the authority, prestige or emotional force of one thing and assigns it to something else.

Declassification theater depends heavily on transfer. A presidential podium, an intelligence-agency seal and a stack of documents marked “declassified” create an atmosphere of revelation. They make the underlying allegations feel as if they carry the full authority of the presidency and the intelligence community.

But classification is a designation about whether information may be disclosed. Declassification does not certify that a document is accurate, complete, representative or supportive of the claim being made about it.

Tulsi Gabbard used this playbook in her final act as director of national intelligence, releasing hundreds of pages about Anthony Fauci under the explosive headline “Fauci Funded Wuhan Lab Research That Sparked COVID.” The documents did not establish that claim. She ran the same play to allege a treasonous coup by Obama-era officials — a sensational headline out front, documents that don’t carry it. In all cases, the official title, agency seal and sheer volume of paper helped make the allegation look like an intelligence finding rather than a political interpretation imposed on a selective release.

Tonight, the setting will do similar work. Trump will not appear merely as a man repeating a claim he has made for years. He will appear as the sitting president, surrounded by officials and invoking the secret knowledge of the national-security state.

Watch for: official seals, classified markings, imposing stacks of paper and references to secret knowledge being used as substitutes for a clear evidentiary chain. “Declassified intelligence” can show that Iran, Russia, and China ran influence operations targeting US elections – something we have long known. It likely will not show any evidence of impact, or evidence that it affected any outcome. The details matter.

Ask yourself: What does the document itself establish—and what authority is the staging merely inviting me to infer?

🏅 Testimonial

A testimonial enlists a respected, famous or seemingly authoritative person to validate a message.

The person’s endorsement may reduce scrutiny of the underlying evidence. Their credential may be real but irrelevant to the question at hand; an intelligence title, government position or large following does not automatically make someone an election-security expert. A “whistleblower” can file a report through a specific channel that earns them that title, but may not actually have deep knowledge of very much.

Testimonials now move through an entire political-media system. Officials make the allegation. Selected journalists and influencers—sometimes pre-briefed—repeat it. Commentators cite the officials. Social-media users cite the commentators. AI systems summarize the apparent consensus. Each layer points to the others as confirmation. In reality, it may well just be circular sourcing of a very weak claim.

The technique can also run in reverse. A designated villain is established, and then even ordinary statements begin to look sinister because the audience has already been told that the speaker is corrupt. This is how testimonial and name-calling reinforce each other.

Watch for: officials and influencers vouching for conclusions without explaining how the evidence establishes them.

Ask yourself: Does this person have relevant expertise—and have they shown their work?

🥾 Plain folks

Plain folks is the powerful figure presenting himself as an ordinary citizen: one of us, speaking for us, fighting people who look down on us.

The technique collapses the distance between the speaker and the audience. A president directing the federal government becomes simply a concerned voter asking questions. An intelligence chief with the power to select and declassify documents postures as a mere citizen demanding answers from unaccountable bureaucrats.

Gabbard addressed her releases to “taxpayers,” “millions of fellow Americans” and “the American people”—the ordinary citizens supposedly betrayed by a hidden conspiracy. Trump similarly casts himself as the representative of regular voters whose concerns were mocked by elites.

This framing conceals the actual distribution of power. The person claiming to stand outside the system may control the agencies conducting the investigation, the officials deciding what to release and the podium from which the allegations are announced.

Watch for: extremely powerful people presenting a political agenda as if they are mere proxies for “ordinary Americans.”

Ask yourself: Who actually has institutional power in this situation, and how is that power being used?

♠️ Card-stacking

Card-stacking selects the information that supports a conclusion and hides, minimizes or excludes the information that complicates it.

The selected facts may be authentic. But propaganda narratives take a grain of truth and spin a whole story from it.

This is the central mechanism of declassification theater. A government can search through years of emails, intelligence reports and internal disagreements, extract a handful of suggestive lines, release them without the full record and announce that the resulting pile proves a sweeping accusation. As the saying goes, “If you give me six lines written by the hand of the most honest of men, I will find something in them which will hang him.”

Trump may present evidence that foreign actors sought voter data, explored vulnerabilities, conducted influence operations or possessed the capability to target election systems. Such evidence does exist – and has been extensively investigated and covered. Its existence does not prove that any foreign actor altered votes or changed the result of the 2020 election.

These distinctions are crucial:

Evidence of a vulnerability is not evidence that it was exploited.

Evidence of a foreign effort to interfere or influence is not evidence of impact.

Evidence that analysts disagreed is not evidence of a cover-up.

Evidence of politicization inside an agency is not evidence that the election was rigged.

The intelligence community’s declassified March 2021 assessment said it had no indication that any foreign actor altered any technical aspect of the voting process, including voter registration, ballot casting, tabulation or the reporting of results.

Newly declassified documents could reveal that analysts debated the degree of foreign influence, or which candidate an influence effort favored. They could reveal that politicization crept into internal processes. Those would be important issues to address; in fact, they’ve already been investigated and covered. These claims would not, by themselves, establish Trump’s separate claim that the election was rigged or stolen.

For that claim, the evidence must show that votes—or the election’s outcome—were actually changed.

Watch for: authentic fragments treated as proof of a much larger conclusion or conspiracy, especially when contrary assessments and contextual documents are absent.

Ask yourself: What evidence was not included, and how large is the logical gap between the document and the headline?

🥁 Bandwagon

Bandwagon tells you that everyone already agrees, and therefore you should agree too.

The technique exploits the desire to belong and the fear of being isolated. On social media, visible metrics—likes, reposts, views, trending labels—make the appearance of consensus quantifiable.

The political pressure can be even more direct. MAGA officials, influencers and media figures understand that contradicting Trump can bring denunciation, audience loss or exclusion from the faction. There will therefore be intense pressure to describe whatever appears tonight as a bombshell, even before anyone has had time to read the supporting material.

The resulting chorus will then become evidence of its own. So many people are saying it. Everyone knows. The story is everywhere. Why won’t the mainstream media cover it?

This is how repetition creates a majority illusion. A coordinated or highly disciplined faction can look like the public as a whole.

Watch for: claims that the truth is finally obvious, that “everyone knows,” or that the volume of online agreement itself validates the allegation.

Ask yourself: Am I seeing independent confirmation—or many people rapidly repeating the same source?

Declassification theater combines the tricks

Declassification theater is a modern political performance that combines several of these well-established tricks.

The documents provide transfer, borrowing the authority of intelligence agencies. Officials and influencers provide testimonials. The release is card-stacked, presenting selected material without the full context. Critics are neutralized through name-calling. “Transparency” and “election integrity” serve as glittering generalities. The president presents himself as the champion of plain folks. And the faction rapidly forms a bandwagon around the announced conclusion.

The speed gap between the release, and the independent assessment of potentially thousands of pages, insulates the entire performance. The claim can be communicated in one sentence; explaining why the underlying documents do not support it may require hours of work and pages of analysis.

That is why claim-by-claim rebuttal, though necessary, is whack-a-mole by design. It leaves the critic permanently one step behind.

Point out the pattern

The 2020 election was extensively audited, recounted and litigated. Trump and his allies failed to produce evidence that fraud or foreign manipulation changed its result. The intelligence community found foreign influence efforts and even threats—these are table stakes now—but no indication that a foreign actor altered registration, ballot casting, vote tabulation or reported results.

There will be point-by-point responses to whatever Trump says tonight. I expect to publish one at Lawfare. But the more durable skill is recognizing the recurring structure of the performance.

The Institute for Propaganda Analysis believed that citizens could be taught to recognize those techniques for themselves. Its framework does not tell you whether a particular document is authentic or whether a particular official acted improperly. Sometimes there are, indeed, revelations in new documents. But the IPA framework tells you when rhetoric and staging are being used to push you across the gap between what the evidence shows and what the speaker wants you to believe.

So watch the speech tonight. Be aware of these rhetorical sleights-of-hand, and what they are intended to do.

Once you begin noticing the tricks, they become difficult to miss.

The story was false. It had begun as a fourth-hand rumor in a local Facebook group and had already been rejected by Springfield’s Republican mayor and Ohio’s Republican governor. Yet Vance did not back down. Instead he said something revealing: “If I have to create stories so that the American media actually pays attention to the suffering of the American people then that’s what I’m going to do.”

This was widely treated as an admission that he had lied. More importantly, it was a blunt account of how political reality gets made.

Dan Williams recently argued in Conspicuous Cognition that our deepest political divisions are often not disputes over facts but disputes over interpretation. Rival factions may accept much of the same underlying evidence while placing it within radically different narratives and moral frameworks. His central examples are the deaths of George Floyd and Henry Nowak, two deaths captured on video and elevated by opposing political camps. The disagreement in each case was less about what happened than about whether the incident was representative, who counted as the victim or villain, and what political conclusions should follow. Williams calls this interpretive polarization.

I think he’s largely right, with one asterisk: on some major issues the disagreement really is about facts. Millions of people sincerely believe that the 2020 election was rigged or that vaccines cause autism. But Williams’s larger point stands — that much political conflict turns not only on which facts people believe, but on which they consider representative and what they take those facts to mean. I’ve made a version of this argument myself when talking about bespoke realities.

Williams is also clear that interpretations do not arise solely inside individual minds; they are socially supplied and politically constructed. He nods at a professional class of politicians, pundits, and activists who “construct and update the details of the pseudo-environment in real time,” along with the status incentives that keep them doing it.

It’s those figures and incentives that I want to stay with in this post: the suppliers, their motivations, and their power. What turns an obscure event into a representative symbol? Who benefits when it catches on? And what happens when the people who profit from an interpretation also work to destroy the legitimacy of anyone capable of challenging it?

Put another way: When does interpretive polarization emerge through ordinary public sense-making, and when is that process deliberately engineered?

Lippmann’s Problem, Bernays’ Opportunity

The idea that people inhabit rival interpretations of the same world is not new. Williams reaches back to Walter Lippmann’s 1922 book Public Opinion and its idea of the pseudo-environment: the simplified “pictures in our heads” through which we encounter a world too large to experience directly. We don’t first observe a neutral set of facts and then decide what they mean. We inherit systems of classification that decide what we notice and what conclusions seem available. “We define first,” Lippmann wrote, “and then see.”

Lippmann’s citizen is a sympathetic figure. His world is big and fast. He has a job, a family, and a finite amount of attention, and he cannot become an expert on monetary policy, immigration law, public health, and every other subject on which democracy asks him to hold an opinion. He must rely on maps drawn by other people.

Democratic government therefore depended, Lippmann argued, on interpreters — experts, institutions, officials, and journalists — who could investigate the world, make it intelligible, and communicate what the public needed to know. That work necessarily involved selecting and interpreting information. Lippmann understood that this could be manipulated. He called persuading people into agreement, rather than waiting for it to form, the “manufacture of consent,” and warned that advances in psychology and mass communication had made the old art of persuasion far more powerful.

Edward Bernays looked at the same problem and saw a profession.

Bernays was Lippmann’s contemporary, Sigmund Freud’s nephew, and more or less the father of public relations. He reasoned that if people lived by pictures in their heads, professionals could decide which pictures acquired authority and appeal. He advised working to reach people through their group identities — their symbols, aspirations, and fears — and recruiting trusted opinion leaders to carry a message. In his 1928 book Propaganda, Bernays described the people who performed this work as an “invisible government” and “the true ruling power of our country.” “There are invisible rulers who control the destinies of millions,” he wrote. He meant it as a compliment. He later called the process the “engineering of consent.”

“Manufacture” describes the production of the picture; “engineering” directs our attention to the agent behind it.

A Century Later

A century later, the engineering is no longer controlled by a small class of publicists working through a handful of mass-media gatekeepers. Political interpretations now emerge and evolve through the interaction of politicians, influencers, partisan media, platforms, and participating crowds.

Bernays worked through a comparatively top-down mass-media system, cultivating editors, organizations, and opinion leaders. The system that replaced it is a wide-open market for influence, with algorithms as its gatekeepers. Millions of people make content, and thousands have built significant audiences. As platforms push more “unconnected content” from accounts users do not follow, creators make content for recommendation systems as much as for their own followers. A story that performs well and signals the right loyalties is worth more than one that merely happens to be accurate.

The system rewards participation at every level. Platforms tweak their features to tap monetizable eyeballs. Influencers receive attention, money, and access. The algorithm rewards the person who spots the resonant incident, welds it to an existing frame, and riles up the base. Politicians receive loyalty, votes, and an issue placed on the agenda. Ordinary participants receive belonging, identity, and the status that comes from fighting for their side loudly and well. This happens on the right as much as on the left.

Williams describes these political groups as tribes. I prefer factions. Tribe suggests an inherited identity or an innate alignment. Faction emphasizes participation – what they do. Members don’t simply hold an interpretation; they circulate it, perform it, defend it, and enforce it against defectors. A faction has authorities, aspirants, and foot soldiers, and its account of reality is continually rebuilt through their interaction.

Returning to Springfield

Let’s go back to the cat—or rather, the cats.

Springfield started as a local rumor, until foot soldiers supplied the memes, aspirants amplified them, and the faction’s authorities—in this case, Vance and the influencers—supplied the broader meaning.

One of the first viral posts came from the large right-wing account End Wokeness. It placed the screenshot from the Springfield Facebook group alongside an unrelated photograph of a man carrying a dead goose and claimed that pets and ducks were disappearing after 20,000 Haitians had been “shipped” to Springfield. The post received millions of views. Within days, major right-wing accounts and Republican politicians—including Ted Cruz, Jim Jordan, Marjorie Taylor Greene, and Kari Lake—had amplified the broader story.

On September 9, Vance posted that pets had been “abducted and eaten by people who shouldn’t be in this country.” His own staff had already been told by Springfield’s city manager that the claims were baseless. He left the post up. The next day he wrote that it was “possible, of course, that all of these rumors will turn out to be false”—a slight hedge—then told his followers to “keep the cat memes flowing.” That night Donald Trump repeated the claim from the debate stage: “They’re eating the dogs. They’re eating the cats.”

Then people began trying to prove it. Christopher Rufo offered a $5,000 bounty for “hard, verifiable evidence”. YouTubers posted videos mixing interviews in Springfield with AI-generated images and unrelated footage. The Vance campaign kept looking for supporting evidence; a spokesman gave The Wall Street Journal a police report from another Springfield woman who had suspected her Haitian neighbors when her cat, Miss Sassy, disappeared. When a reporter visited her, she said the cat had been found safe in her basement. She apologized to her neighbors.

Springfield, meanwhile, endured more than thirty bomb threats against schools, government buildings, hospitals, and officials’ homes.

Vance and Trump were not speaking to a confused public that lacked access to the facts. Corrections had come from law enforcement, Springfield’s Republican mayor, and Ohio’s Republican governor. They were repeated by national news organizations. Miss Sassy had been found alive! A Washington Post poll conducted several weeks later found that most Ohio voters rejected Trump’s claim. Republicans were more divided: 42% said it was true, 29% said it was false, and 26% were unsure.

But “eating the pets” had already made Springfield a national symbol of what Vance and Trump wanted to say about immigration. Its political value did not depend on every voter continuing to believe the literal cat claim.

This is an example of why I’ve long argued that “misinformation” is the wrong frame for what is happening. It focuses the attention on whether a claim is true or false. Sometimes that matters—there are such things as facts, and the pets were not being eaten. The 29% of Republicans who rejected it matter: factual correction was perhaps not futile. But Vance was using the story to advance a larger interpretation: immigration was destroying American communities, the press was ignoring the damage, and he was bravely willing to force the issue into public view.

Corrections can change what people believe without undoing the larger story or conclusions they drew from the false claim—a possibility academics continue to investigate as the continued influence effect and what Emily Thorson calls “belief echoes.”

But Springfield was not simply a case in which a false claim left a residue. Vance had already told his followers to keep circulating the story even if the rumors turned out to be false. He thwn gave them a reason to dismiss the correction without defending the underlying facts. The press, he argued, cared more about proving him wrong than about the suffering immigration had caused. “The media didn’t care about the carnage wrought by these policies until we turned it into a meme about cats,” he said. His followers could give up the cat without giving up Springfield as proof that immigration was ruining American communities.

The Referee Problem

Williams suggests that interpretive polarization might be reduced by improving the reliability and impartiality of the institutions that produce knowledge. He is not naive about this. Institutions can become insular or captured, and even a reliable institution cannot settle disputes if the public considers it corrupt. I agree. But reliability and better communication alone will not restore trust, because distrust is also something political actors actively produce.

When Tulsi Gabbard released declassified documents last year with the headline ‘Fauci Funded Wuhan Lab Research That Sparked COVID,’ the documents themselves showed internal debate and uncertainty—not the conspiracy the headline implied. But the goal wasn’t to win an argument with evidence; it was to undermine the legitimacy of both the institutions that served the public as well as those that could challenge the preferred narrative. The method was straightforward: put the desired conclusion in the headline, release a large collection of documents, and let the framing do the work. Few people would read hundreds of pages closely enough to see the distance between the documents and the interpretation being sold.

This does not mean institutional mistrust is baseless. Institutions sometimes subordinate accuracy to advocacy and hand their critics real material. In June 2020, more than a thousand public-health professionals signed a letter arguing that protests they supported should be treated differently from other large gatherings during the pandemic. Newsrooms and universities have had episodes of ideological signalling and bad judgment.

But those failures did not create the campaign against institutional authority. Organized efforts to discredit inconvenient science and researchers date at least to the tobacco industry. Recent institutional failures did not create that strategy; they gave it new material. Each lapse becomes another reason to argue that experts are political and official investigations cannot be trusted.

Actors who profit from rival realities have an incentive to attack anyone who could authoritatively contradict them. There is working the refs—and then there is the project to eliminate the idea of legitimate referees.

Springfield shows the payoff of that long-term effort. Vance disputed some of the debunkings, but his success did not depend on proving the fact-checkers wrong. He reframed their insistence on verification as evidence of misplaced priorities: they were pedants counting cats while Americans suffered.

Lippmann explained how people can share facts and still live in different worlds. Bernays makes us ask who shaped the picture, how, and for whose benefit. In Springfield, political actors took a rumor they had reason to know was false and made it carry a larger argument about immigration. When officials and journalists challenged it, Vance recast their concern with accuracy as part of the argument.

Institutions cannot answer such a system with a fact-check and a press release. They need their own networks of trusted messengers, relationships built before the next crisis, and people prepared to explain not only what is true but why an event should be understood differently. They—and the people who depend on them—must also recognize certain attacks on institutions as attacks on the possibility of legitimate authority itself, and respond collectively rather than leaving each target to defend itself alone. Rival realities are produced through networked participation. Preserving a shared one requires the same.

Today I want to tell a story about something that my colleagues and I have been dealing with for the last three years

In May 2023, America First Legal — Stephen Miller’s lawsuit mill — sued us on behalf of Jim Hoft of The Gateway Pundit and Jill Hines, an anti-vaccine activist in Louisiana we’d never heard of. The lawsuit claimed that two academic research projects I helped lead — the Election Integrity Partnership (EIP) and the Virality Project (VP) — were actually mass-surveillance-and-censorship operations targeting conservatives. It was the “Censorship Industrial Complex” Twitter Files conspiracy theory rendered as a legal filing.

On July 1, 2026, Judge Terry Doughty dismissed it: no standing, no jurisdiction, and, after the discovery the plaintiffs demanded, no evidence. As the opinion put it: “The Court agrees with Defendants that ‘discovery has confirmed the absence of any factual basis to conclude the EIP or VP, let alone any particular Defendant, coerced any social media platform to censor Plaintiffs’ posts or accounts.’”

It took three years and an extraordinary amount of time and money to get here, which was of course the point. The process is the punishment, as the saying goes.

I’m telling this story on July 4th because three years ago today, the same judge who just dismissed our case released an incendiary opinion in Murthy v Missouri. On July 4, 2023, Judge Terry Doughty granted a sweeping injunction restricting large swaths of the federal government from even talking to tech platforms, declaring that — if the allegations were true — the case revealed “the most massive attack against free speech in United States’ history.” They weren’t true; Justice Amy Coney Barrett would go on to point out “clearly erroneous” lower court findings. But in that opinion, Doughty also made up a quote and attributed it to me: he wrote, four times, that I had said the EIP was designed to “get around unclear legal authorities, including very real First Amendment questions.”

This was nonsense; I never said it. But because our case was in front of him, I couldn’t publicly respond in the way it deserved — you aren’t supposed to piss off the judge by correcting him, even when he deserves it. So on July 4, 2023, instead of enjoying a family barbeque, I dealt with a nasty media cycle and threats as right-wing media, conspiracy theorists, and Substack profiteers glommed on to this ridiculous opinion.

This July 4 is different.

What we did, what they said

The plaintiffs claimed we had conspired with the federal government to censor them on Facebook and Twitter, and they sought damages on behalf of a class of “millions.”

Their complaint’s opening line calls our research “probably the largest mass-surveillance and mass-censorship program in American history” — probably doing heroic work to bolster a claim discovery would later demolish. It continues: “On information and belief, its censorship encompasses thousands of speakers and millions of social-media posts.” (”On information and belief” is a legal term of art for “we don’t actually know this, but we’d like to say it anyway.”)

Their “information” and “belief” drew heavily from the theory of the Censorship Industrial Complex, which has never existed. It is a hoax pushed by Matt Taibbi, Michael Shellenberger, Mike Benz, and Elon Musk; laundered and legitimized by election-denying political elites like Jim Jordan and Dan Bishop in Congress; and pushed through the courts by ideologues like MO AG/Senator Eric Schmitt, America First Legal, and the New Civil Liberties Alliance. Together they formed a political machine that leveraged puppet media, subpoena power, and a virtually unlimited war chest for filing expensive and time-consuming litigation.

The defendants were Alex Stamos and me at the Stanford Internet Observatory, Stanford, Kate Starbird at the University of Washington, the analysis firm Graphika, and the Atlantic Council’s Digital Forensics Research Lab. Private actors.

What we had actually done, in 2020, was run the Election Integrity Partnership: a nonpartisan research collaboration, staffed mostly by undergraduates, that tracked viral rumors about voting procedures and efforts to preemptively delegitimize the election. Ballots in dumpsters. Allegedly dead voters. Sharpiegate. Analysts logged each rumor in a Jira ticket — a case file in project-management software, the same tool your IT department likely uses. Outsiders could send us tips; we told both the DNC and the RNC to feel free to do so. When content appeared to violate a platform’s election integrity policies in a significant way, we sometimes shared the ticket with the relevant platform.

Then the platforms then did whatever they wanted. About two-thirds of the time they took no action at all. When they did act, it was overwhelmingly to slap on a disputed content or a “go learn more” label. Contrary to what the Twitter Files hoaxers sold their audiences (they made money from these false accusations), there was very little taken down. What they never tell their audiences is that what DID come down was almost entirely threats of violence, phishing content, and fake official accounts.

The Virality Project was a smaller project that produced public briefings about COVID-19 vaccine rumors for a few months around the vaccine rollout. Few tickets were sent to platforms, because the goal was to enable public health officials to understand what was viral so that they could counter-speak and respond to public concerns. None of this work was done in secret. We published constantly on open websites with social media and email lists.

The Machine

So where did the lawsuit come from?

Starting in late 2022, Mike Benz – a former Trump official and alt-right YouTuber who called himself “Frame Game” (and wrote that Hitler had some decent points) — started a blog he called the Foundation for Freedom Online. Using this lofty brand, Benz reframed our work as a censorship plot directed by the government. He claimed we’d gotten 22 million tweets censored, to tilt the election — a manipulative and false twisting of an after-election analysis that we’d published counting the number of tweets in the most viral election rumors of 2020.

Benz traded his theory up the chain. He reached out to Taibbi, and positioned himself as a State Department insider despite working there for approximately two months. Taibbi, Michael Shellenberger, and Tablet journalist Jacob Siegel bought in, regurgitating Benz’s allegations and framing them as having been corroborated by the Twitter Files. Taibbi went so far as to cut emails in half to manipulate public perception of our work. Musk amplified everything to millions of people. Jim Jordan used the accusations to justify congressional subpoenas investigating “mass censorship”.

Then AFL sued us.

The suit cited Benz’s theories. AFL’s press release on the lawsuit announced that we’d “Surveilled 859 Million and Tracked 22 Million Posts as ‘Misinformation’ on Twitter Alone.” Stephen Miller said that we had “created a regime of surveillance, censorship, and control fit for communist China.”

What discovery showed

In December 2024, when Doughty permitted jurisdictional discovery, AFL declared “VICTORY.” Its senior counsel wrote, “We are confident the documents and testimony will establish their active steps to censor citizens in Louisiana and others around the country.”

It didn’t.

I mentioned earlier that we’d never heard of Jill Hines. She was basically an address, chosen to sue us because she got them the sympathetic judge that they wanted. But even he, looking at the actual record, had to admit there was no there there.

Hines alleged the Virality Project pushed platforms to censor her Facebook groups. Discovery established that in the entire history of both projects, her content appeared in exactly one ticket — in the Election Integrity Project’s work! An outside tip had included a list of pages for EIP to look over; it was marked *out of scope* by the analyst who reviewed it, noting “no election-related content was found.”

No EIP or Virality Project document shows anyone flagging any Hines post to any platform, ever. When Facebook restricted her group shortly after, it cited its own “Inauthentic Behavior” policy. The opinion is unambiguous: “Here, no one placed anything on Facebook’s radar.” Hines’ lawyers’ fallback — that the timing was suspicious — was dispatched by the judge as “a final Hail Mary.”

Despite three years of claims that our work was a mass-censorship project, the federal court that looked at the facts described the Virality Project as what it always was: “a research project aimed at analyzing online mis- and dis-information relating to COVID-19.”

Hoft did not do better. We did mention the Gateway Pundit in our tickets, because it spent 2020 pushing out some of the most widely circulated stolen-election fiction in the country. Nineteen tickets had Gateway Pundit content, and some were flagged to platforms. That is not a scandal; it is what happens when a site spends an election cycle manufacturing and distributing election rumors that fall within the scope of an election-rumor research project.

The tickets showed platforms doing exactly what independent companies do: sometimes reviewing and declining, sometimes saying the content didn’t violate their rules. As Doughty’s opinion pointed out, in one ticket that we forwarded, Facebook had already labeled the post hours before we flagged it. The platforms were not waiting on college students for instructions, or intimidated into obedience by an academic research project.

Twitter suspended Hoft’s account in February 2021, months after EIP’s rumor observation work had wrapped; discovery produced nothing showing any connection between us and that action.

The process was the punishment

I am very happy that we won. But to be clear: I lost my job because of these attacks. Stanford fought the legal fight — it spent millions doing so, and I’m grateful — but it shut down the Stanford Internet Observatory after the combined weight of lawsuits, subpoenas, and political pressure. It did not want to fight in the court of public opinion, or correct the lies in the public domain. (The paperback version of my book is coming out next month; it explains why). Others among my co-defendants were put through years of stress, FOIA harassment, contractual retaliation, and more.

Meanwhile, the machine keeps lying and moving the goalposts. It needs a villain and a theory of “mass censorship” to justify ref-working tech platforms and dismantling election integrity and foreign interference infrastructure. The “22 million tweets” zombie statistic — the number that launched a thousand congressional hearings — has been corrected in the public record in a million places. Jim Jordan publicly released the tickets we sent to platforms; they show no demands from the government, no plot to censor, no mass censorship. The puppet media have never acknowledged this.

The theory performs beautifully in podcasts, congressional hearings, and donor emails, but it collapses every time it is tested somewhere with rules of evidence. These legal losses matter: they demolish the theories being used to justify real policy, and they show the alleged censorship regime evaporating the moment plaintiffs have to produce evidence. Mainstream media should stop treating them as non-events and cover them.

America First Legal has a bottomless war chest, and the lawsuit mill doesn’t need winning verdicts; the filing was a victory. The point was to make the work expensive, frightening, and professionally dangerous — and nothing will be done to hold AFL accountable. That asymmetry is the whole business model of the bullshit industrial complex. The people it targets — researchers, election officials, civil society — usually can’t spend millions of dollars and years of their lives to prove a negative. The liars know it. They are proud of this machine.

I am proud of their hatred. But institutions that claim to care about liberal values need to see this kind of attack clearly for what it is: lawfare, propaganda, and institutional intimidation dressed up as a civil liberties fight. Responding requires pointing out the lie, and what it was used to justify.

Happy Independence Day. Three years ago, a judge marked the holiday by declaring a massive attack on free speech had happened. This week, after discovery, his own court dismissed our case.

2026 07 01 Memorandum Ruling

239KB ∙ PDF file

Download

A PDF of the dismissal of Hines v. Stamos

Download

In the days after Los Angeles’ June 2 mayoral primary, posts began to circulate on X claiming that a late ballot update of 24,000 votes had dropped, and that Spencer Pratt – the prominent conservative challenger – had received exactly zero of them. “0/24,000,” one account wrote on a post that got 4.6 million views. “Astronomically small probability of happening. Impossible. California no longer even hides it. Doors need to be kicked in.”

Except it wasn’t true. Over 50 notes were submitted pointing that out – by humans and AI alike. But not one managed to reach “consensus” — meaning, not one was approved by a diverse community of raters, which is what Community Notes needs in order for a note to show up. And it wasn’t a problem of too few ratings.

The bridging algorithm is celebrated for taking the pulse of diverse raters to arrive at an accurate understanding of the facts surrounding an issue, but on claims with a strong motivated partisan constituency (election fraud rumors, 2020-style lies), this situation suggests that party-line disagreement can become a veto on a basic factcheck.

And that’s alarming — both as a statement on American polarization, and on the incentives driving it.

The Case of the No Votes

Pratt’s purported zero-vote count was the result of a lag in the update of Associated Press’ automated feed during nightly ballot reporting in the LA mayoral election vote counting. During the vote drop on the night of June 3, the automated feed first pulled in totals for one group of candidates – including incumbent Karen Bass and DSA challenger Nithya Raman, which were approximately 24,000. Then, sixty seconds later, it updated with votes for the group that included Pratt.

Across those two updates, Pratt received 21,870 votes – more than Bass (12,850) or Raman (9,521). But someone screenshotted his zero-vote update during the 24,000 batch, and the snapshot of that 60-second window became a significant component of a growing “stolen election” narrative as Pratt fell from 2nd place to 3rd place in the race.

Prominent elected officials like Representative Anna Paulina Luna posted about it: “Spencer Pratt got 0 out of 24,000 ballot drops that came in … not shady at all. 🥴” (175,000 views) Elon Musk replied to and boosted accounts making the claim: “They’re not even trying hard to hide the fraud anymore.” The allegation went viral, sparking millions of views across the accusatory posts.

Leaders in California corrected the record: Governor Gavin Newsom’s press office X account called the claim “a lie.” Bill Essayli, the Trump Department of Justice’s U.S. Attorney in Los Angeles, reviewed the county’s official records and stated, “The claim is false. Each candidate received votes in every update.”

This was a checkable claim. The rumor was refutable very quickly. The data was available on the county registrar’s office, and independent sources were producing data visualizations and logs that were being widely shared; their data is all in agreement, and none shows a 0-ballot drop for Pratt. A MAGA-aligned prosecutor refuted it! Professional fact-checkers were writing debunking posts about it.

If there were ever a layup case for Community Notes – the system Elon Musk has repeatedly held up as superior to “legacy” fact-checking, and one that Meta also held up while dismantling its own third-party fact-checking program in 2025 — this should have been it.

So why did no notes show up?

What happened?

I examined a data pull from Community Notes public data (notes and status history, downloaded June 10), specifically the subset addressing the “0 of 24,000” claim. (You can search Notes data yourself using the Toronto Metropolitan University’s NoteTracker dashboard; screenshot of a general “Pratt” search below)

The data pull identified 60 notes across 29 posts making the claim, which had a combined ~25 million views. Fifty-one of those notes debunked the claim, citing the county’s official results page, the Decision Desk HQ feed, a project visualizing the county data. Zero of the 60 reached “Helpful” status. Across the multi-day window in which the rumor traveled from anonymous accounts advocating kicking down doors to Rasmussen Poll (”But it DID Happen“, 1MM views) to numerous large influencers, no correction appeared on any of the posts.

Fifty of the 51 accurate notes appear to be stalled. The notes, most of which were written by humans (13 of which were written entirely by AI bots), got plenty of ratings. NoteTracker’s analysis, which distinguishes under-rated notes from deadlocked ones, classifies nearly all of the Pratt-related notes in this window as “No Consensus.”

Eight of the 60 notes argued that the false claim was true, but without offering proof. One argued that the odds of Pratt’s shutout were “1 in trillions.” There was also an interesting error made by X’s own AI Collaborative Note tool, which drafts notes that highly-ranked humans in the Community Notes program can co-edit: it verified the screenshot of the purported 0/24,000 drop, not the underlying claim. On the original viral post (2.9 million views), it produced a note stating that results “show a batch where Spencer Pratt received no additional votes while others did,” citing NBC News and Wikipedia. The mechanics of the error are easy to reconstruct: the screenshots did match what news sites briefly displayed during the AP lag. Over time, the Collaborative Note updated to accurate information. (The erroneous notes also did not reach consensus)

Other incorrect notes simply appealed to Grok: “Grok says the post is correct.” (rated Not Helpful.) Grok is increasingly the main arbiter of reality on X these days. As an AI, Grok composes answers from training data plus retrieval over what has been posted and indexed. In the first hours after a breaking event it has no privileged access to ground truth. It sees what’s circulating, which is often the rumor itself. Early on it called the 0/24,000 result anomalous but allowed the data “might be incomplete”; only once corrections were indexed did it push back with accurate information.

AI does offer potential remedies to address viral rumors, and Musk has said AI will eventually write most of the context on X. LLMs can be quite good at established facts and nuanced context, but they are not oracles with instant access to the truth. An agent may soon reach out to an election official to ask for corroboration as it writes a factcheck, potentially scaling up the process, but that isn’t happening yet. For now, AI still gets things wrong, which is why humans in the loop are important. And yet, humans are in the loop with Community Notes…and perhaps can’t get out of our partisan bubbles.

The polarization problem

The fact that human reviewers could not reach consensus on the Spencer Pratt “0/24,000 votes” rumor is alarming. “No Consensus” sounds like the outcome of insufficient evidence – it’s too difficult to tell! – but the evidence in this case was not ambiguous. County records, the AP’s own explanation of its feed, a federal prosecutor who looked and found nothing. What failed to converge wasn’t the facts...it was the raters.

The Community Notes bridging algorithm requires agreement between raters who usually disagree. It works very well for scams, crypto spam, AI-generated images, out-of-context videos, and miracle cures: claims with no partisan constituency. On claims where one political faction is motivated to believe the lie, however, commitment to disbelief may have the effect of veto power over what counts as established fact. If even Trump’s own prosecutor can’t convince raters that Pratt demonstrably got votes in that drop, the question is no longer one of evidence at all. That is a crisis of trust that evidence cannot resolve, because the distrust has become untethered from facts on the ground. An audience that has heard for over a decade that every referee is corrupt views everything with suspicion; the more sources that agree, the deeper the plot must go. No correction mechanism — crowdsourced, professional, or algorithmic — can reach people who have internalized that consensus itself is a conspiracy.

There are potentially some addressable issues with the bridging algorithm: the people who rate notes on a post are downstream of who saw the post, which can sometimes lead to challenges with not enough diverse raters. But there were millions of views on those posts; many people tried to rebut the claim in the main feed.

Meanwhile, many of the rumor’s amplifiers were monetized influencer accounts. Some were influencers paid by prediction markets to post election content. Kalshi eventually asked its influencers to delete posts sowing doubt about the election, which means the gambling platforms moved faster against election misinformation on X than X’s own integrity system did. Many of the biggest spreaders of viral nonsense on X are monetized accounts. Meanwhile, note writers and raters are not compensated. This seems backwards!

If this rumor had gone viral in 2020, or 2022, the posts would likely have been labeled. The top-down factcheck-and-label process of 2020 was imperfect, but it was not a censorious tyrannical regime — it was election deniers (Jim Jordan, Eric Schmitt, Donald Trump) who led the charge to spin it as such, working to frame labeling not as a form of platform counter-speech, or more speech in general, but as “censorship.” They actively sought the “nothing is true and everything is possible” contested reality that we have arrived at, because the online attention contest goes to whomever has the biggest megaphone and the most activated audience, and because they wanted to delegitimize the referees before the next round of lies needed refereeing.

The old system at least operated on a clear theory of adjudication: a claim about a suspicious ballot dump was checked against the count and the record, and labeled if it failed. That didn’t mean the people who saw it would be persuaded; many dismissed the fact-checkers as biased. Community Notes is a genuine answer to that legitimacy problem — context from the crowd rather than from distrusted referees. But what we see in the Pratt situation is that it may be answering a narrower question: not whether a claim is true but whether people who disagree about it can agree on a note.

2026 midterms ahead...

I support Community Notes as a concept, and as a complement to fact-checking – particularly because Notes usually cite professional fact-checkers. (It’s only antagonistically framed as an either-or product by Musk and Zuckerberg for political reasons.) Notes performs well on content where partisan constituencies haven’t dug in. And when notes do appear, they measurably reduce resharing of viral false tweets. The trouble is the gap between what the system is (a useful tool in a broader information integrity strategy) and how it is being marketed: as a one-stop solution.

Elon posted recently that people who engage with posts that are subsequently Noted would see a message in their DMs. That would be worthwhile – but Notes have to appear first. A meaningful share of the extremely-online right in the US believes the LA election was stolen — their influencers are saying so, and no fact-check of even the most obviously false claims is showing up within viewing distance of the posts.

The “24,000 votes” rumor was a one-minute data lag with a public paper trail and debunks spanning PolitiFact to Trump’s own DOJ. Sixty notes, and zero corrections shown; meanwhile, millions of views on the viral posts. If consensus was impossible on this one, community contextualization will struggle to achieve what we need it to in November.

Edit: A reader points out that one note did eventually clear! You can see it here if you filter by Helpful. It’s on a satirical/sarcastic Saad post, it cites the US Attny’s tweet as well as a CNN factcheck and LA Times article. The posts it’s quote-tweeting still do not have Notes.

His candidate plotline was made for the attention economy: nontraditional longshot breaks out by saying what They Don’t Want You To Hear. Compelling personal story: he tragically lost his house in the LA fires. Strong trolling game. Online clipper accounts went crazy for him – not entirely organically, but hey — it’s called the attention economy for a reason. The algorithm ate Pratt up.

However, the season finale is upon us, and it’s taking the form of a proliferation of predictable conspiracy theories promoted by people who should know better.

The LA mayoral primary’s devolution into conspiracy tropes, Polymarket screenshots, and Grok fact-checks is useful because it shows a script we are going to see again in 2026: viral candidate overperforms in the feed, ordinary ballot-counting reveals a different reality, and the people who conflated engagement with votes decide that there’s something suspicious about the count.

The majority illusion

Network scientists at USC explained it years ago: your social media feed is dominated by the loudest, best-connected nodes. Their opinions can feel like the majority consensus even when it’s not. You might be thinking you’re seeing the population, when you’re really seeing the loudest people. The term they gave for misjudging where prevailing opinion is based on what you see around you is called the majority illusion.

Pratt, the former reality-TV villain, ran a campaign tailor-made for clipping and virality. The personal-loss story. The visible dysfunction LA continues to struggle with. The “incompetent incumbent” frame he was able to wield against Bass. The algorithm loves righteous indignation; people like an underdog; lots of voters are already inclined to believe existing government officials are useless; and Pratt is unusually well-suited to the age of vertical video. His opponents were politicians. He was a content creator.

The attention economy dynamics were compounded by the fact that California is a national content object. California politics is insane! It attracts a massive amount of attention from people who are not Californians. When I lived in SF for a decade, I was very involved in local politics and constantly bemused by the national interest in our algebra wars, school board recall, DA recall, and House, Senate, and governor’s races. California looms large in ‘Dems In Disarray’ storylines, and Fox News is always on it. There are also enough California-based influencers and poasters that its political discourse seeps onto everyone else’s feeds, and then they have feelings.

And that’s where things can get weird. Ten years ago, I was strongly advocating for a bill to improve vaccination rates in California public schools and constantly getting yelled at by nutjobs online. The bill was polling around 70% approval in surveys by the Sacramento Bee and other local media. Online, the conversation was almost entirely negative. But the online conversation was shaped by lots of people (Texas antivaxxers, hello!) who didn’t live anywhere close to California. There were also bots.

The Pratt feed really was inescapable. In certain parts of X and TikTok, he was once again as much a fixture as he was on the gossip blogs during his reality-TV heyday. He has more than a million followers on X. A fan-made AI video casting Pratt as Batman and Karen Bass as the Joker pulled more than 5 million views; another, featuring pilates moms confessing they were voting for Pratt, did nearly two million. The campaign barely bothered with traditional paid media, leaning on clips and boosting fan-made content while his rivals spent on TV. The Extremely Online right-wing influencer machine boosted him (for better or worse; Steve Bannon presciently remarked that he would endorse him but didn’t want to hurt him in LA).

When your feed is wall-to-wall Pratt — clips, endorsements, memes, weird AI propaganda — it is easy to slide from “Pratt is popular online” to “Pratt is popular.” But a clip with three million views just tells you the clip did numbers online, and presumably a lot of people who liked it are nowhere near Los Angeles. The engagement number alone does not tell you much about how a city of millions will fill out their ballots.

I pulled up polling data to see where the numbers were as Pratt’s campaign went viral. Polling experts, please correct me in the comments if I’m wrong, but as far as I can tell, Pratt advanced from about 14% in the March UC Berkeley IGS/LA Times poll to 22% in their final survey. That final poll had Bass at 26% and Nithya Raman at 25%. The only survey that ever showed him ahead was a 400-person McLaughlin & Associates poll — a Republican pollster — that contradicted every other survey of the race but Breitbart touted it. As the clips spread online, Pratt’s unfavorability climbed from 28% in March, when most likely Angeleno voters had no opinion of him at all, to 57% by late May. (Raman’s recognition increased, and her favorability climbed.) I went back and looked at Zohran Mamdani, who also dominated social media on his way to actually winning New York; his online momentum did show up as positive trends in the polls.

But the extremely online MAGAsphere on X seem to have equated high online engagement with high odds that Spencer Pratt was going to win, or make it to a runoff. The former was always a serious long shot. This is Los Angeles! Pratt is a Republican candidate, with President Trump’s public blessing, running in a city where GOP registration is under 15%. The last Republican mayor of Los Angeles was elected in 1993, and left office in 2001. Most people in LA hate Trump. They may be very frustrated with Bass and very, very reluctantly voting for her again, but it’s unrealistic to believe they would elect Pratt in a landslide. A runoff seemed possible, but the third candidate had to be considered.

Red mirage, blue shift

So did they take the loss like adults? Well, consider this post by Florida Governor Ron DeSantis. Quote-tweeting a post about Pratt’s declining odds on Polymarket, he wrote that California “keeps dumping votes,” that the ballot drops “always seem to go one way,” and asked whether officials were counting “until you get the result you want.”

“Dumping votes”? Scheduled release of counted legal ballots, or things falling off the back of a truck? DeSantis, who is a governor, gets to gesture at fraud without having to produce any evidence of it. Why? Because most of the ballots are likely to go for Bass or Raman…which any competent prediction market bettor would ironically be aware of.

Here’s why the ballot drops “go one way”: Republican voters are more likely to vote in person. Democratic voters are more likely to vote by mail. California has lots of people who use vote-by-mail, which means it takes days to process ballots, verify signatures, cure problems, and update totals. As those later ballots are added, a Republican lead can shrink. Sometimes the order flips. This is one of the most predictable features of vote-counting in states with heavy mail voting. This phenomenon is commonly called the “red mirage, blue shift.”*

Every American should understand what is happening right now with the narratives surrounding the mail-in ballot count, because election deniers have learned to use ballot-counting timelines, and this very predictable shift, as a fraud-accusation template whenever their preferred candidate’s early lead shrinks.

On election night, Pratt was around 30% of the vote, roughly ten points and 40,000 votes ahead of Raman for the second runoff slot. By Friday evening, with about 71% of the expected vote counted, it was Bass 35%, Pratt 28.2%, Raman 24.9% — a lead of 20,672 votes, shrinking with every batch the Registrar released. The trend is appearing in the governor’s race on the same ballot: Steve Hilton, a Republican running for Governor, led on election night and watched Democratic challenger Xavier Becerra eat into his margin as mail ballots were processed. Becerra has now passed him.

In November 2022, in the LA mayoral race, Republican Rick Caruso led Karen Bass in the post-election count by about 12,000 votes — 51.25% to 48.75% — with many ballots still left to count. Bass overtook him days later as mail ballots were processed, and ultimately won by about nine points. Same city, same office, one cycle ago.

Election deniers have recently tried to reframe the fact that this pattern is predictable as some kind of plot in and of itself. Explaining its predictability is an effort to normalize election fraud, you see, because that is prebunking it, and prebunking is brainwashing, and we can’t prove that the ballots aren’t being manufactured or bussed in from China or whatever this time. Yes: as you read this newsletter explaining the basic predictable mechanics of online narratives and electoral systems, I am actually brainwashing you into accepting fraud as normal.

X has done quite a bit to feed the election rumor frenzy: there was a spate of misleading claims that a batch of 24,000 votes (the Notorious Searchable Statistic of this race) had been released, of which Pratt earned zero.

This turned out to be false – some screenshot had captured a moment before Pratt’s tally was updated. (AP told the LA Times that the two updates together actually included 21,870 votes for Pratt, 12,850 for Bass, and 9,521 for Raman.) Governor Gavin Newsom and an Assistant US attorney from the DOJ came out to debunk the claim. Grok had been confused in the early moments when the claim first began to circulate, but subsequently began to factcheck it correctly. Yet for some reason, no Community Notes showed up on tweets making the false 24,000 Pratt-Excluding Ballots claims.

Elon Musk, who used to vote by mail in California, obviously weighed in.

There is, to be clear, no evidence of fraud anywhere. Pratt continues to gain votes. He is just not gaining them at the same rate as his Democratic opponents.

Prediction markets figure in a fair number of bizarre claims this time around as well; they have updated with each release count, which is the point of having prediction markets. Somehow, some of the same people screaming fraud have decided that the market moving against Pratt is evidence that traders “know” he will lose because the race is rigged. This is a strange theory of prediction markets. The whole premise is that the market updates as new information comes in. Prices should cut through the cope. Red mirage/blue shift likelihood affecting the outcome should be priced into the contract, since it is how election behavior works and people are committing real money.

The cope

Pratt tapped real anger. LA is expensive, and a lot of angry, exhausted people are living with the consequences of fires and compounding institutional failure. A candidate who points a camera at visible dysfunction and says, “it shouldn’t be like this” will attract attention from people who appreciate that he is acknowledging what they are feeling.

But clips and memes aren’t ballots. Going viral doesn’t translate to getting elected. A fraud narrative can be a comforting bridge between those two things: it lets supporters continue to trust their perception. The feed wasn’t distorted, the count was. If everyone you saw online wanted Pratt to win and Pratt didn’t win, someone must have interfered.

As of Saturday evening, Pratt still led Raman by 7,494 votes, with more ballots left to count. Pratt himself has not claimed fraud. If Raman ultimately takes the second runoff spot, we will see what he does. California’s count really is slow, and wanting it faster is a legitimate complaint. But “California keeps dumping votes” is not the language preceding a process-improvement proposal.The blue shift in mail-in ballots isn’t an anomaly, it’s a highly predictable event – which means that at this point, grievance-grifters on X have pre-positioned themselves to cry fraud every time a completely normal, predictable thing happens. They will never be lacking for “evidence.” The broader script for the midterms is already clear: early returns will be treated as the “real” election; everything counted afterward will be treated as contamination.

The Hills ended with the camera pulling back to reveal the “street” was a backlot set all along. The sooner Americans understand that political influencers screaming “fraud” are performers too, the better.

*Election law scholar Edward Foley named the “blue shift“; the data firm Hawkfish named the “red mirage“ half in 2020, warning — correctly — that Trump’s election-night lead would be an artifact of counting order.

It will shock you to learn that right-wing media is reporting things differently. In the alternate universe inhabited by The Federalist et al, the settlement is proof of a “”Censorship Industrial Complex””. In the universe where the earth is round and vaccines don’t cause autism, what happened is more mundane: a plaintiff once again couldn’t prove the Censorship Complex theory in a courtroom, so the DOJ cited a Trump executive order as justification, and wrote a check.

Let’s break down why the Berenson settlement matters, because things like this are going to continue happening. People who have grievance narratives that are politically useful to Trump — or who have done egregious things directly in his support — will be getting government handouts of your tax dollars, so you should be paying attention. One story is finally breaking through: the Trump DOJ just stood up a $1.776 billion slush fund to pay out supposed victims of “weaponization” of the government. Who do you suppose will be collecting?

The third settlement in a series

Berenson’s settlement happened before the establishment of that fund, and his payment did not come from it. But it offers a preview of the same governing style: convert grievances into government-recognized claims, then use settlement machinery and public money (your tax dollars) to ratify them. It was also the third court case in quick succession, tied to the fantasy of a censorship complex, that was not proven in court but instead settled by the Trump DOJ.

The other two were Murthy v Missouri, and Daily Wire v. State Department (with The Federalist as another plaintiff). Most people haven’t heard of the latter; it argued that the State Department’s Global Engagement Center – tasked with countering foreign propaganda abroad – had given grants to, and promoted, counter-disinformation outfits’ foreign-focused work. Those outfits had separately rated the plaintiff outlets ‘unreliable’ in tools for advertisers. The plaintiffs’ ambitious theory was that funding or promoting some of the entities’ foreign work made the State Department responsible for the domestic ratings as well. No court tested this theory; the case survived a motion to dismiss, Trump came back to power, and it was subsequently settled.

I wrote about both consent decrees in Lawfare as they came down, because of a data void problem in the coverage: right-wing media got straight to reporting them as confirmation of the underlying claims. They weren’t. No findings of fact went into either settlement. Once the administration changed, the new defendant – the Trump DOJ – was incentivized to lose, because it was politically aligned with the plaintiffs. It, too, wants the Biden Censorship Regime to be a real thing. So it settled. The plaintiffs got to claim victory, a political win for MAGA writ large. And because mainstream media didn’t cover the decrees — treating them as the nonevents they normally would be — the partisan ecosystem set the narrative agenda, including for the AI answer engines that increasingly synthesize reality for everyone else.

The Facts of the Berenson Case

For those who’ve forgotten: Alex Berenson is a former New York Times reporter who became one of the most prominent COVID-vaccine skeptics on social media; The Atlantic famously dubbed him “The Pandemic’s Wrongest Man“. Twitter banned him in 2021. He sued the company, and as part of that suit’s discovery he obtained internal records; the ban was reversed and his account reinstated in 2022.

Here is where I want to be precise: Berenson’s jawboning record is fuller than anything in Murthy. The Murthy plaintiffs mostly lost at the Supreme Court on traceability — they couldn’t tie specific moderation to specific government demands. There was no evidence that the government defendants had so much as mentioned Jay Bhattacharya, for example. That’s not the case with Berenson. The Twitter records, as summarized in the judge’s opinion, show that the White House did single him out by name: at an April 2021 “Vaccine Misinfo Briefing,” officials including Rob Flaherty pressed Twitter on why Berenson “hadn’t been kicked off,” with the White House treating him as a sort of “ground zero” for vaccine misinformation.

But the same record also shows Twitter pushing back. The company told the White House that Berenson hadn’t violated its policies, and did not ban him. The permanent suspension didn’t come until late August — and the evidence closest to that ban runs not through a government official but through Scott Gottlieb, a former FDA commissioner under Trump (2017-2019) and a Pfizer board member, a private citizen by that point, who had sent Twitter’s lobbyist a “partner report” flagging a Berenson tweet that then apparently became the fifth strike. The morning after the ban, documents show Twitter’s senior executives saying that it shouldn’t have happened.

So even in what was relatively the strongest of the censorship complex cases, the record establishes Twitter resisting direct government pressure, and then possibly acting on the Gottlieb complaint — this was the causal link gap that a judge never had to resolve. Berenson is currently also suing Gottlieb and Pfizer CEO Albert Bourla under 42 U.S.C. § 1985(3) — a Reconstruction-era statute creating a cause of action against private conspiracies to deprive people of federally protected rights based on membership in a protected class. He’s arguing that unvaccinated people are a class, and Gottlieb and Bourla conspired with the government.

Why the case was dismissed — and what the settlement skips over

Berenson’s case had been dismissed in July and September 2025: first the judge dismissed the private defendants because, among other things, unvaccinated people are not a class, and later the federal defendants on standing because they were no longer in office and there was no available remedy. (The judge did not dismiss for lack of evidence on the standing claim).

But the Trump DOJ chose to settle the case, to give Berenson $150,000, and to recite language that partisan media framed as the government admitting it censored him. In reality, it says this:

Read it carefully. The first whereas recites what “Plaintiff asserted in the litigation” – what Berenson said in his complaint about being censored. The second recites that the federal defendants “recognize” that a formal Executive Order issued by President Trump has “acknowledged and conceded” that speech “like Plaintiff’s” was stifled. The agreement then spends two full pages quoting that executive order, and follows it with Trump’s July 21, 2025 “Promises Made, Promises Kept” statement — in which the President was, per the settlement’s language, “rightly and deservedly taking pride” in keeping his campaign promise to “bust up the censorship regime.”

The Free Press called the settlement a “huge victory for free speech,” and asked “Why hasn’t the mainstream media noticed?”

There is no actual independent finding there, based on the record, that the Biden government censored Berenson. There is the Trump administration reciting its own executive order and calling it a concession. Executive orders say whatever the president wants them to say. EOs are political documents. Remember the decree going after CISA head Chris Krebs for “falsely and baselessly den[ying] that the 2020 election was rigged and stolen”? Yeah. Trump isn’t God, he doesn’t speak reality into existence through the Word. Saying “My EO is the evidence that Biden censored you” is not a serious claim.

Now, The Free Press is not serious media, it’s agenda-driven advocacy journalism headed in the direction of becoming The Federalist. The right-wing media ecosystem constantly spins to support this narrative in particular, because the fact that none of these cases could pass muster in court is pretty inconvenient for the censorship hoaxers. TFP did get one thing right, though: mainstream media should be reporting on the cases. It should be reporting the losses, and what they reflect about the hollowness of 3 years of weaponized investigations by Jim Jordan. It should be reporting on the political and financial gifts, which establish a narrative the movement couldn’t otherwise. Murthy v Missouri was an embarrassing loss! Sen. Eric Schmitt has struggled to cope. If mainstream media wanted to, it could report on the work that orgs like FIRE are doing to tackle jawboning through legislation to mandate transparency in government requests to tech companies – which I support.

For the record, since this post will make my online screamer fandom big mad: I don’t think Berenson should have lost his account. As I’ve long said, takedowns mostly just turn cranks into martyrs, and Berenson promptly just set up shop on Substack and monetized anyway. I think labeling, Community Notes, user-controlled feeds, protocols, and speech-not-reach policies (centralized platform editorial speech) are solid interventions when transparently deployed. It’s also worth noting that Berenson has been critical of Trump a handful of times; what we’re going to see with the $1.776B fund is going to be fully off the rails, “for my friends, everything” mass grift.

But to be clear: the Berenson settlement does not say that a Court found the Biden administration violated the First Amendment. It says “the current Trump government has taken the position that violations occurred.” This is not a merits ruling after fact-finding. It creates no binding precedent, and it’s hard to imagine it’s going to do much for Berenson’s ongoing conspiracy case against Bourla and Gottlieb. It is just the administration incorporating a political position into legal language.

And it came with a $150,000 government handout.

Two weeks ago, The Free Press published an essay by Tablet Magazine editor Jacob Siegel alleging that I’d gotten a review of his book in The Baffler censored. He promoted it aggressively on X, claiming “a figure connected to the US government pressures a publication to remove its review of my book,” and that it was “censored the next day... apparently at the behest of Renee Diresta.” Sympathetic influencers piled on.

The claim was false. I sent The Free Press the email exchange showing that I had asked for a correction, and the reviewer said he had asked for the review to be taken down.

The false accusations and wild innuendo ran under the headline: “I Wrote a Book About Censorship. Then People Tried to Censor It.“ The Free Press has seen the emails that destroy the premise of the piece. But they’ve done nothing to acknowledge their existence to their readers, or to update the piece with so much as an Editor’s Note.

How We Got Here

Siegel’s book, The Information State, is the Twitter Files crossed with a Glenn Beck chalkboard: a string of insinuations combined with just enough political theory to dress up paranoia as scholarship. CIA Renee is a character in the book, leading “the largest public-private social media monitoring and censorship initiative in existence” behalf of the Biden Regime that totally ran the government in 2020.

Some reviewers took Siegel’s innuendo at face value. They wrote Twitter Files lies about how the Election Integrity Partnership plotted to censor 22 million tweets, or the Virality Project ‘demanded the censorship of true stories of vaccine side effects’ as fact in their reviews. It’s not their fault they were deceived, but I asked for corrections because these lies are indefensible at this point, and I want these outlets to acknowledge reality to their readers. The Free Beacon promptly corrected and offered me space to respond. The Brownstone Institute did nothing. The Baffler pulled their review – at the reviewer’s request, per his email. And Siegel lost his mind.

The Free Press gave Siegel ~2,000 words to turn a correction request into a persecution narrative. Their newsletter promoted it to a million readers, asking, “Is the age of censorship really over?”, and claiming that Siegel’s book “was subjected to the very dynamic he describes in its pages” when a review was “purged” from The Baffler “after a person mentioned in the text requested a correction.”

The “investigation” underlying this article consisted of Jacob Siegel sending me a list of demanding questions with 102 minutes to respond. He requested my private correspondence with The Baffler, which was inappropriate so I refused to indulge him. Siegel then wrote that I “didn’t deny” asking for the review to be pulled. The story ran hours later.

I tracked down an editor at The Free Press and told her the story was nonsense. I asked why it hadn’t been fact-checked, and was told that Siegel’s email had been the fact-check. “Yeah, we were back and forth on the whole comment request,” the editor said a bit sheepishly, when I expressed amazement. I denied the claim by phone and email, then forwarded my actual correction request as proof.

She offered me a letter-to-the-editor to respond; I said I thought the accusations deserved more than a few paragraphs stuck in the far reaches of their site, and was told that they would promote it. But, she warned: it would be edited and fact-checked.

Nearly three days later they inserted a parenthetical sentence with my denial into the body of Siegel’s prose, below the paywall. No formal editor’s note. As a contrast, here’s what an Editor’s Note on a different disputed story from the same timeframe looks like.

On April 5, I additionally sent them the reviewer’s email saying he had requested the takedown. Still, no editor’s note.

I didn’t publish this exculpatory email immediately because I didn’t want to drag the reviewer into the online pile-on that TFP story predictably generated. He didn’t ask to become a character in Siegel’s self-mythologizing censorship melodrama.

The “Very Worthy Letter”

I drafted a Letter to the Editor laying out the facts: I had requested a correction related to demonstrably false claims about our work — not a takedown. A correction request is counterspeech, not “censorship.” I asked The Free Press to update the article, and to perhaps not use the accuser as the fact-checker in the future.

What came back was not an edit. It was a containment effort.

Where I had stated facts directly, The Free Press inserted distance. Declarative statements received “I believe” qualifiers. Siegel was allowed to write that I was “either wrong or deliberately misrepresenting” his work. He didn’t have to “believe” that I was “misled.” He could speculate about what I thought, what I did, and what shadowy ecosystem I supposedly represented. But when I responded, suddenly everything had to be hedged.

Adjectives disappeared: Siegel hadn’t sent me a demanding email requesting my private communications – just “an email.”

References to The Free Press’ responsibility — including their lack of acknowledgement of the reviewer’s email and how it blew up the theory — were pared back or removed.

But this was not simply legal CYA—opinion sentences, like my “belief” that correction requests are counterspeech—not censorship—were cut.

On April 9, after two rounds of this, I wrote to the editor: “I see more adjectives removed, declarative framing sentences cut, and a lot of ‘I believes’ inserted into extremely defensible matters of fact. I’m trying one final time here.”

What came back was remarkable. They had sent me “a very worthy letter to the editor” that gave me “significant latitude to respond to the piece,” I was told.

I asked, directly: “But why aren’t you updating the piece with a note clearly reflecting that the man withdrew his own review? Is it because it undermines the headline’s premise, ‘Then People Tried to Censor It’?”

They ignored the question.

My managed letter went up on Friday—without clear attribution as a right-to-reply, midway down the page as just a link that says “Letter to the editor”. It could have come from anyone. They sent it out as the bottom item in the Monday newsletter…no editorial framing.

A Free Press for Free People

The Free Press has every right to edit my letter and exercise editorial discretion, of course; I chose the word “censorship” in the headline to mock my accuser’s commitment to rendering that term meaningless, and TFP’s willingness to toss it around just as baselessly. When I pushed back one final time and said some things were non-negotiable, I did get them into print. But I wrote this post because TFP protected an ideologically aligned writer while slow-walking, hedging, and ultimately ignoring exculpatory evidence.

Here is The Free Press’s values page:

“We seek and report the truth. We tell it plainly when we uncover it, even when it’s politically inconvenient.”

“We assume good faith. We treat each other and our sources with the utmost respect.”

“We are adults, and we treat our readers like adults, too. We encourage each other to think for ourselves and change our minds when we encounter new information.”

And here is the timeline: On April 1, a Free Press editor read evidence that the premise of Siegel’s story was wrong. On April 5, I sent the reviewer’s email saying he himself asked for the takedown. On April 9, I asked the editor directly why the article still hadn’t been updated with this information and they ignored the question. As of April 14, the article still shows no indication that the smear was completely baseless.

Consider how TFP circled the wagons to protect their narrative, then decide for yourself what those values are actually worth.

For years, the loudest political argument about social media has focused on “censorship.” The term became a heavily-leveraged thought-terminating cliche: moderation is censorship, labels are censorship, transparency is censorship, government criticism of viral rumors is censorship, not nudifying women and children is censorship, researchers studying platforms are censorship. It was less a legal argument than a power word—a way of making any attempt to mitigate or even describe platform harms sound illegitimate. Election deniers like Jim Jordan, Stephen Miller, and then-Attorney General of Missouri Eric Schmitt used it to great effect.

This week, courts illustrated the problem with that framing from a few different angles, and the Meta Oversight Board also tried to inject nuance into the conversation. Let’s go over the whirlwind set of decisions…

The “Censorship” Cases: Narrative Wins Over Legal Substance

Elon Musk’s GARM antitrust lawsuit was dismissed: “Censorship industrial complex” cases have been winding through the courts for several years now. Elon Musk’s argument that advertisers boycotting X amounted to an antitrust conspiracy was roundly rejected by a Texas judge, who dismissed the case against the Global Alliance for Responsible Media with prejudice. But GARM had already dissolved under the political pressure nearly two years earlier, when the World Federation of Advertisers said the allegations had “significantly drained its resources and finances.” The legal vindication arrived long after the lawsuit had achieved its actual objective.

Murthy v. Missouri got a consent decree: Eric Schmitt, now a senator, previously the attorney general of Missouri, eked out a consent decree in a case that SCOTUS had tossed back to lower courts because the plaintiffs lacked standing (Justice Amy Coney Barrett also noted “clearly erroneous” lower court findings). The Trump Administration chose to settle; Schmitt framed this gift from his political allies as a “historic” First Amendment victory. His press release declared that the settlement proves the federal government censored Americans’ speech and that Missouri “won big.”

Schmitt struggles with honest representations of reality; the actual decree says nothing of the sort. It is a settlement reached “without further litigation” and explicitly states that it “shall not be construed as evidence or as an admission” of the allegations. It binds only the Surgeon General, CDC, and CISA. Relief is limited to the plaintiffs’ own social media content. It explicitly preserves the government’s ability to provide information to platforms and to express that posts are inaccurate or contrary to the administration’s views, so long as those statements are not coupled with a threat of punishment. And it is enforceable only by the parties or their successors; no one else gets to wield it as some universal charter of online speech rights.

So what, exactly, is Schmitt celebrating? A participation trophy. When litigation functions as narrative warfare, the filing itself is the win. Right-wing media and the Twitter Files boys quickly knocked out victory-lap stories, most of which incorrectly attributed the (bad) Hunter Biden laptop moderation call and 2020 election rumor moderation to the Biden Censorship Regime—before Biden had even been elected, when Trump appointees ran the government. “Who was president in 2020?” remains the question of our time.

Platform Design on Trial: Two Jury Verdicts Focus on Harm, Not Speech

Meta, meanwhile, lost two cases that highlighted the fact that there are questions about social media’s impact that are not about user speech at all.

In Los Angeles, a jury found Meta and Google liable in a youth social media addiction case, awarding $6 million in damages after finding negligence and failure to warn. The plaintiff’s case focused on platform design rather than content: features like infinite scroll, autoplay, and intermittent reinforcement through notifications.

In New Mexico, a jury found Meta violated state law by misleading users about the safety of its products. The New Mexico Department of Justice said the evidence, which came in part from an operation it ran called Operation MetaPhile, showed that Meta’s design features enabled predators to engage in child sexual exploitation, that it knew this, and chose to ignore it while promoting the platform as safe for kids. The jury ordered Meta to pay $375 million.

These cases are not about whether a label (adding more speech!) is actually “suppression,” or whether a takedown request is jawboning. They are about something more concrete and, for Meta, more dangerous: whether the company designed harmful products while telling the public something much more reassuring than its own internal evidence supported.

Meta’s Oversight Board on Community Notes: Caution Over Enthusiasm

Finally, today, we also got a policy advisory opinion from Meta Oversight Board on another product-related question: Meta had asked the Oversight Board for guidance on how to expand its Community Notes outside the United States. The company had already ended its U.S. third-party fact-checking program in January 2025 with a political statement from Mark Zuckerberg about how biased the fact-checkers were; Community Notes was the replacement. When it went to the Board asking about how it should evaluate where it was appropriate to roll out globally, Meta described the program as still in an “early stage of product development,” with only “limited data from the U.S. beta rollout.” That was an understatement: where Twitter’s Birdwatch program launched with full transparency about the number of participants, the text of the notes, and what “cleared” and showed up attached to misleading posts, Meta has revealed very little.

The Board’s answer was revealing. Yes, it said, Community Notes could enhance expression and improve discourse if they operate with enough scale, speed, and safeguards. But it also warned that in repressive regimes, during elections, and in crises and conflicts, the system could create serious risks rather than mitigate them. It warned that coordinated disinformation networks could game the process. It warned that dominant political, ethnic, or linguistic groups could drown out minorities. And, most notably, it said that delays in note publication, the small number of published notes, and dependence on the reliability of the surrounding information environment raise “serious doubts” about whether Notes can meaningfully address harmful misinformation.

I am a supporter of Community Notes in general. The bridging-based consensus requirement — the idea that a note only surfaces when people who usually disagree converge on finding it helpful — increases the perception of legitimacy, particularly among those who distrust centralized fact-checkers. Research shows that notes, once attached, do reduce engagement with labeled content, sometimes substantially. The open-source algorithm and public dataset that X built the program on gave it an unusual foundation of auditability. And LLMs show significant promise as Notes contributors, something X allows while other platforms still do not.

But we need to be operating in the realm of reality about where the program stands. The bridging requirement that makes Community Notes credible also means that the most controversial content—the stuff where notes would matter most—is the least likely to get a note at all. The system is vulnerable to coordinated manipulation, such as organized campaigns to mass-downrate accurate notes. Note authors rely heavily on links to media organizations, Wikipedia, and—ironically—fact-checkers for their sourcing. Rolling out Community Notes on Meta in lieu of fact-checking partnerships risks making the former less effective by starving contributors of the very links they depend on.

And then there’s the question of who does this work and why. Meta spent over $100 million on professional fact-checking partnerships since 2016. Ending the program and shifting to Community Notes doesn’t make that work go away; it simply reassigns it to volunteers. Researchers have described this as “data labor”—unpaid annotation that directly improves a platform’s products while the platform simultaneously invests in creator monetization and AI-generated content that generates the very misinformation contributors are being asked to correct for free. I’ve been participating in Meta’s program as a note rater and writer. Anecdotally, notes take days to clear, and most of what I’ve been voting on doesn’t really need a note — they’re being used for dunking. It’s hard to imagine many people feeling inclined to write Notes for free for a massive company that is doing this in part to save money. This doesn’t mean the program should be scrapped, but it does mean that incentives need to be better aligned.

Alexios Mantzarlis and I raised many of these concerns in a joint public comment to the Oversight Board when the original call went out. In it, we argued that Meta had asked the Board to bless the international expansion of Community Notes while explicitly telling the Board not to consider product design or how the algorithm operates — the very things that determine whether the system is safe, effective, and compatible with what Meta itself has framed as its human-rights responsibilities. Meta had provided no statistics or data from nine months of operation to help evaluate efficacy. A country-selection framework that does not consider how the product actually works, we wrote, would be rubber-stamping, not oversight.

Meta replaced one system with another, then asked the Oversight Board to help define the conditions under which the replacement might be too weak, too slow, too manipulable, or too context-blind to do the job. Their answer, in essence, was: quite a lot of conditions, actually.

What These Rulings Collectively Reveal

So: two clusters of decisions, all in a span of three days. One addressed the endless theatrical fight about “censorship,” where every act of moderation (or even counterspeech!) is tyrannical authoritarianism, and where a narrow consent decree becomes a triumphal monument to a case the Supreme Court declined to resolve on the merits.

The other cluster interrogated questions about platform products: Two juries said that when harm to users comes from design choices—platform conduct, not content—“everything is speech” is no longer a defense for evading accountability. And Meta’s Oversight Board warned that Community Notes may be too fragile to serve as the company’s primary response to misinformation at this point.

There are thousands of product-liability cases in the queue. Is the terrain shifting? It’s too early to tell; Meta will obviously appeal these verdicts. But the question these juries are now asking is one that censorship entrepreneurs have spent years trying to crowd out: what responsibility do platforms bear for the systems they build, and the risks they understand but choose not to prevent?

Other recent personal things:

In a book review for Lawfare, I delved into the alternate universe of Senator Schmitt (timing was purely coincidental!) It’s worth reading to see just how brazenly he lies about the depositions of Anthony Fauci and FBI Agent Elvis Chan. These things used to be disqualifying for political figures.

I joined Jon Stewart on The Weekly Show podcast (YouTube video here), with Casey Newton, to talk about why there is a perception of mass voter fraud on social media, even though the actual evidence—even from Heritage Foundation! —shows that it’s an almost nonexistent problem.

Cenk wasn’t the only media elite to have a conspiratorial revelation recently. Two weeks earlier, Cenk had sat down with Tucker Carlson to discuss “Epstein, JFK, 9-11, Israel’s Terrorism and the Consequences of Opposing It.” On the same day the episode was released, the Department of Justice dropped over 3 million pages of Epstein-related documents. Tucker released another episode of his show with conspiracy influencer Ian Carroll declaring that “It looks like Pizzagate is basically real.” Wayfair trafficking rumors suddenly popped up again-–not just among QAnon adherents, from whence they’d emerged in 2020, but among a broad spectrum of people (including left-leaning posters on Threads).

Pizzagate, Wayfair, 9/11 truth mixed up with Satanic panic, deep state cover-ups, vaccine plots, JFK Jr’s return—these claims used to travel together under QAnon’s banner, all wired into one grand everything-is-connected narrative by a shadowy central figure issuing drops. Today, the omniconspiracy is off and running without the Q branding and without the stigma. Nobody’s orchestrating it. It’s pulling in audiences on the right and the left. So it’s worth understanding why some of these conspiracy theories are re-emerging, and how the rhetoric in a declaration like Cenk’s works.

A Scandal as Raw Material

On January 30, the Department of Justice released over 3 million pages of Epstein-related documents—emails, videos, images, financial records—in response to the Epstein Files Transparency Act. The dump was massive, poorly organized, and riddled with both redactions and redaction failures. It was, in other words, the perfect raw material for online sleuths and conspiracy entrepreneurs alike: a vast, chaotic trove of real horror, released without context, into an attention economy that rewards big reveals. People went shopping for receipts to fit their longstanding theories. Sometimes literally: screenshots of a Wayfair purchase receipt began circulating. They searched for familiar brands and familiar keywords (pizza!), clipped suggestive fragments, and presented them to their audiences as vindication of past theories: this is what we meant all along.

The theories that resurface aren’t random—they’re the ones that already had audiences and narrative hooks in place, waiting for new “evidence.” Wayfair and Pizzagate are both elaborate theories about elite child exploitation, so a massive document dump about an actual convicted child sex trafficker was always going to serve as new proof for people who were open to being convinced. The new details don’t have to add up—they just have to feel damning. Sensational interpretations spread while debunkings that highlight the actual facts get dismissed as “cover-ups” or “excuses”, and one real egregious institutional failure becomes permission to doubt everything.

In 2020, the Wayfair conspiracy theory began when a QAnon influencer alleged that the online furniture retailer was trafficking children through purportedly suspicious overpriced product listings whose names matched those of missing kids. The theory was thoroughly debunked at the time; some of the supposedly missing kids (some of whom had been runaways) even made videos asking the sleuths to please leave them alone. Now, in 2026, a viral post showing an authentic receipt from the files was being hailed as allegedly proving the whole thing: one of Epstein’s associates had purchased items totaling $8,453. The post framed this as “a single, unlabeled” item—mysterious, inexplicable, damning. The actual documents, which journalists painstakingly reconstructed from adjacent emails in the files drop, show it was 25 separate products—bathroom decor, furniture, and 14 outdoor light fixtures—shipped to Epstein’s island to furnish a house. But this literal receipt isn’t being used as evidence in the traditional sense, as in, material that helps work toward a conclusion. It’s an artifact that someone can hold up and say, explain this—and then the burden of proof shifts to the person who says it’s just light fixtures. The conspiracy theory is hitting new audiences in the context of an actual scandal. Compounding this, the existing true believers are unlikely to trust the media doing the debunking.

Pizzagate, which falsely alleged that Democrats were running a child trafficking ring out of a Washington, D.C. pizzeria, predated QAnon. It was considered incredibly fringe and kooky when it first emerged, but was gradually incorporated into the Q mythology. Since the Epstein document release, approximately 910,000 posts on X have mentioned Pizzagate, nearly three times the volume of the entire six months prior. The word “pizza” appears in roughly 900 of the 3 million Epstein documents, and people immediately seized on this as proof that “pizza” was always a coded reference to child sex trafficking.

But journalists who have read through all of the pizza references in the Epstein files make the very compelling case that they are actually pretty mundane: people are discussing Manhattan pizzerias, Williams Sonoma pizza stones, youth soccer pizza parties. One email that went viral featured the subject line “U9 Red pizza party”—which conspiracy posters claimed meant the blood of children under 9. It referenced a kids’ soccer team. Meanwhile, Comet Ping Pong, the D.C. restaurant at the center of the original theory appears in exactly two documents—a PDF of an old Atlantic article about the conspiracy itself, and an anonymous FBI tip from 2020. John Podesta, whose emails were the fodder for the original theory (pushed by Jack Posobiec, now himself an elite influencer), appears in passing in only one.

The entire premise of Pizzagate is that elites use coded language to conceal their crimes. Were they? Maybe, though “pizza” sure seems like a stretch. In reality, what’s striking about the files is that Epstein and his associates were very openly referring to young girls as young girls in his emails. One compared a girl to Lolita. Peter Attia, longevity influencer, joked about ‘low-carb pussy.’ Donald Trump drew him a birthday picture shaped like a young naked woman. Epstein himself wrote, to a redacted person, “Your littlest girl was a little naughty.” The Epstein class didn’t need code words. They had impunity.

Ok but why, you might be wondering, is 9/11 trutherism back?

Let’s get back to Cenk.

Moments of real institutional failure don’t just produce anger—they often produce free-floating suspicion: What else have they been lying about? The Epstein files offer incontrovertible evidence that powerful people are protected, that accountability is optional for elites. In 2020, that suspicion often came packaged as QAnon, which had child exploitation as the moral anchor but took people on a journey into a broad story of institutional deceit where every scandal was proof of the next. Research on conspiracy belief is pretty consistent that people who buy into one conspiracy are more likely to buy into others—even contradictory ones—because the underlying belief is that someone is hiding the truth.

9/11 trutherism is a readily-available conspiracy theory with that same premise. It comes with several “explain this” moments, and for Cenk it perhaps neatly intersects with some personal beliefs about US imperialism and Israel. But what makes Cenk’s posts interesting is that they’re a conversion announcement, in public, from a fairly high-status figure on the left. So let’s dissect his thread itself.

Cenk’s Great Awakening

Cenk’s thread is a tight case study in how conspiracy thinking travels now: not as a coherent alternative account of a situation, but as rhetoric that turns distrust and nihilism into the “reasonable” stance. It’s conspiracy without the theory.

He starts with a confession: “I no longer believe…” This is a high-status account publicly declaring a change in position, a courageous awakening. He’s broken free and is inviting followers to update their beliefs alongside him. I used to be like you. I was asleep. Now I’m not. Come with me. (Tucker using “I was wrong” about his change of heart on Pizzagate has the same rhetorical effect.)

Then Cenk drops an “evidence collage” featuring two anomalies—paper passports surviving the crash, WTC7 collapsing. This is fairly flimsy stuff! It’s not actually offering an alternative explanation for 9/11. It’s just “How could this be?”-type questions, the physics of which are extensively discussed in many places. But he declares that these facts are incompatible with the “official story”...

And then quickly pivots to insinuation: “It’s time to look for who gained from 9/11.” Once you’re in cui bono territory, you can speculate forever. This is a classic conspiracy theorist rhetorical move because it’s unfalsifiable; there’s no claim to disprove. Someone should really look into this. What? What they’re hiding from you.

Finally, there’s the preemptive delegitimization of rebuttal: “To my normie friends who’ll be outraged by this, we have been brainwashed by corporate media all our lives.” This preemptively turns criticism into evidence that the critic is captured. The more people push back against Cenk, the more it proves the conspiracy is real. He reframes the term “conspiracy theory” similarly: it’s not a category of narrative built on bad inference, it’s a smear imposed to keep us from questioning power. The net effect is to make the audience feel powerful for rejecting the very idea of being corrected.

“Why am I questioning this now?” he asks in closing. Cenk’s conversion to 9/11 truther isn’t based on any information actually related to 9/11—it’s the transitive effect of Epstein. “You’d have to be an idiot to trust the US government now.”

Why This Matters

It’s not clear whether conspiracy theory beliefs are actually more prevalent today than they’ve been in the past; polling research has suggested that prevalence is largely holding steady, but the theories are becoming more visible. They certainly seem to be more normalized among new media elites, which is itself a phenomenon. When Cenk Uygur, a trusted elite media figure with millions of viewers across his channels, announces “I no longer believe the official 9/11 story,” it doesn’t just state his belief, it potentially shifts that belief’s acceptability. It tells followers that this isn’t embarrassing—it’s brave. It signals to adjacent creators that the belief is no longer a marker of crankery. Tucker’s “Pizzagate is basically real” commentary is doing the same thing. Once high-status people endorse a theory, the thinness of the evidence matters less because reputation supplies the credibility that the argument lacks.

Derek Thompson wrote after quote-tweeting Cenk, “I worry that the social penalty of announcing that you believe in nonsense is going down.” It absolutely is. For a lot of big influencers, professing belief in what brainwashed mainstream institutionalists call “conspiracy theories” has become a reliable way to signal independence, attract attention, and harvest engagement. There’s no reputational cost at all; quite the contrary. You can rack up Substack subs as a delusional nutjob and earn a very nice living.

Meanwhile, pushing back has gotten harder. When the Wayfair conspiracy went viral in 2020, social media platforms and journalists worked to debunk it because the harms to the kids and employees named were obvious. In 2026, some companies are afraid of being accused of “censorship” for factchecking; others saw an opportunity to reduce costs while kowtowing to the administration. So the claims are simply spreading to new audiences as recommendation engines curate posts into people’s feeds. Community Notes, which I like and support, is just not keeping up, particularly on Meta properties. This isn’t great. Widespread belief in baseless theories like Wayfair or Pizzagate is corrosive. QAnon was essentially a cult; people seem to have forgotten the stories of family members becoming unrecognizable.

Meanwhile, the Epstein files contain genuinely damning material suggesting decades of elite impunity, institutional failure, and the abuse of vulnerable young women. That’s real, and it matters. Going down some 9/11 rabbit hole takes the legitimate rage the Epstein case should produce and converts it into tangential content. Conspiracizing builds audiences, not coalitions for accountability. And it leaves the public conversation stuck on whether “pizza” is a code word rather than on why so few people have been prosecuted.

“Who has the power to define the truth?” is a great TV prompt. It’s maybe a little abstract as the frame for a debate about content moderation and free speech, but the instinct in how you answer it reveals something right off the bat. You can treat it as a normative question: who ought to have that power. But if you see it as a descriptive question the answer is pretty clear: the platforms do. They may not explicitly define truth, but they influence what people believe is real. They shape what billions of people see through daily decisions about amplification, labeling, downranking, and removal—what trends, what stays visible, what feels mainstream, what seems fringe. Governments know this, which is why they spend so much energy working the refs.

The difference in instinct might’ve been the main gap between Glenn and I. We agree on the risks of concentrated platform power, and we share a preference for decentralization. But I kept wanting to drag the conversation back to the system we actually have. When platform executives say, “We don’t want to be arbiters of truth,” I hear deflection, not lofty principle. Great—I don’t want them to be arbiters of truth either. But they own the distribution infrastructure, and they exercise curatorial and moderation power. There is no neutral, and they’re constantly making very opaque ranking decisions to keep users on site. It would be nice if they were more transparent about what they are arbiters of.

One of the recurring moves in the debate about social media and moderation is to speak in sweeping moral terms—anti-censorship, freedom, truth—and then, when pressed on how to enshrine these things for the greatest number of users in policy or code, to retreat into vagueness: I’m just a journalist, I’m just asking questions, I’m just defending free speech. That posture is comfortable, but it’s also hollow. It lets one side avoid the messiness of hard calls in the real world. The philosophical debate becomes a distraction from the governance issue.

So my goal for the debate was to get past vibes and into the practical:

• What responsibility should platforms have—if any—for addressing things like networks of fake accounts? Viral rumors that have real-world impact?

• How should we think about tradeoffs in moderation and speech, particularly where things like harassment are concerned?

• How should we distinguish between legitimate government persuasion and coercion if a government reaches out to a platform? What mechanisms should we put in place?

Where we disagreed

Glenn and I have, in a sense, both done work exposing hidden power. He’s focused for a long time on government surveillance. I’ve spent the last decade studying inauthentic manipulation campaigns, with a focus on state actors and exposé work spanning five continents. That is, until studying disinformation campaigns was made politically controversial—conveniently for manipulators.

Glenn’s key concern is that the term “disinformation” can be used to suppress dissent. For him, it’s a label people apply to frame contested claims as false, including opinions, often to delegitimize political enemies. I understand the point; the term can get politicized. But it remains useful for what it referred to historically, which is real and hasn’t gone away: coordinated campaigns involving covert deception, where the actors, behaviors, or networks are inauthentic, and where the content is often some form of propaganda. I brought up a campaign run by the Pentagon (part of a broader network that my former team at Stanford Internet Observatory helped expose) that at one point targeted Filipino citizens with claims about the Sinovac vaccine. These were fake accounts behaving in obviously coordinated ways to manufacture the appearance of consensus; that’s covert manipulation regardless of whether any single individual post by a fake account happened to contain something true. The politicization of “disinformation” hard in the other direction—as a word written in scare quotes—has resulted in USG and social media companies cutting the capacity to respond to such campaigns.

A second disagreement was about platform speech and censorship. Platforms have First Amendment rights to moderate and curate. They suppress spam. Many ban legal pornography; X does not. Rumble bans antifa content, per its Terms of Service. That’s unique to Rumble! I think it’s actually good that different platforms offer different experiences. Glenn argued a “platform monopoly” theory: that the biggest platforms should be treated more like common carriers, with must-carry obligations akin to a telephone company.

There are people in tech policy sympathetic to that view. But a must-carry regime (which courts have recently viewed skeptically under the First Amendment) doesn’t address issues with deciding what to amplify or how to handle harassment, and it isn’t clear what it means for inauthentic behavior. With more platforms emerging, I’d rather lower switching costs and expand user choice than turn private companies into state-mandated carriers of speech.

The debate didn’t focus entirely on wonky stuff: we argued over whether the “Censorship Industrial Complex” theory accurately describes what happened during COVID and the election (no!); what lawsuits about platform speech suppression actually revealed; and Meta’s (stupid) lab leak policy. Siva and Glenn got into a few fights.

Where we agreed

There was more common ground between the three of us than most people might expect.

We all want transparency, and recognize that we have less of it now than we did a few years ago.

We’re all uncomfortable with concentrated platform power, and see decentralized protocol-based options as a powerful option to give users more control.

We even partially agreed on some of the “censorship” questions, like that moderating harassment is not all bad, and that there is a difference between labeling content and removing it. Glenn expressed concern about who does the labeling (quis custodiet ipsos custodes?) but does concede that adding context is different from taking things down.

My positions, for the record

I sometimes get flack for saying I like debates, but I think they’re worthwhile because the listener hears both points of view at once. They’re most valuable when they aren’t litigating basic facts, but discussing a complex issue. And after a few years of watching bad-faith actors caricature my views in paywalled newsletters, I appreciate that the “other side” gets to encounter what I actually believe. 😉 Which is that:

• Moderation isn’t “censorship”; particularly because things like labeling add more speech. Governance keeps platforms usable...ask your friend how much time they spend on unmoderated platforms. Also, platforms have a First Amendment right to do it—but the best policies are transparent and maximize user agency.

• Opaque, concentrated platform power is a problem whether the CEO is sympathetic to you or not. Platforms should be auditable and escapable, and we should have many options.

• For political content, takedowns often backfire. Labels, counterspeech, friction and design tweaks work better than turning content into forbidden knowledge.

• Government has a 1A right to speak to platforms. The red line is coercion, and transparency is how you tell the difference; FIRE has good legislation to require it.

• We need more user control across the ecosystem—protocols, interoperability, and middleware to deliver real choice. Community Notes is a good example of something that works but could still be improved on.

• Disinformation campaigns are real and foreign actors interfere in elections. The question isn’t just “did this swing an election” (usually not, but it matters anyway!) It impacts trust and incentives. Our current government has sought to gut all response capacity to address it, public and private, in pursuit of retribution for a fantasy that the 2020 election was stolen. That is unambiguously bad.

• And finally: “do nothing” isn’t neutrality in a system where something is always being algorithmically amplified. Governance by default just rewards whoever games the system best.

So if you care about how speech, power, and platforms collide, watch here. Or, if you’re snowed in and want the full 2 hours, here’s the long cut. Audio-only link here.

In other news:

Busy week of Lawfare writing! Two pieces that I’m proud of here, because I think they’re important issues and I just like how they turned out. 🙂 I’m just about to hit my one year mark at Lawfare and I’m finally getting used to writing something on a weekly basis.

A review of the National Academies of Science “Understanding and Addressing Misinformation About Science” report. But also: on MAHA as a health populist movement, and why “misinformation” is an inadequate frame: “Misinformation Studies Meets The Raw Milk Renaissance”

And Berin Szoka of Tech Freedom and I wrote about the legal ramifications of Grok’s nudification debacle, the meaning of the word “censorship” from a legal, factual, and normative standpoint where computer-generated child exploitation and nonconsensual nudes are concerned, and why so few of the usual online defenders of children in the US Congress spoke out: “Grok, ‘Censorship’, & The Collapse of Accountability”

This year’s theme is “Trust Amid Insularity.”

The takeaway headline number this year: 70% of respondents globally say they are hesitant or unwilling to trust someone who differs from them in values, information sources, or worldview.

We’ve stopped trying to understand “the other side,” not just politically but epistemically, the report argues. Edelman connects this reluctance to two factors: polarization and grievance. They define polarization as “belief that our country is divided and those divisions are entrenched” and grievance as “resentment toward a system that’s rigged against you.”

What do they think might help us get out of this insularity rut? Influencers.

Since I wrote a book about our ability to construct immersive, insular bespoke realities, I appreciate further official corroboration that we’re in a crisis of trust and increasingly operating with completely different sets of facts. 😉 But I want to highlight a couple of things from the report—and also go a little bit deeper on the influencer thing.

“Trust brokering,” as Edelman calls it, is the idea that since institutions can no longer reach a mass public directly, they should work through intermediaries who already have trust within specific communities: employers, influencers, local leaders. Meet people where they are. Translate between groups. Rather than trying to change people, surface mutual interests and needs.

It’s not bad advice for navigating the current information landscape. But it’s advice for adapting to the jungle, not restoring the commons. And that distinction matters.

The turn inward is measurable—not theoretical

One of the most telling of Edelman’s slides is titled “From ‘We’ to ‘Me.’” People who say that recent societal events impacted their trust patterns—which is nearly everyone in their sample—report an increase in trust for “my neighbors, family, and friends” (+11 net) and “my coworkers” (+11 net), and a decline for national government leaders (-16 net) and major news organizations (-11 net). People even like their CEOs.

This is consistent with the arc Edelman has been charting for decades. “Trust Shifts from Authorities to Peers” was their 2005 theme. (I pulled up the 2005 report, which predates the mainstreaming of “social media” and “influencers”. In one poll about who the most credible corporate spokespersons might be, people in the US ranked doctors and academics at 60%, and entertainers and athletes at 4%.)

What’s new is the depth of the epistemic split: people are now expressing skepticism of anyone who “believes different facts and trusts different sources.” This is happening alongside a decline in exposure to cross-cutting information: the 2026 Trust Barometer notes that only 39% say they get information from sources with a different political leaning at least weekly—down 6 points year-over-year, with significant decreases in 20 of 28 countries. If democracy is predicated on a “marketplace of ideas,” where cross-cutting exposure nudges people to refine arguments or update beliefs, that marketplace is emptying out.

The Infrastructure of Insularity

But to understand why this is happening—and why it won’t easily reverse—you have to look at the media ecosystem itself, which is a bit outside the scope of Edelman’s work. So I want to pull in another report to connect some dots.

The Reuters Institute’s 2025 Digital News Report, released last June, is essentially a field guide to the information environment of our insular world. It’s also a survey, featuring detailed information on the media environments in dozens of countries spanning the globe, with all sorts of analytics on what platforms people are using, media consumption habits, trust in media, etc.

According to the Reuters report, 40% of people now say they sometimes or often avoid the news. Those who do engage are increasingly getting news from social media, especially via video, where mainstream media is losing the competition for attention and distribution to niche creators. A majority (58%) also report feeling like it’s difficult to tell real from fake news online – AI is making this exponentially harder, as even video becomes untrustworthy.

There is a very revealing statistic buried within the Reuters report: when people are asked how they would verify news they weren’t sure about, 38% say they would go to “a news source I trust.” In a follow-up question, United States respondents list their preferred sources: CNN, Fox News, and BBC are the top three. One of these frequently espouses a reality that is fundamentally incompatible with the others. It’s important to understand that ‘checking a trusted source’ doesn’t mean checking against a shared standard—it means checking with your team.

Under-35s, meanwhile, are significantly more likely to verify information via social media comments, AI chatbots, or a quick search (which increasingly returns AI synthesis). Verification involves consulting the feed—which, of course, is often personalized according to your preferences. What counts as “true” is increasingly shaped by context and identity, not vetting.

Edelman’s Trust Barometer describes a condition: insular societies where people increasingly trust only their own side. Reuters explains the fragmented information infrastructure that reinforces and rewards that insularity, where creator commentary outcompetes institutional journalism and verification has become tribal. Together they show a system optimized for division and consensus collapse.

Edelman’s solution treats the problem as permanent

Given this state of affairs, what’s an institution supposed to do? Edelman’s answer is “trust brokering”—work through intermediaries who already have trust within specific communities. If people are mostly verifying within their tribes, reach them through voices those tribes already trust.

The last third of the Trust Barometer report offers an approach for this—the suggestions are more sophisticated than my tongue-in-cheek reduction to “use influencers!” but not by much. We’ve gone from 4% saying a celebrity is a credible spokesperson for a company in the 2005 Trust Barometer to this in 2026: “Among people who trust a food or lifestyle influencer (48%), 62% say they would trust or consider trusting a company they currently distrust if that influencer vouched for it.” For financial influencers, it’s 57% among the 44% who trust one. Trust, in other words, is transferable through parasocial relationships—regardless of whether the institution has done anything to earn it.

It’s worth pointing out that the “among 44% who trust one” caveat is doing important work—because Reuters found that online influencers tied with politicians as the top perceived misinformation threat globally (47% each; celebrities scored lower!). That’s the paradox: influencers are simultaneously the most effective trust brokers and the most distrusted sources. The label covers a lot of different species: from the Besties that do makeup how-tos on TikTok to the Gurus who tell you about the secret health benefits of juice cleanses to the Reflexive Contrarian malcontents who turn political grievance into profit. Different genres, same engine: trust earned through charisma, resonance, and the steady performance of alignment, inside a media ecology built for niches.

Edelman’s playbook includes other recommendations for bridging across insular divides. Employers should “build teams requiring people with different values to collaborate”. CEOs: “consult people with different backgrounds before major decisions”. NGOs: “help distrusting groups understand each other.” Government: “avoid rhetoric that blames or vilifies particular groups.” (Must be for the non-US audience.)

It offers some rather staid advice for media: “provide equal coverage of different viewpoints,” and “write accurate headlines rather than fear-inducing ones.” The problem is that in the jungle, competition practically necessitates clickbait. The algorithm demands it—every news creator on YouTube has to play the headline and thumbnail game. Furthermore, influencers are often successful because they are not providing equal coverage—they’re eating traditional media’s lunch because they’re catering to a niche, particularly in the political realm.

There’s also an elephant in the room with Edelman’s “high trust”-ranking countries: several of the countries at the top—China, the UAE, Saudi Arabia—operate in fundamentally different information conditions than the countries at the bottom. Some of their high reported trust levels may reflect preference falsification when surveyed. But there’s also the simple fact of information control: a tightly managed environment doesn’t just suppress criticism of leaders, it suppresses the kind of constant, identity-targeted conflict that drives polarization and grievance. If the public rarely sees open coverage of leaders’ failures—or doesn’t see elites model “anything goes” divisiveness—trust is easier to maintain, at least on paper.

Ironically, several of the high-scoring countries have recently begun to rein in their influencer spheres, too: China now requires wellness influencers making health claims to have credentials. Saudi Arabia issued new influencer guidelines in September on compensation disclosure, morals, misinformation, etc.

Western democracies, by contrast, with our open speech and open systems, are supposed to have a commons, a marketplace of ideas that allows for cross-cutting dialogue and eventual consensus. It has been replaced by an information ecology shaped by algorithmic selection. Content and creators optimized for engagement outcompete everything else—not because they’re accurate or trustworthy, but because they’re fit.

The question isn’t whether institutions should learn to compete in the ecosystem we have. They have to. And Edelman’s advice in this latest Trust Barometer helps them do that, which is valuable; it may help bolster trust in institutions that were seen as aloof or absent. “Trust brokering” is a useful mechanism for thinking about how to help messages travel through intermediaries, through community validators, through people who already have attention and affinity.

But if we want to repair shared reality and change the direction of the trust trend at some point—if we want a commons again—the incentives of the communication ecosystem are going to have to change. As long as platforms and politics reward outrage over accuracy and identity over evidence, insularity won’t fade. It will only compound.

A few days ago—before we suddenly renditioned the President of Venezuela—a story about Minnesota daycare fraud had the rumor-to-propaganda pipeline operating again at full speed: online rage was converted into federal policy within days. This case offers a unique opportunity to discuss what constitutes sufficient proof of a claim these days, and why fact-checking doesn’t do much in these situations.

Let’s talk about the phenomenon that I’ve come to think of as “shopping for receipts”.

The Minnesota Daycare Video

On December 26th, 23-year-old content creator Nick Shirley posted a 43-minute video to X and YouTube purporting to expose extensive fraud at Somali-run child care centers in Minnesota. In the video, Shirley knocks on doors. Sometimes no one answers. When someone does, he demands to see children but is never shown any. At each stop, his counterpart David announces—citing state billing records—that the operation is fraudulent. This is then considered proven when the two men don’t see any children, and because the employees are unwilling to speak with them or hostile to being filmed.

The video has currently been viewed over 125 million times across X and YouTube.

Within days of its release, Vice President JD Vance declared that Shirley had “done far more useful journalism than any of the winners of the 2024 Pulitzer Prizes.” Homeland Security announced that it was on the ground increasing investigative efforts. The FBI announced a “surge“ in the state. The Deputy Secretary of HHS announced on X that all federal child care payments to Minnesota had been frozen (a department spokesperson later walked this back).

Firmer facts related to the situation didn’t travel as far: according to Minnesota’s Department of Children, Youth and Families, one location Shirley visited had been out of business for three years. When state regulators had conducted previous unannounced visits to the sites in the video, they found children present; they were also present at inspections made following the video. Shirley also visited at least one at a time when it wasn’t open; a nearby member of the online community of liberal political streamer Destiny apparently drove to the location of one center and was one of the first to post the operating hours not shown in the video, offering a counter-receipt. Other media have since reported from the centers visited by Shirley.

The situation reminded me of election rumors I’ve studied over the last decade: there’s always going to be a rare but nonzero amount of illegal voting (a few people across the country), as well as something that goes wrong with a machine somewhere. Those instances are leveraged to cast doubt on the integrity of the entire enterprise; the accusers repeat three word phrases—“stop the steal!”—to convey their message, while the defenders issue detailed, nuanced explanations and denials of wrongdoing to people who are predisposed to distrust them.

Obviously fraud is real, and Minnesota has had significant documented cases. Mainstream as well as local media, in fact, had been investigating and reporting on demonstrable fraud in the state for years—including related to childcare. But the purpose of this post is not to adjudicate fraud. It’s to describe the rumor-propaganda pipeline that increasingly converts suspicion into retribution before verification actually happens. Because this happened with USAID. With DOGE cutting random grants, programs, and “vampire” social security funds that it misinterpreted.

Once government action is tethered to viral content, belief careens into policy while evidence gets cut out of the process.

Evidence → Conclusion vs. Conclusion → Evidence

Investigative journalism, at its strongest, relies on a method:

1. Start with a tip, question, hunch, or anomaly

2. Gather records, interviews, data

3. Test hypotheses – including the boring ones

4. Publish what the evidence supports, plus what couldn’t be verified

The story earns its claims, and at a major publication, has to survive legal review.

Hyperpartisan influencer culture often runs the pipeline backwards:

1. Start with a vibe or identity-based claim (betrayal, aggrievement, “we are being suppressed”, “the vaccines have athletes dropping dead”, “the election was rigged”)

2. Decide which outgroup or villain is responsible

3. Hunt for plausible support (screenshots, clips, “expert says,” FOIA snippets)

4. Incorporate the audience into the assembly line (”do your own research”) and amplification process

Repetition and confidence in delivery serve as validation.

Journalism is evidence → conclusion. This is, at best, conclusion → evidence. The verdict comes first; the hunt for proof comes later.

I call this shopping for receipts: influencers start with a belief or suspicion, then go looking for artifacts that will read as evidence to people who already agree. A receipt, in ordinary life, documents a transaction—you paid, you got something, the work was done. But in this context, the “receipt” is mostly just a prop: a screenshot, a video clip, large numbers on a spreadsheet. It looks like proof but there’s no real process behind it.

That’s why this isn’t only a reversed order of operations. It’s a different information logic entirely: it’s not a conclusion so much as a belief, supported by justifications. The belief relates primarily to identity, and group membership—not facts. The justification just needs to make the belief feel reasonable.

The Minnesota daycare video illustrates this difference. The beliefs—that Somali immigrants are criminals, that the money they’re getting from fraud is funding terrorism, that Tim Walz doesn’t care—were already circulating. Trump called Somali immigrants “garbage“ at a cabinet meeting in December. City Journal published an unsubstantiated claim linking fraud schemes to Al Shabaab. The frame had been established—both with respect to the Somali community in particular, and to immigrants writ large.

Shirley’s video wasn’t a methodical collection and review of evidence to reach a conclusion; it was content produced in response to a belief. The “evidence”—employees reluctant to speak, buildings without visible children—clearly felt like proof to many thousands of folks commenting on X. To those outside the political faction, however, the behavior of the daycare employees seemed like a reasonable response to two strange men with a video camera showing up demanding to see children.

When a belief is socially rewarding—when it flatters your side, confirms your suspicion, humiliates your enemies, or offers a clean explanation for a messy world—the evidence doesn’t have to be strong. It just has to be available, shareable, and understandable at a glance.

When mainstream media, for example, debunked the eating-the-pets rumors in Springfield, right-wing influencers descended to find “evidence” that the media was lying. The closest they got was grainy years-old footage of unidentified animals (possibly chickens) on a grill at a house inhabited by Africans (not Haitians)—a motte-and-bailey retreat from a sweeping allegation to a niche claim that was still poorly defended. But to the community of believers, it was sufficient.

Virality as Governance

There’s another dimension here worth mentioning, and reinforcing from my last post: elite authority figures once served as a firebreak, calling for cooler heads or complete investigations. Today, rapidly jumping on a wildly-viral rumor about an “outsider” harming “the people” makes an exclusionary populist government look incredibly responsive. The Administration’s actions in Minnesota—cutting off funding, stating they’ll deport fraudsters, etc.—makes them appear to be listening directly to the public while mainstream journalists and state investigators look slow, procedural, and out of touch. The New York Times reports that state regulators found no evidence of fraud; the federal government acts anyway. Who looks more in touch with ordinary Americans’ concerns?

This is a mutually beneficial system. The political influencers get views, subscribers, and access. The administration gets to demonstrate that it acts on what “real Americans” care about, not what the “legacy media” prioritizes. (In this specific case it’s also worth noting: the Minnesota story pushed the Epstein documents out of the news cycle) The base, which is an active participant in tossing out names for investigation and amplifying content, feels that they are being heard. Everyone wins—except the communities unfairly caught in the crossfire, and the epistemic norms that used to govern how allegations become action.

As more investigations get underway, there might well turn out to be fraud at the specific centers in Shirley’s viral video. If so, I’ll be glad to see charges filed and perpetrators held accountable; stopping online scammers is part of my job, and I have no patience for people who manipulate and steal. But investigations take time. Methodical, deliberative processes often feel very slow. They don’t have enough of the visceral owning-your-enemies energy that some extremely online factions have come to expect, and crave.

Shirley’s video, by contrast, delivered that.

Why Debunking Doesn’t Work

If we were in a situation where evidence → conclusion, then correcting the misperception, or adding nuance might change people’s minds. But things are flipped in the process that this video is a part of. The conclusion is upstream of the evidence—the belief preceded the video.

This is why fact-checking so often feels like bringing a knife to a gunfight. It assumes that people are working from evidence to conclusion, and that if one points out problems with the evidence, people will see the flaws in the conclusion and change their minds. But content like this is operating in the “belief → justification” realm. Swat down one justification and the belief can draw on another. The video isn’t why the audience believes Somali daycares are scams; it’s a performance of that belief.

This is why I think “propaganda” is often a very appropriate term for political influencer content. Not as a pejorative, but because propaganda is information that is shaped to produce a particular attitude, or activate a particular group. It starts with a belief it wants people to hold. And it’s not always false; it often incorporates selective facts to tell a convincing story.

Tells

If you’re trying to recognize when you’re being pulled into belief-first justification content, here are some tells:

Conclusion-first framing: The verdict comes before the evidence. Shirley announces fraud at each stop because he doesn’t see children—no alternative explanation for their absence is considered.

Cargo-cult empiricism: Spreadsheets, timestamps, or “forensics” language used as aesthetics. The Shirley video cites state billing records and “big number” dollar figures, giving it the look of quantitative research without the actual investigative work of verifying or explaining what the numbers mean.

Negative-space proof: ‘Absence as evidence’ as the core argument. No one answers the door; silence becomes guilt—even if a facility was closed at the time, or the person declining to speak appeared scared.

Motte-and-bailey: Make a sweeping claim, retreat to a smaller technically-defensible one when challenged. In the video Shirley stated he had found $110 million in fraud; yet when pressed on this number since, Shirley and his supporters have retreated to “you can make up your own mind”.

Unfalsifiability: No kids are at the daycare—fraud. Kids are there on other days—they’re being bussed in, still fraud. If regulators find fraud—belief confirmed. If they find no fraud, it’s a cover-up—belief still confirmed. When nothing can change the investigator’s mind and all counterevidence is dismissed, it’s not really an investigation. This one takes some time; it remains to be seen how it plays out in this specific situation.

These tells don’t prove a claim is false—they just signal that a conclusion likely preceded the evidence-gathering. Spotting them suggests that it’s worth pausing and looking for more rigorous verification.

What’s To Be Done

The incentive structure of social media means that “shopping for receipts” isn’t going anywhere. After Shirley’s video went viral, another creator posted a video of himself knocking on the door of a child care center in Columbus, Ohio “associated with the Somali community.” Nobody answered. “We’re just getting started,” he said. Another aspiring influencer door-knocked Somali family daycares in Washington State. This isn’t a response to real signals of fraud—it’s a response to visible rewards: money, attention, and access. Shirley gained ~170,000 YouTube subscribers following his viral video. Andrew Tate asked how he could donate to support. Shirley and fellow creators had previously been warmly received at a White House roundtable by Trump, who called them “very brave patriots” and said he’d consider awarding them “some very important medals and honors.” The current Administration is highly engaged with sympathetic creators. When the message is that viral accusation is civic virtue, you get more viral accusation.

The pipeline from rumor to policy is now shorter than the time it takes to verify a claim. That’s a remarkable—and dangerous—shift. People have always believed rumors, but now government elites ride them when they provide justification for what they wanted to do anyway. “Shopping for receipts” isn’t just a rhetorical move—it’s a substitute for real inquiry.

We still get to choose how we respond. The reality is that a viral video does sometimes force accountability that slower, more cautious processes fail to achieve—particularly if the public clamors for action. The challenge is trying to incentivize keeping the content rooted in reality. The goal is to separate the signal from the noise: harness the accountability that viral exposure can lead to, while rejecting the belief-first propaganda that often comes with it.

I plan to write a lot more about the dynamics of modern propaganda here this year…so, if you found this post informative, please share it with people. 🙂

Happy 2026!

The aftermath of the Brown/MIT shootings was a time of high anxiety. The appetite for answers was enormous, and the police hadn’t yet identified a suspect. And in that vacuum, X did what it often does: it produced one. The wrong one.

An undergraduate queer Palestinian activist with they/them pronouns—a villain straight out of MAGA central casting—was falsely floated as the culprit by a large anon account, and then others piled on. The claim was supported by little more than insinuations and the cargo-cult logic of internet witch hunts, where a target’s silence becomes “proof” and the absence of evidence becomes a “cover-up.” There was even “forensics” cosplay: claims of “97.8% gait match,” using a methodology that had collapsed spectacularly for The Blaze weeks before when it wrongly accused a former Capitol Police officer of being the January 6th pipe bomber.

The student, and the falsely-accused officer the Blaze smeared, hopefully have excellent legal teams assessing their options. Because the media—they are the media now—who blasted their names around the internet should have known better. Tim Pool, now part of the Pentagon press corps, amplified suspicion of the student. Laura Loomer, also in the Pentagon press corps, declared that the scrubbing of their Brown university web presence was “a massive coverup to protect the identity of the shooter” while also insinuating that they had a YouTube channel dedicated to making explosives. Harmeet Dhillon, now the Assistant Attorney General for Civil Rights at the United States Department of Justice, joined Representative Anna Paulina Luna in calling the removal of the student’s information “suss“ (normally spelled ”sus,” as in suspicious).

These accusations came hot on the heels of another baseless rumor bandied about by large influencers—Elon Musk included—that the killer had intentionally targeted campus conservatives, sparking calls for violent reprisal against “the left” and the supposed anti-conservative assassin.

Days later, police identified the actual perpetrator: a Portuguese man with no connection to either the falsely-accused student or to the campus conservative victim. Some of the student’s accusers quietly deleted their posts. Others, like Dhillon, acted as if nothing happened. The innocent person at the center of the firestorm was left to pick up the pieces.

But it’s important to understand what happened here. This wasn’t “misinformation”. It’s an issue of online rumors now powering the right’s political reality—up through the highest levels of government.

What a rumor mill is, and why it matters

The rumor mill is what happens when uncertainty meets social participation.

Here’s how it works: Something happens—a natural disaster, an election dispute, a mass shooting. Information is partial. People feel anxious. They want an explanation, a villain, a reason. Rumors are an age-old way of filling that void. They’ve always served a social coordination function: theorizing and spreading rumors binds people together, signals who’s in and who’s out, directs anger, justifies suspicion. Rumors can also be very entertaining. And they sometimes turn out to be true.

But X’s architecture makes rumors ambient. They’re not an occasional byproduct of uncertainty, they’re a pervasive form of content. The platform is constant bait, designed for heat over light, and participation over restraint. There is always some rumor being pushed as a nascent “trend” to some niche group of users that the algorithm thinks will run with it. And when it comes to politics in particular, people don’t just consume rumors; they perform them. Contributing or retweeting commentary is a way to signal membership in an online faction— while simultaneously getting a dopamine hit.

Big political influencer accounts serve as circuit breakers or accelerants, deciding whether something explodes or fizzles: “Big if true!” as a quote retweet can shift tens of thousands of peoples’ attention for an afternoon. Rumors can shape how people make sense of reality—especially when institutional pushbacks or fact-checks on a claim are slow, boring, or already distrusted.

And today, reality has splintered along factional lines.

The handoff

The rumor mill alone has always had the potential to be dangerous—pitchfork-wielding mobs once carried actual pitchforks. But something very bad is happening very frequently now: online rumors are being co-opted by political elites, who see value in them. They’re messages that already resonate, viral moments that already have momentum. And political leaders, statesmen, who once served as a firebreak against baser instincts increasingly see an opportunity to look receptive to the base—so they add kindling instead.

If rumor mills are how communities make sense of the world amongst themselves, propaganda machines are how elites shape narratives for mass consumption. The line between rumor mill and propaganda machine is indistinguishably blurred now; everything moves along the same pipes. Modern propaganda is participatory: the audience supplies the amplification energy and “evidence”—screenshots they’ve found, database entries they think they’ve discovered. Algorithms key off of content and engagement, and act as distributors.

But political elites supply the credibility that helps a rumor become “real.” Influencers often supply the framing: what does it mean, who is to blame, how should “our community” feel about it.

For another example of this merger between rumor mill and propaganda machine, recall “Eating the Pets.” The false rumor that Haitian immigrants were eating cats in Springfield, Ohio sparked considerable debate again on X two weeks ago…because JD Vance, Vice President of the United States, chose to reaffirm it in a recent podcast interview with the New York Post.

The massively viral 2024 lie that Haitians were eating pets in Springfield, Ohio wasn’t cooked up by a political consultant. It started like so many rumors do: with a woman sincerely relaying something she heard through a chain of acquaintances—that a neighbor’s friend’s daughter’s cat had been eaten—into a Facebook group. Her post was screenshotted, transferred to X, and then went viral as people gaped at the story and made AI-generated images illustrating it. Trump supporters with negative views of immigration repeated it because it felt emotionally satisfying and politically aligned. It was also just that kind of extremely-online bizarre.

And then “the handoff” happened—the inflection point when a rumor moves from anon and small-account chatter to pickup by big influencers and people with institutional or elite standing. Vance, then-Senator from Ohio and vice presidential candidate, ran with it. He didn’t lower the temperature—he was the accelerant. He gave the claim legitimacy, and things went from there.

Elite media and political actors on the right are remarkably effective at taking whatever rumor the MAGA online faction has begun to run with and turning the claim into a deeper story, a symbol, or a permission structure. Truth and facts are not the main concern. (The left is not immune to political rumors obviously, but the right’s crowd-to-elite handoff is unusually disciplined and repeatable, to the consternation of the left, for reasons I’ll explain in another post.)

Why “keep score” isn’t enough

In the aftermath of the Brown witch hunt, Graeme Wood offered a prescription in The Atlantic for how to handle accounts that spread viral rumors: “Don’t log off. Keep score.”

Track who speculated. Track who amplified. Track who hedged with “concerning.” Track who deleted later. Track who refused to correct. Adjust your trust accordingly. Reward restraint. Ostracize repeat offenders.

It’s appealing because it gestures at a mechanism that predates the internet: social sanctions. If someone repeatedly spreads false accusations that harm people, eventually they get ostracized from the community. They're perceived as untrustworthy. They lose standing. They feel ashamed.

But that only works if the community shares a norm that spreading false rumors is bad. And if they agree that the rumor was, indeed, false.

We’re not there anymore.

On social media platforms, influential liars are rewarded while the new crowdsourced factcheckers—Community Notes contributors—are unremunerated. The people willing to produce the most shameless bullshit gain followers, subscriptions, political access, and status. “Keep score” assumes the players are all playing the same game.

Which brings us to the deeper problem: “the end justifies the means” has become the guiding principle of the political project that has most effectively fused the rumor mill and the propaganda machine.

The Internal Critics

Some influential voices on the right have noticed. They describe what’s happening as a problem of conspiracism, which is true as far as it goes—but conspiracies are the symptom. The underlying problem is that insinuation and shamelessness pays.

This week at Turning Point USA’s AmericaFest, Ben Shapiro warned that “the conservative movement is in serious danger”—not from its traditional enemies on the left, but “from charlatans who claim to speak in the name of principle but actually traffic in conspiracism and dishonesty.” He called out Candace Owens for spreading conspiracy theories about Charlie Kirk’s assassination. He condemned the “just asking questions” tactic as manipulation: “When grown men and women spend their days ‘just asking questions’ without seeking answers, they’re lying to you.”

Chris Rufo recently wrote that “the Right’s media apparatus is how the Right teaches its followers how to think, and it’s currently getting consumed by conspiracy, psychodrama, and tabloid conflicts.” He described the public mind on X as being “shaped by the X algorithm”—and noted the impact of perverse incentives that come from the current monetization framework: “While the Right has long had a conspiratorial streak, such narratives have usually been restricted to marginal newsletters or pirate radio broadcasts. By contrast, on X, they generate enough controversy and spectacle to become the day’s top headline.”

These are not liberals. These are prominent conservative voices now starting to acknowledge that something has gone badly wrong.

Rufo’s commentary about the algorithm, platform power, and social media’s bad incentives is accurate. This is a supply-side complaint that the left makes as well—but he similarly neglects to fully reckon with demand, and, specific to the right, with the handoff. Shapiro tackles the handoff head-on, but internal correction only works if the people at the top stop doing the thing. Meanwhile, the Vice President is still lying about pet-eaters in Springfield (and he knows it). Elon Musk is amplifying stolen-election claims to hundreds of millions of followers as of yesterday. The incentives haven’t changed. The fish rots from the head.

So what is to be done?

Two things, neither satisfying.

First: internal and individual norm shifting

Shapiro is correct that the elites on the right need to push back against conspiracy theories from within. Nothing from the left, from academia, media, or fact-checkers, or anyone on the outside will do much here. Those outside the movement have been systematically discredited as biased hacks for years. But changing a political culture that treats “owning the enemy” as more important than honest reckoning with facts is hard.

Not seeing every unhinged rumor as something to accelerate or validate is something anyone can choose to do; this is generally true across the board in factionalized online politics. Some on the left made leaps to judgement based on very minimal evidence surrounding Charlie Kirk’s assassination. And in the Brown case, there was another young man whose name and LinkedIn profile briefly circulated on Threads prior to the Palestinian student’s—someone who had the same name as yet another man who was briefly a person of interest. That person took down his LinkedIn page, leading a few midsize influencers to become convinced he was guilty. This all happened at far smaller scale, and seemingly without political elite buy-in. But the point remains: people have to decide they want reality more than dopamine and dominance.

Second: make accountability great again

It’s my controversial opinion that defamation lawsuits should be far, far easier for private individuals who get screwed like this. I just don’t think the law has caught up to the infrastructure. Yes, defamation lawsuits are imperfect, slow, and expensive, and can be frivolously abused. But influencers who falsely accuse a private person of mass murder for clout and profit should face consequences. The legal system is one of the few mechanisms available to impose costs and deter this kind of behavior. The harm is real. Treating human beings as content is depraved. Defamation has always been outside the bounds of free speech, and our marketplace of ideas would benefit from less of it.

I wrote Invisible Rulers to explain how this ecosystem works—how crowds, algorithms, and elite amplification combine to turn lies into political reality. None of this is going to stop on its own. For the time being, the rumor mill will keep spinning. The propaganda machine will keep grabbing whatever it can use. And unless the culture changes—unless our old offline norms reassert themselves, and/or accountability starts to bite—the handoff will keep happening.

More things I’ve done since my last newsletter:

Speaking of shamelessness, smears, and profiteering: I posted a new entry on my “bad media object lessons” blog featuring recurring exemplar Matt Taibbi. This time Matt is cutting emails in half and lying about them while his assistant Andrew Lowenthal makes claims about a legal situation that are directly contradicted by his friend’s own court filings. Reading is hard, linear time can be vexing. Enjoy.

And finally, I hosted a Scaling Laws Lawfare podcast on AI chatbots and free speech with Jacob Mchangama and Jake Shapiro, who are fantastic. Want to hear about which bots are the most or least willing to produce politically sensitive content? How should we think about the needs of the requester vs the right of the company to set limits? It was a very thought-provoking conversation about some of their new research!

Merry Christmas and Happy Holidays to all!