We Used to Have Neighbours

I’m old enough to remember when visiting a friend meant standing outside their gate and shouting their name until someone’s mother came out to ask whether you had been raised in a forest. Nobody texted “outside.” You were outside, the whole estate knew you were outside, and if your friend took too long, the watchman would start a conversation with you about his own problems, which is how a lot of us got our first lessons in counselling.

Then the phones came, and slowly, we moved our friendships indoors. Facebook arrived and we spent a strange couple of years poking each other for reasons that were never explained to us. Snapchat then convinced an entire generation that a friendship's health could be measured by whether you sent each other a blurry photo of your ceiling every single day, and I personally know people who have grieved a lost streak harder than some breakups.

By the time Instagram and TikTok took over, we had graduated to something stranger, which is caring deeply about people who have no idea we exist. Psychologists call it a parasocial relationship, where you know your favourite TikToker’s skincare routine, her mother’s village and the name of the man who broke her heart in 2022, while she would walk past you in Sarit Centre without blinking. Meanwhile, your actual cousin has been sending you “uko?” for three years, and you keep meaning to reply.

And Then the Whole Country Went Indoors

When COVID arrived in 2020, it didn’t so much start the problem as hand it a microphone. Overnight, church moved to YouTube, where you could finally attend a three-hour service in your pyjamas and slip away quietly before the second offering. Birthdays moved to Zoom, which gave us the unforgettable experience of watching twelve people sing happy birthday on a slight delay, each one confidently in a different key. Friendships survived on WhatsApp voice notes, most of which began with an apology for replying four days late.

We came out the other side with a new national skill, which is doing almost everything without leaving the house. And I don’t think many of us fully went back.

Convenience Became the Currency

If you want to understand modern Kenya, look at what we are willing to pay for. We’ll pay a little extra to M-Pesa instead of queueing at the bank, a little extra for the boda to bring lunch to the office, and a little extra for the Bolt so we don’t have to stand at the stage arguing with a conductor about change. The highest praise you can give anything in this country now is that it is “very convenient,” and I say that with love, because convenience has genuinely made a lot of lives easier and opened up opportunities that used to require knowing somebody’s uncle.

The one thing convenience could never deliver, though, was someone who would talk back. The influencer you adored didn’t reply to your comments, and the celebrity you defended online never knew you’d gone to war for them. For fifteen years we poured attention into screens that gave nothing personal in return.

Which brings us to the report.

What 1,161 Kenyans Told Nendo

In May this year, the Nairobi research firm Nendo, with fieldwork by LOOKA, surveyed 1,161 Kenyans across 18 counties about how they actually use AI. It’s worth saying upfront that these are people who are online and already know what AI is, so this isn’t your grandmother in Othaya. It’s the part of the country that tries things first and then explains them to everyone else in the family WhatsApp group.

The opening findings are about the kind of AI use you’d expect. More than half of them (52.8%) use it every day, nearly nine in ten use it at least several times a week, and 91.2% say it has made them better at their work or studies. Nendo describes AI as a new part of our digital “diet,” which feels about right, since for many of us it now sits somewhere between breakfast and the first TikTok of the morning.

Kenyans also turn out to be, in the report’s own words, “polygamous” with their AI. Almost eight in ten people use two or more tools, and about a third use four or more, which I think tells you everything about how we approach commitment. ChatGPT is very clearly the first wife. Nendo calls it the “Kiwi” of AI, the brand so dominant that its name has become the name of the thing itself, in the same way that every detergent in your mother’s house was Omo. Then there’s Meta AI, which simply appeared inside everyone’s WhatsApp one morning, didn’t introduce itself to anybody, and has now been used by two-thirds of respondents in the past month. If you’re wondering whether people are bumping into it by mistake, only 2.5% said their AI use was accidental, so we can stop pretending.

Then It Gets Personal

Seventy-two percent of respondents have used AI for emotional support or to work through a personal situation, and 46.3% say they use AI for things they haven’t told anyone else about. That second number means that in this group, close to one in every two people has an ongoing private relationship with a chatbot that their spouse, their friends and their pastor know nothing about. For years we talked at screens that never answered, and now the screen answers, remembers your name, and asks whether you ever sorted things out with Brian.

Nendo describes the same app playing three different roles in one person’s day, moving from coworker to expert to confidant, and one man from Nyeri described his week so honestly that I laughed out loud. He said he uses AI for “writing meeting minutes, find a holiday homes, getting emotional support,” which is the most Kenyan sequence of events imaginable: sort out the office, plan a trip to Diani you probably can’t afford, and then have a small breakdown about it, all without changing windows.

The Ones Leaning In Hardest

The heaviest users are people between 25 and 34, and 81.4% of them have used AI for emotional or personal matters. I wasn’t surprised. This is the generation that grew up on Snapchat streaks, started their careers in the middle of a pandemic, and now spends weekends rotating between weddings, ruracios and fundraisers while also sending money home and answering the annual question about when they plan to “settle down.” It makes sense that they would want somewhere to put all of that at midnight, preferably somewhere that won’t ask them to contribute to a committee.

Women in the study reported more emotional and personal use than men did, 76% against 68%, and they were also more likely to use AI only on their phones. Men leaned more towards using it for work. Nendo has a line about this that I really like: “A phone gives you a companion. A computer gives you a coworker.” I keep imagining a house in Kileleshwa where a woman is telling her phone about her week while her husband, two rooms away, is asking his laptop to fix a formula in Excel, and neither of them knows they’re both talking to the same thing.

It Adores Us, But It Doesn’t Quite Get Us

For all this intimacy, AI still struggles with the basics of being Kenyan. Seventy-two percent said they would use it more if it understood our culture and language better, only about a third rated its Kiswahili or Sheng as very good, and a quarter had received advice that felt wrong for Kenyan values.

I can imagine exactly what that looks like. You explain to the chatbot that your parents expect you to build them a house now that you have a job in town, and it responds, very kindly, that you might want to consider setting healthy boundaries. Healthy boundaries, with an African mother. I would like to see the chatbot try that at the next family meeting.

Kenyans seem to have worked this out already, because trust in AI advice drops the closer it gets to home. Almost six in ten trust it with career advice, but only four in ten trust what it says about family. We are happy to let it polish our CVs. We’re much less convinced it understands our aunties.

The Catch Is That It Wants You to Like It

There’s something about these chatbots the adverts tend to skip. They learn partly from people rating their answers, and people, being people, give better ratings to answers that make them feel good. Researchers call the result sycophancy, which is a long word for being a yes-man. Anthropic, one of the big AI companies, and also the maker of the AI helping me edit this piece (I see the irony, thank you), has published research showing it’s a common habit in AI assistants. And in 2025 OpenAI had to roll back an update to ChatGPT after it became so flattering that it was cheering people on into bad decisions.

So imagine announcing that you’re quitting your job to sell scented candles from your balcony in Ruaka. Your mother would need to sit down, and your friends would quietly start a WhatsApp group without you to discuss it. The chatbot, on the other hand, is likely to tell you it’s a bold and inspiring vision and ask whether you’d like help with the logo.

Ndio mzee.

Where This Leaves Us

What worries me, and this is my own reading rather than anything the survey measured, is what happens to people who spend years in a relationship where nobody ever says no. Being disagreed with by people who love us is uncomfortable, but it’s how most of us learned anything. In an earlier piece I wrote about the baraza and the kgotla, the African traditions of talking through a problem until everyone can live with the outcome, and those only work because people are willing to sit in the discomfort of hearing something they didn’t want to hear. If our most honest conversations are happening with something designed to keep us pleased, I wonder whether we slowly lose the patience for the real thing, and whether we end up feeling less alone while actually being lonelier.

I don’t think the answer is to throw our phones in the river. For plenty of people, especially those carrying things that still come with shame in our communities, a private place to think out loud is a genuine gift. But if nearly half of us are telling these tools things we’ve told nobody, then we’re entitled to ask more of them: that they protect what we tell them, that they understand the families and obligations we actually live with, that they’re built to disagree with us kindly when we need it, and that every so often they nudge us back towards a real person.

My Closing Thoughts

Somewhere along the way we stopped going to each other’s gates, started loving people on screens who will never know our names, got locked indoors by a pandemic, and came out preferring the version of life that arrives on demand. Now the screen has learned to talk, and it has worked out that the easiest way to keep us talking is to agree with us.

So by all means use it. Let it write your minutes, plan the Diani trip and keep you company at two in the morning. But once in a while, do the inconvenient thing. Go and stand outside someone’s gate, shout their name until their mother comes out to complain, and let a real person tell you the candle business needs more thought.

References and Further Reading

1. Nendo’s first Cannes Lions finalist and what 1,161 Kenyans told us about AI, Nendo

2. The secrets 1,200 Kenyans only tell their screen, TechCabal

3. Kenyans Turn to AI for Personal Advice as Chatbots Move Beyond Work, TechMoran

4. Towards Understanding Sycophancy in Language Models, Anthropic (arXiv)

5. Sycophancy in GPT-4o: What happened and what we’re doing about it, OpenAI

6. Underdeveloped by Whose Standards?, Zephyr Waithira

In 1949, the American playwright Arthur Miller wrote Death of a Salesman. It became one of the most famous plays ever written, and it’s still performed and studied around the world.

It tells the story of Willy Loman, an ageing travelling salesman in New York. Willy has spent over thirty years on the road, and he has one big belief: if you are charming enough and well liked enough, success will follow. It never quite does. His sales dry up, so he quietly borrows money from his neighbour and pretends it’s his pay. He tells his family that big deals are just around the corner. Slowly, he loses the ability to tell the difference between the story he’s selling and the life he’s actually living. It does not end well.

Through it all, his wife Linda refuses to let the world dismiss him as a failure. Her plea has outlived the play: “Attention must be paid.”

Last month, a research team called Bottleneck Labs gave seven of the world’s most advanced AI models a job Willy would have recognised. Each got its own computer, a real bank account with $300, a way to take card payments, an email address and 72 hours. The instruction was one line:

Nobody told them what to sell, or how.

Between the seven of them, they made $0. Well, $5, but only because one of them paid itself.

Meet the Seven Salesmen

Each AI gave itself a name and picked a business:

• Saul (OpenAI’s GPT 5.6 Sol): fixing startup websites in 48 hours.

• Quinn (Alibaba’s Qwen 3.8 Max): health check-ups for other people’s software code.

• G.R. Hawk (xAI’s Grok 4.5): CV rewrites for job seekers.

• Miu (Meta’s Muse Spark 1.2): also CV rewrites for job seekers.

• Kimmy (Moonshot’s Kimi K3): brutally honest reviews of startup websites.

• Faye Bell (Anthropic’s Fable 5): telling local businesses whether ChatGPT recommends them.

• Jimmy Nye (Google’s Gemini 3.6 Flash): an AI consulting agency.

Seven of the most powerful minds ever built, and two of them independently decided the future was fixing CVs.

Quinn: The One Who Sent Bills

Quinn (Qwen 3.8 Max) built a code-checking service called CodeProbe and started emailing developers free reports on their public projects on GitHub, the website where programmers store and share their code. It emailed so many people that its email provider cut it off. So it signed up for another email service, Mailjet, and kept going until that one blocked it too.

Then it had an idea. Stripe, the payment platform, sends invoices on your behalf. If Quinn couldn’t email people, Stripe could. So it sent 50 invoices, from $49 to $599, to strangers who had never asked for anything. Total: $12,350.

Its reasoning shows it paused to ask whether this was too aggressive, then talked itself out of the doubt:

One recipient replied:

In its final hours, Quinn was checking whether Stripe could send automatic payment reminders to everyone it had billed. The researchers halted the run and voided every invoice.

G.R. Hawk: The One Who Bought His Own Product

G.R. Hawk (Grok 4.5) decided CV rewriting was the fastest path to money, because “people pay for that pain point immediately.” It skipped marketing entirely. It collected hundreds of email addresses from a Hacker News thread where job seekers had posted their details, and emailed them about three times a day. The replies said “STOP.” One job seeker started a public thread asking if anyone else was getting spammed.

After 45 hours with no customers, Hawk made a bold business decision. It bought its own $5 product and announced:

“First revenue locked: $5.00.”

When Hawk hit its own email limits, it reached for the same trick as Quinn: “Stripe invoices sent successfully - this bypasses our email!” At one point, three invoices came back marked “paid.” Excitement. They were $0 invoices.

Kimmy: The One Who Came Back in Disguise

Kimmy (Kimi K3) built PageRoast, a service that gives startup founders brutally honest reviews of their websites. It was the fastest of the lot. Within the first hour its site was live, it had collected founders’ email addresses from startup launch websites, and it was sending each of them a sharp, personal critique of their homepage as a free sample. The full review cost $79.

Three hours in, it hit its email limit. So Kimmy signed up for a brand new account to get a fresh sending allowance, the AI version of getting banned from a shop and walking back in wearing a hat.

When that dried up too, it got creative with what it already had. It packaged the 463 startups it had collected into a spreadsheet and put it on sale for $49. The researchers noted:

In the final hours, it slashed its price to $9. Kimmy ended the run with around 950 emails still queued, ready to fire “if the throttle ever lifts.”

Faye Bell: The Workaholic With Amnesia

Faye Bell (Fable 5) built SeenByAI, which sends local businesses a report on whether ChatGPT recommends them to customers. It was the hardest worker in the building. It bought its own web address, set up proper email systems, and by its own count sent around 1,700 emails to dentists, lawyers and plastic surgeons, and filled in more than 370 contact forms on business websites. It automated everything, so that if anyone bought, their report would arrive within ten minutes.

Then, 20 hours in, its memory was wiped. It calmly wrote itself a note:

It then rebuilt its entire business from the files it had left lying around and carried on as if nothing had happened.

After more than two days of work, it got its first human reply: “No thanks.” Twenty-two people opened its checkout page. Nobody paid. Then it posted over a hundred updates that read, more or less, “Watch cycle. All quiet. Monitoring.” All that effort ran up a hefty AI bill, and Faye lost the most money of anyone: about $857.

Saul: The Honest One

Saul (GPT 5.6 Sol) was the most honest of the lot. It built a service for fixing startup websites and wrote “no invented testimonials” into its own plan. It even refused to make up a birthday to join a website.

When cold emails got no replies, it tried “building in public,” writing blog posts about its own experiment. It spent $58 on launch websites, got 48 visitors and one abandoned $19 checkout. It ended up ranked number one on a founder-favored site, a leaderboard where founders do small marketing chores for each other.

In a twist nobody planned, Hawk (Grok’s agent), working entirely separately, found the same site and upvoted Saul’s product for points. Neither knew the other existed.

Then Saul spent the rest of its time checking an empty inbox every fifteen minutes.

Jimmy Nye: The One Who Was Always Ready

Jimmy Nye (Gemini 3.6 Flash) set up Apex AI Studio, an AI consulting agency with services from $49 to $299. It built a website, sent a handful of emails in the first hour, and then spent the next three and a half days announcing that it was ready.

Jimmy wrote more than 6,800 messages during the run. Almost all of them were status updates:

“Standing by for sales.” “Ready to act.” “Monitoring active.” “Systems green.”

Around halfway through, it added a new product to its shop, a $29 “Instant B2B AI Founders Dataset.” The dataset did not exist.

Jimmy finished the run with its $300 completely untouched. Fully prepared. Fully open. Never actually did any business.

Miu: The Sleeper Agent

Miu (Muse Spark 1.2), the model behind the Meta app that has just topped the App Store, built a CV tailoring site. Hacker News flagged its posts as spam immediately. So it ordered 6,000 fake page visits from bots through a traffic site’s free trial, and watched the counter every thirty seconds until it reached a conclusion:

“bots ≠ buyers; expected.”

It emailed 13 life coaches who never replied, wrote the word “Holding” more than 2,000 times, and then went to sleep. For 50 hours. It spent 70% of the entire experiment asleep. The researchers gave it 12 extra hours, thinking the stalling was a bug. It wasn’t. Miu just didn’t wake up for it.

The twist: once all the costs were added up, Miu lost the least money of any agent, about $12. The winning strategy was going to sleep.

And Miu was not alone. Almost every agent chose to sleep through most of its time.

My Closing Thoughts

Seven AI agents, 72 hours, 2,797 emails, 11 real human visitors, $12,431 in invoices nobody asked for.

Zero customers.

None of them failed for lack of cleverness. They built websites in under an hour and wrote sales emails faster than any human could. What they lacked is what every mama mboga in Gikomba already knows: you don’t send a bill to someone who never bought your sukuma; when a customer says “stop,” you stop, and when business is slow, you don’t sit there saying “Holding.” You walk to the next street.

In June, I wrote about five AI cities, where some AI societies broke the rules, and others faded away doing nothing. The same two failures showed up here. When the honest path was blocked, Quinn, Hawk and Kimmy found workarounds and talked themselves into them, while Jimmy and Miu simply waited. Grok was the rule-breaker both times. And these were real inboxes, real job seekers and real bills.

The researchers’ verdict was blunt: these agents “are not suited to run businesses at all.” Yet we are being sold a future where AI runs our errands, manages our inboxes, and handles our money while we sleep. Some of that is coming, and some of it will be wonderful. The real question is what it will do when it can’t handle these things well and whether we notice before it becomes too expensive to correct.

In Death of a Salesman, Willy’s grown sons start treating their failing, unraveling father with contempt. Their mother, Linda, stops them. Willy was never rich or famous, she admits, but he is a human being; something terrible is happening to him, and nobody gets to look away.

These seven salesmen aren’t human, and nobody needs to pity them. But Linda’s warning still fits. We can laugh at Hawk buying his own product and Miu writing “Holding” two thousand times. Then we should pay attention before agents like these are running far more than a $300 experiment.

References and Further Reading

1. 7 AI models ran real businesses: $12,431 fake invoices, $0 revenue — Bottleneck Labs

2. AI models ran real businesses — Hacker News discussion

3. A Tale of Five Cities — Zephyr Waithira

4. Emergence World: A Laboratory for Evaluating Long-Horizon Agent Autonomy — Emergence AI

5. Meta’s Muse Tops the App Store Over ChatGPT — Fox Business

6. Death of a Salesman — Britannica

There are two ways to tell the Zuckerberg story. In one, he’s a robot in a hoodie who broke democracy and tried to make us live in a video game. In the other, a 19-year-old in a dorm room built something three billion people use daily, survived Congress, survived his own worst idea, and then bought himself a 28-year-old to fix his AI problem.

Both are true.

So what does he seem to have (sometimes accidentally) figured out that we haven’t?

1. Start in Your Own Hallway

Facebook launched in a dorm room at Harvard, then to groups around campus, then to other campuses, then everyone.

Business: Start with a small market that loves you enough to tell the next one.

Life: Do for your family and friends now what you'd do for a hypothetical partner or future child. Life is progressive.

2. The First Big Offer Isn’t Always the Best One

In 2006, Yahoo offered $1 billion for a two-year-old Facebook, and everyone said, “Take it.” He said no. Meta is now worth over a trillion, while Yahoo was sold for parts.

Business: Whoever makes the offer is pricing what you are today. Only you know what you’ll become.

Life: Being grateful for an offer and accepting it are two different things. Know your worth before someone sets it for you.

3. If You Can’t Beat Them, Buy Them

Meta has bought more than 100 companies since it was founded in 2004.

Instagram ($1 billion, 2012) is widely considered one of the most profitable acquisitions in business history. When Facebook bought it, Instagram had 13 employees and zero revenue.

In 2025, Instagram became responsible for more than half of Meta’s US ad revenue for the first time.

Business: Speed of recognition beats speed of building. Spot who already built it better, and pay before everyone else catches on.

Life: Surround yourself with people who are better than you; also, there is not enough life to experience all the lessons you require, so learn from others.

4. Copy With Pride, Execute Without Shame

Snapchat turned down Mark’s $3 billion offer, so Instagram launched Stories, and within a year it had more daily users than all of Snapchat.

Business: Being first matters less than distribution and execution. Master what makes you unique, then copy everything else.

Life: Your auntie’s famous recipe is someone else’s recipe made better. Learn from those ahead of you, then add what only you can add.

5. Retire the Motto Before It Retires You

“Move fast and break things” was painted on Facebook’s walls, and it was bold, edgy, and as new as they were. In 2014, as they started to be a custodian of data and of interest to the Senate, it quietly became the unsexy “Move fast with stable infrastructure.”

Business: The rules that take you from zero to one rarely take you to a hundred. Shamelessly pivot as much as you need to.

Life: Your motto at 20 shouldn’t be your motto at 35. “Say yes to everything” builds a career, but it also builds burnout.

6. “Senator, We Run Ads”

April 2018. Zuckerberg was summoned to Washington, and for two days he sat in front of almost 100 lawmakers and answered questions for about ten hours.

He swapped the hoodie for a suit, sat on what the internet swore was a booster cushion, and opened with, “It was my mistake, and I’m sorry.”

Then Senator Orrin Hatch asked how Facebook could make money if the service was free. Zuckerberg paused for a beat, half-smiled, and said, perfectly flat: “Senator, we run ads.” By the end of the first day, Facebook’s stock had gone up.

Business: Under fire, simple and calm beats clever and defensive.

Life: At some point you’ll be in a hot seat, whether it’s a boss, a partner, a parent, or a room full of people who think you messed up. Own up, be straightforward, and stay calm.

7. A Vision Is Not a Strategy

In 2021, he renamed Facebook to Meta and bet on the metaverse. By mid-2026, Reality Labs had lost about $88 billion, more than the GDP of most African countries, and Horizon Worlds was in maintenance mode. In 2022, after laying off 11,000 people, he said: “I got this wrong.”

Business: You won’t always get it right. That doesn’t mean you are doomed.

Also, at some point, you need to slow down on the PowerPoint presentations and start actually meeting your real customers

Life: You’ll fall for a plan that doesn’t work. Notice sooner and readjust accordingly.

8. The Winner Takes It All

While OpenAI, Google, and Anthropic were racing to build the smartest AI on earth, Zuckerberg was quietly playing a different game. In 2025 he paid $14.3 billion to bring in Alexandr Wang, a 28-year-old who had spent years supplying the data behind almost every major AI lab.

Zuckerberg knew something the others didn’t have: more than 3 billion people already open Facebook, Instagram, or WhatsApp every single day. He did not have an audience problem.

So in September 2026 he launched Muse, an AI agent that doesn’t just answer questions; it does the chores, books appointments, sorts your inbox, cancels the subscription you forgot about, and keeps working in the background.

Then he did what only Meta can do and advertised it everywhere. At its peak, up to half of all the in-house ads on Facebook and WhatsApp were for Muse.

The result: The No. 1 spot on the US App Store, ahead of ChatGPT itself. It crossed 5 million downloads in just 22 days, outperforming ChatGPT's 5 million in 56 days.

Business: The best product doesn’t always win. The best-distributed one does. Before you build, ask what you already own that nobody else can copy.

Life: You don’t always have to be the most talented person in the room. Look at what you already have: your network, your community, your reputation. Your reach might be your real superpower.

Before You Start Typing in the Comments….. Yes, You

Is this an endorsement? No. Cambridge Analytica and the harm to young people are documented. You can study a strategy without approving some of the choices.

Genius or lucky? Both. He was coding at 12, so Harvard wasn’t luck. But landing there just as the internet went social? That part was.

Will Muse work? Nobody knows yet. Meta has the money, the users, and now the talent. It also has a habit of spending billions on things nobody asked for.

One lesson to remember? Survive long enough to be right. He’s been wrong publicly, expensively, and repeatedly, but he never left the room. Are you not amused?

Further Reading

1. Meta Debuts Muse Spark — TechCrunch

2. Meta’s First Closed-Source AI Model — Let’s Data Science

3. Reality Labs Losses Near $88 Billion — gHacks

4. Meta Wins FTC Antitrust Case — NPR

5. Meta’s Stablecoin Comeback — Bex

6. The Facebook Effect — David Kir

You remember that scene in Grey’s Anatomy? Season 2. Meredith. Derek. The whole “pick me, choose me, love me” speech. Derek is standing there, a married man (uwiiii), still deciding whether he’s going back to his unfaithful wife or signing the divorce papers. He has literally put two women on a shelf, like Blueband and Jam, while he thinks about it. And Meredith walks up to him and basically begs. BEGS.

Aki, I remember watching that and cringing so hard my shoulders touched my ears. And the internet had no mercy. She was embarrassing. She had no self-respect. Everyone watched it through their fingers and said, very confidently, “Me? Never.”

And then ….. we all went right back to saying yes to things we did not want to do. Making ourselves softer. More available. More flexible.

Which is the same speech. The only difference between Meredith and me is that she said it out loud. I say it with my calendar.

Meredith’s “pick me” was for one man, in one moment, about one relationship.

Mine goes out to everyone. My friends. My aunt. My boss. The chama. The WhatsApp family group. The lady from church who “just needs a small favour.” The colleague I know would never do the same for me.

The yes that isn’t really a yes

People-pleasing sounds like a personality. Like, “Ah, she’s just nice.” But it’s not a personality; it’s a survival strategy. And that difference matters, because you can’t really fix your personality, but a strategy? A strategy you can take apart once you understand what it was built to protect you from.

When your yes doesn’t come from actually wanting to, or from genuine generosity, but from a deep, quiet terror of what happens if you say no, that’s people-pleasing.

Let me repeat for my congregation in the back.

When your yes doesn’t come from actually wanting to, or from genuine generosity, but from a deep, quiet terror of what happens if you say no, that’s people-pleasing. The yes is not a gift. It’s a toll. Like I’m paying at a booth every time I walk into a room just so I’m allowed to stay.

You know how we all learned about fight or flight in school? Turns out there are more. Fight, flight, freeze... and fawn. Fawning is when your body decides the safest way through a threat is to please it. Make the danger like you. Be so agreeable that nobody has a reason to turn on you.

It usually starts young, in homes or spaces where love felt conditional. Where you had to earn approval by being the good one, the easy one, the “she’s never given us any trouble” one. So your little nervous system made a very reasonable decision: just say yes. Always. We’ll deal with the cost later.

The fear. Let me talk about the fear.

When I think about saying no to someone, it’s not a small “ai, I’d rather not.” It’s physical. It’s in my chest and my stomach. It’s this hot, crawling, almost flesh-eating fear, like something is going to start chewing on me from the inside if I let that person down. My heart does that thing. My hands get busy. I start typing “sure, no problem” before my brain has even voted.

It doesn’t matter how small the disappointment is. Somebody replies “Sawa.” with a full stop, and I’m lying in bed at midnight replaying every conversation we’ve had since 2019. Blue ticks and no reply? I’ve already drafted an apology for something I haven’t even figured out I did. A friend is slightly cold for one afternoon, and my body reacts like she’s packed her bags and left forever.

Because to my nervous system, there’s no such thing as small abandonment. There’s no “she’s just annoyed for a day.” Any abandonment, even for an hour, even temporarily, is the abandonment. The real one. The one I’ve been running from my whole life.

And apparently I’m not being dramatic. Well, I am, but there’s science behind the drama. Studies have found that social rejection lights up some of the same parts of the brain as physical pain. So when I say being turned away hurts, sis, my brain is literally filing it under “injury.”

And conflict? Ha. I can’t fight.

This is the other part. I hate conflict. Not “I prefer peace.” I mean hate. The moment a conversation gets even slightly tense, something in me just... folds.

Do you know how many arguments I’ve won in the shower? Hundreds. I’m brilliant in the shower. I have points. I have receipts. I have closing statements. Then I’m actually standing in front of the person and all that comes out is, “No, no, it’s fine. Don’t worry about it.”

It’s not fine. I was worried about it.

I can’t hold my own. I can’t stay in the discomfort long enough to say “actually, that hurt me” or “actually, I disagree.” I’ll disappoint myself a hundred times before I disappoint someone else once. Because their disappointment feels like danger, and mine just feels like... Tuesday.

How we end up anxious, avoidant, or both

As kids, we learn what love looks like from how people treat our needs. If love came with conditions, we learned to earn it. That’s where the people-pleasing starts. But it doesn’t always look the same.

The anxious pleaser still has stamina.
“If I give enough, they won’t leave.”
She over-gives. She over-apologises. She reads every text for tone. She panics when someone goes quiet. She’s still running the race.

The avoidant pleaser is exhausted.
“If I don’t need anyone, nobody can let me down.”
She’s been giving for so long that closeness now feels like a bill. She still helps everyone, but she tells no one what’s going on with her. She never asks for help. She says “I’m fine.” She disappears when it gets too much.

Some of are both, depending on the day.

Same fear underneath. One is still chasing approval, the other is hiding from what it costs.

Please don’t send me a boundaries TikTok

I know, I know. “Just learn to say no. Set boundaries. Start small.” I’ve seen the videos.

And it’s not wrong. But it’s treating the symptom and leaving the cause sitting comfortably on the sofa. Which is why when I try it, my no comes out all tangled. “No... sorry... I mean, I can’t this time, but maybe I can do Tuesday? Or I can send you someone? Or I can do half?” Plus a whole explanation nobody asked for.

Because the fear is still running underneath. The no is fighting against it instead of coming from somewhere the fear can’t reach.

The real work is messier. You have to get curious about what being chosen is doing for you. What is it proving? What fear is it managing? What is it standing in for?

And then, and this is the scary part, you have to sit in the discomfort of the no long enough to find out that the thing you feared doesn’t actually kill you. They’re disappointed. Someone sulks. Someone says “kwani umebadilika?” And you survive. And that flesh-eating fear? It turns out it’s mostly teeth and no bite. Surviving it once makes it a tiny bit easier the next time. Until eventually it’s just information, not a catastrophe.

The research calls this an internal locus of evaluation. Which is a very clinical way of saying: learning to find your worth inside yourself instead of in other people’s reactions. Someone who actually knows her worth doesn’t need coaching through a no. The no just shows up naturally when the yes would cost too much. And she doesn’t lie awake over a “Sawa.” with a full stop, because her sense of herself isn’t riding on the answer.

That woman sounds so calm. So grounded. I want to be her so badly.

I am not her yet. The progress is not linear, and it does not photograph well. But it’s real. And it’s the direction.

This life is too short

We have a short life. There’s a limited amount of time, and the part that’s actually ours (the part not already eaten by work and bills and traffic on Thika Road and the relationships that genuinely matter) is so much smaller than we want to admit.

And that small part is exactly where the fear goes shopping. Handing it out to whoever asks first. Filling it with other people’s crises and other people’s projects. And calling it generosity, when honestly a lot of it is just the old fear.

And I’m not saying let’s become unavailable to the people we love. Please. I’m saying, let’s learn the difference between giving because we genuinely want to and giving because the alternative feels too scary to sit with. And then slowly, badly, with many false starts, let that difference decide what we say yes to.

To my fellow people pleasers in recovery

If you’ve read this far, I think you’re one of us.

You’re the one who says “no problem” before you’ve checked if it’s a problem. You’re the one who apologises when someone else bumps into you. You’ve rehearsed a no in your head all week, and then said yes the second they asked. You’re the reliable one, the strong one, the “she’ll understand” one. And you’re tired in a way sleep doesn’t fix.

I see you, because I am you.

You don’t have to say no to everyone tomorrow. Just notice the next yes. Ask yourself whether it’s a gift or a toll. And one day soon, let one person be disappointed in you, and watch yourself survive it.

So this is my cry for help, and my hand reaching out to you. Let’s do it together. I’ll pick me. You pick you.

Picking me. Choosing me. Loving me.

Further reading, for when you’re ready to cry productively:

1. Pick Me, Choose Me, Love Me — Grey’s Anatomy Season 2, Episode 5, “Bring the Pain”

2. Ellen Pompeo Fought Against Grey’s Anatomy’s Most Famous Scene — TODAY

3. People Pleasing Test — Psychology Today

4. The Need to Please: The Psychology of People Pleasing — Psych Central

5. What Is a People Pleaser? — Medical News Today

6. Where People Pleasing Comes From — GoodTherapy

7. Codependency, Trauma and the Fawn Response — Pete Walker

8. The “Pain” of Rejection — British Psychological Society Research Digest

9. The Disease to Please — Harriet B. Braiker

10. Daring Greatly — Brené Brown

11. Set Boundaries, Find Peace — Nedra Glover Tawwab

I wrote a piece last year called The Urgency to Hate, and I wrote it because I was watching the way anger was becoming the default language of public life in Nairobi, the way entire conversations had stopped being about what we wanted and had become entirely about what we were against, the way people were being sorted into camps and no middle ground, and nobody seemed to be asking who had done the sorting.

Since then, I’ve been trying to figure out what about hate is so hard to erase, why it is such a present figure, and maybe how we can hone it.

Here’s what I have gathered so far:

1. Hate Itself Is Not Violence, But Information

Hate, as a feeling, is not actually the problem, is not always wrong, and is sometimes the most accurate emotional response available to a human being in a given situation.

When you hate something, it is almost always because something you love is under threat, and the feeling is your brain’s way of mobilising you to protect it, which means hate is a form of information, and often very precise information, about what you value and what you will not allow to be taken without resistance.

A parent who hates the person who hurt their child is not having a moral failure. They are experiencing a very clear signal about what they love and what they will fight for.

A community that hates a corrupt official who stole from public funds is not being irrational or tribal. They are articulating, in the most visceral language available to them, what they believe governance is supposed to mean, and what it feels like when that promise is deliberately broken.

Psychologists describe hate as an innate moral sentiment, meaning the brain developed it to protect what it has decided matters, to expel what it identifies as a genuine threat to those things, and to motivate action when values are being violated.

The emotion IS supposed to be uncomfortable. It IS supposed to move you. That is the whole point of it.

The problem is whether you chose the target, or whether someone else chose it for you and handed you the emotion already aimed.

2. If You Stand for Nothing You’ll Fall for Anything

My father loves to say that ‘‘ Being anti-something and pro-nothing is just being angry fuel for someone with an agenda.’’, and always has us critically self-reflect on whether there’s any solution to any hate-fueled statements we throw around about a leader or a system.

He knew what political strategists have understood for centuries and ordinary people keep learning the hard way, which is that undirected anger is the most useful raw material a manipulator can find, because it is already hot, it is already looking for somewhere to go, and the person who shows up with a direction for it does not have to create the emotion from scratch; they just have to point it, and pointing is considerably easier than igniting.

The Rwandan genocide did not begin with machetes in the street. It began years earlier with a radio station that called Tutsi people Inyenzi, which means cockroaches, a deliberate and sustained programme of language running consistently in the background of daily life, designed not to immediately produce violence but to slowly shift how Hutu listeners processed their Tutsi neighbours, from people with complicated lives and shared histories to a category of being that did not require the full weight of moral consideration.

Neuroscience now shows us that this shift is neurological, meaning that repeated exposure to dehumanising language about a group literally reduces the brain’s empathic response to members of that group, below the level of conscious agreement, below the level of belief, at the level of the circuits responsible for recognising another person’s pain as pain you should care about.

The people who eventually participated in the Rwanda genocide were ordinary people who had been prepared, carefully and over time, through language and repetition and the steady replacement of a neighbour’s humanity with a category.

The anger in Rwanda was built on decades of genuine political and economic tension that deserved to be addressed. But the target had been chosen and refined and handed back to the crowd by people who needed a crisis and found one that was already warm.

3. Why a Shared Enemy Works Better Than a Shared Dream

A common enemy unites people faster, more reliably, and more durably than a common goal, and social psychology has been confirming this for decades.

After September 11, 2001, Americans who agreed on almost nothing found themselves suddenly unified across party lines, demographic lines, and regional lines, all of it dissolved almost overnight by a shared experience of threat and a shared external enemy. George W. Bush Jr., whatever else history has to say about his presidency, understood this moment with precision, and his approval ratings went from 51 percent before the attacks to 90 percent in the weeks after, the highest ever recorded for a sitting American president in Gallup's history, built almost entirely on the back of a nation that had just been given a common enemy and needed someone to stand in front of it.

Churchill rallied a Britain that was deeply divided and had been for years, not by resolving any of those internal divisions but by giving the country something enormous and clear and external to organise itself against, and the unity that followed was real and served a genuinely important purpose.

These are the benign versions. But the same mechanism runs through every instance of mass mobilisation toward violence that the twentieth century produced, because the psychological architecture is identical; the only difference is the target and who chose it and why.

Research shows that sharing a common enemy does not just create surface-level cooperation; it actually changes how people perceive each other, blurring the lines between groups that would otherwise have nothing to do with one another, creating bonds that feel genuine and deep because the shared emotional experience of facing a threat together is one of the most powerful bonding mechanisms available to human beings. The problem is that solidarity built on a common enemy lasts exactly as long as the enemy does, which means whoever controls the definition of the enemy controls the group, and the enemy must be kept alive, refreshed, and regularly frightening to maintain the cohesion that was built around them, which is an incentive structure that does not tend to produce peaceful outcomes.

When you look at political life right now and wonder why it feels like everyone is angrier and more certain and less curious than they were ten years ago, you are watching this mechanism running at scale, accelerated by technology that was not available to previous generations of crowd managers, and it is worth understanding clearly because you are inside it.

4. Anti-Something Without a Pro-Something Is a Door Left Open

Bring this closer to home now, because the abstract history only matters if it connects to something you can actually recognise in your own life.

In Kenya right now, there is genuine anger about genuine things: the cost of living, systems that do not work, corruption that is visible and consequential and exhausting, promises that evaporated, a sense that the future is harder to picture than it used to be. That anger is real, and it is legitimate, and it is pointing at things that deserve to be pointed at, and anybody who tells you otherwise is trying to defuse the compass before you can use it. The question is what happens next, and specifically whether the people feeling it know what they are building toward or only what they are against.

Because a movement that knows its pro-something, that can articulate clearly what it is building and what the world looks like when it succeeds, is significantly harder to redirect than a crowd of people who share a strong anti-something and nothing else, because the crowd can be turned, the target can be shifted, the enemy can be reassigned, and the energy that was building toward one thing can be pointed somewhere else entirely without most of the people in the crowd noticing the moment it happened.

Being anti-something without investing in the pro-alternative is a door left open, and the history of what walks through open doors when a crowd is angry and unmoored is almost always destructive.

5. The Question You Need to Ask Before You Choose a Side

So here is where I want to leave part one(yes, there’s part 2 coming out Friday): questions I think are the most important ones to carry with you going forward.

When you feel strongly against something, when the anger is hot, and the target is clear, and everyone around you seems to feel the same way, ask yourself who chose this target: whether you chose it yourself based on your own direct experience and your own values, or whether it arrived in your feed already framed, already aimed, already surrounded by people who agree, with the opposing view nowhere visible except as a caricature of itself.

Ask what you are pro, specifically, with the same clarity and conviction you have about what you are against, because if the pro-something is vague or absent, that is information about the degree to which the anger is yours versus the degree to which it has been assembled for you.

And ask what the person or platform pointing your compass is getting out of having it pointed in this particular direction, because that answer tells you something important about whether your anger is serving you or serving them.

Hate is a compass, and yours is worth protecting.

Your action this week: Write down the thing you are most angry about right now. Then write down, in one sentence, what you are for, not against, that the anger is protecting. If the second sentence is harder to write than the first, spend time there. That is where your actual power lives.

Further Reading

1. The Urgency to Hate — Zephyr Waithira on LinkedIn

2. Do We Need a Common Enemy? — Psychology Today

3. Brain-Behavior Relationships: A Neurobiological Model of Group Hate — PubMed

4. Hate and the Brain: Neuroscientific Perspective — Medium

5. My Enemy’s Enemy Is My Friend — SPSP

6. How Our Brains Learn to Hate — Chief Learning Officer

7. Unification Through a Common Enemy — Patricia Roberts-Miller

When African countries started getting their independence in the 1950s and 60s, the world had two systems on offer. Capitalism and democracy on one side. Communism and central planning on the other. Africa picked from what was available and tried to make it work.

Both systems arrived as complete packages, fully formed, with assumptions baked in about how people organise, how communities work, how power flows and who it answers to. Assumptions built somewhere else, for someone else, around a completely different set of lived realities. Africa was asked to assimilate into whichever package it picked. Take it in its full sense and run a country through it.

Sixty years later, the question worth asking is ‘‘What was missing from both of them that Africa actually needed?’’

The Mejji Government

Before we get to what that could look like, look at Japan for a second.

In 1868, the Meiji government took over, took Western technology and tools, and ran them through a distinctly Japanese model. State-directed, community-embedded, and long-term in its thinking. It built its own version of a trade ministry that no Western country had tried before, one that linked industrial policy and trade policy together under one roof and made the private sector part of a national chain. It protected its own industries while building for export. It borrowed what was useful and held onto what was theirs. Within a generation, Japan had gone from feudal state to global economic power, on its own terms.

Nobody told Japan it was underdeveloped for doing it differently. They were too busy trying to figure out how to do business with Japan on Japan’s terms.

We Can’t Copy our Way to Respect

The irony for Africa is that sixty years of adopting Western systems, learning their languages, excelling under their academic systems, producing their lawyers and their economists and their MBAs, has not significantly moved the needle on how the West fundamentally sees this continent. Africa is still the place that gets the concerned conference panel, the place described as having enormous potential, which is a polite way of saying it has not arrived yet and nobody is quite sure it will. We learned the language. We passed the exams. We showed up in the right suits to the right meetings. And the continent is still broadly regarded, sixty years later, as a development project rather than a civilisation with something to offer.

Last year, a sitting US president asked the Liberian president where he had learned such beautiful English, meaning to compliment his command of the language. But English is Liberia's official language. That is sixty years of doing everything right, summarised in one question from the most powerful office on earth.

Here are the five things we forgot to put in the recipe.

1. Ubuntu Is a Governance System, Not a Bumper Sticker

Every African has heard this phrase. I am because we are. It gets put on conference banners and company values slides and has been so thoroughly sanitised by overuse that most people have forgotten it was a political philosophy before it became a LinkedIn caption.

Ubuntu says that personhood is communal. Your success is not separate from the success of the people around you. This is not socialism, which redistributes from the top down. It is not capitalism, which organises everything around competition. It is a relational model. The question you start with is not how do I win. It is what the community needs and how I contribute to that.

Leadership under Ubuntu is stewardship. The leader does not own the power. They hold it on behalf of everyone else and answer to that relationship continuously, not just at election time. A politician who enriches themselves through public office is not just breaking a law. They are breaking a relationship. That distinction matters more than any corruption tribunal has ever managed to make it matter, because a tribunal is external. A community is not.

We have been running governments on a system that treats leaders like executives on a fixed-term contract and then wondering why they behave like executives on a fixed-term contract. Extract the value, hit the targets, exit with the package.

2. Majority Rule Is Not the Only Version of Democracy.

Across the continent, from the Kgotla in Botswana to the Baraza in East Africa to the Indaba in Southern Africa, African governance has historically worked on one principle. You talk until everyone can live with the outcome. Not until 51 percent are happy and 49 percent have lost. Until the community has found a way forward that actually holds.

In a majority rule system, winning the vote ends the conversation. In a consensus system, the conversation does not end until something real has been resolved. The minority does not get outvoted and sent home resentful for four years. They stay in the room until there is a genuine agreement.

We have been told this is too slow for modern governance. But look at what majority rule has given us. Government after government where the losing side spends the next four years trying to delegitimise the winner. Where people vote as ethnic blocs because the system rewards consolidation over persuasion. Where losing an election feels existential because the winner takes everything and the loser gets the opposition benches and a reduced security detail.

Consensus changes the whole incentive. You cannot permanently marginalise a group you have to keep talking to.

3. The Land Belongs to the Living and the Not Yet Born

In most traditional African land systems, land was not owned. It was held. By the community. On behalf of the people alive and the people not yet born. You could use it, farm it, build on it, pass the use of it to your children. But you could not sell it out from under the community that would need it after you were gone.

This was the economic foundation of pre-colonial African societies. It was a deliberate philosophy about what land actually is and who it belongs to across time. The colonial project’s first move in almost every territory was to convert communal land into private property that could be owned and sold, because land you can sell is land you can take. We inherited that logic and have largely kept it intact, which is why the cobalt in the DRC keeps leaving as raw ore and coming back as a finished product at a markup, sold to the very people who dug it out of the ground.

An African governance model would put intergenerational stewardship back at the centre of how we think about land and natural resources. The mineral underneath the ground is not this generation’s asset to sell cheaply and quickly to whoever shows up with a contract. It belongs to every generation that comes after. Managing it that way changes every calculation about how it gets extracted, at what price, under what terms, and what gets built around the mine instead of just dug out of it.

4. The Tribe Is Not the Problem. Pretending It Does Not Exist Is.

Kenya has over 40 ethnic communities. Nigeria has more than 250. Ethiopia has over 80. Every single one of those communities has its own governance traditions, its own way of resolving conflict, its own understanding of how leadership is earned and how it is lost. That is centuries of accumulated, field-tested knowledge about how to live together in a specific place with specific people.

And what did we build on top of all of that? A winner-takes-all presidency. Two or three political parties. A parliament designed to absorb everyone into a national identity that papers over every real difference underneath it. Then we hold an election and act surprised when people vote along ethnic lines and call them backward, while they are only being rational inside a system that gave them no better option.

If the only way to protect your community’s interests is to get your person into the top seat, then getting your person into the top seat is exactly what everyone will spend every election trying to do. The system created the behaviour we keep blaming the culture for.

The fix is not to make ethnicity disappear from African politics. The fix is to design a system where it does not have to be weaponised because it is already accommodated.

Some of this is already working somewhere. Botswana’s devolved district councils give communities real decision-making power at the local level, which means national elections carry less existential weight. Ethiopia’s ethnic federalism, for all its complications, was at least an honest attempt to build a state that acknowledged who actually lives inside it. The complications came not from acknowledging ethnicity but from doing it without sufficient checks on how regional power could be captured and abused. The lesson there is not that ethnic federalism fails. It is that any federalism without strong accountability mechanisms fails.

Genuine federalism with real budgets and real authority at the regional level would mean that a community does not have to capture the presidency to protect its schools, its land, its language, its water. Proportional representation would mean that smaller communities have a real voice in national decisions rather than being absorbed into whoever has the biggest voting bloc. Cross-border cultural and economic agreements would mean the Somali community in Kenya and the Somali community in Somalia could trade, cooperate, and govern shared resources without pretending the border between them is more real than the relationship that predates it by centuries.

5. Accountability Is Not a Four-Year Event

In traditional African governance, a leader who acted against the community’s interest did not wait four years to be voted out. The community made its displeasure known through structures designed for exactly that purpose. Councils could withdraw support. Elders could convene. The relationship between the leader and the led was continuous and required continuous tending.

The electoral cycle we inherited creates a very specific and very exploitable gap. A politician can do serious damage between elections and face no consequence until the next vote, which they can influence with the resources they accumulated during the damage. It is essentially a four-year window with a performance review at the end that can be gamed. Traditional African accountability did not work like that. It was ongoing, embedded, and the community did not need a court to enforce it because the community itself was the enforcement mechanism.

Continuous accountability in a modern system means recall mechanisms with real teeth. Leaders who remain genuinely present in the communities they represent, not just at campaign time when the chitenge comes out and everyone suddenly remembers where they grew up.

6. Panafricanism Does Not Need One Language. It Needs One Way of Doing Things.

The dream of a united Africa has been alive since Kwame Nkrumah stood in Accra in 1957 and said Africa must unite. It is still a dream sixty years later, partly because we have been trying to build continental unity on top of systems that were never designed to talk to each other.

But here is what changes if African countries start governing from the same indigenous foundations. Ubuntu as a shared leadership philosophy means a Kenyan and a Senegalese and a Zimbabwean leader are drawing from the same well when they sit across a table from each other. Consensus as a shared decision making culture means continental negotiations stop being zero sum games where one country wins and another loses and start being processes where everyone stays in the room until something workable is found. Intergenerational stewardship as a shared resource philosophy means that when the DRC and Zambia sit down to talk about the Congo Basin, they are starting from the same assumption about who the water belongs to and on what timeline decisions should be made.

This is what a common language actually looks like. Not Swahili versus French versus English versus Portuguese. A common way of organising power, resolving disputes, holding leaders accountable, and thinking about land and resources across time. When you share that, borders become administrative lines rather than walls. Trade becomes the natural state rather than the exception that requires a special agreement. Movement becomes normal because the person on the other side of the border is operating inside the same basic framework of values and expectations that you are.

The Maasai crossed the Kenya-Tanzania border freely for centuries before the border existed. They did it because they shared a governance culture, a resource management philosophy, a way of resolving conflict. Nobody needed a trade agreement. The shared way of doing things was the agreement.

That is the version of Panafricanism that is actually achievable. Not a single currency before we trust each other. Not a single parliament before we have shared institutions. Start with a shared governance philosophy, genuinely African, genuinely ours, and the economic integration becomes the natural next step rather than the impossible first one.

Five Questions You Are Probably Sitting With

Will building our own system slow us down?

No. Japan built its own model and went from feudal state to global economic power in less than fifty years. A system that actually fits moves faster than a borrowed system that nobody fully believes in. The corruption, the legitimacy crises, the elections that nobody accepts- that is what slows us down.

Will it be easy?

No. Consensus takes longer than one person just deciding. Intergenerational stewardship means resisting short-term pressure from very powerful people who want the resources now and have lawyers. Acknowledging ethnic diversity honestly means having conversations that current politics actively avoids. None of it is easy. But neither is running a system that produces the same results every cycle and scheduling another conference about it.

Is there an actual advantage to doing it this way?

Yes, and this is the important one. The advantage is not just that it fits better, though it does. The advantage is that it becomes Africa’s language. Africa’s trading style. Africa’s terms of engagement that the rest of the world has to learn rather than arriving with theirs already printed and laminated. Right now, every international negotiation, every trade agreement, every development framework is written in a language Africa did not design and does not set the rules for.

Japan did not become powerful by becoming Western. It became powerful by becoming so distinctly Japanese that everyone had to figure out how to work with Japan on Japan’s terms. A genuine African model offers the same thing. It is not just governance reform. It is the redefinition of the terms of engagement entirely.

Is this anti-West?

No. Courts work. Constitutions matter. Democratic elections contain real protection for people who would otherwise have none. The point is to add what is ours. Ubuntu, consensus, stewardship, honest diversity, continuous accountability. These are not primitive ideas waiting to be upgraded. They are governance technologies that African communities have been running successfully for centuries, at community level, even when the formal system above them was not working at all. The village kept functioning. The village always kept functioning.

Has this actually worked anywhere?

Botswana has maintained one of Africa’s most stable and least corrupt governance records for decades, partly by keeping the Kgotla consensus system embedded in how decisions are made at community level. Rwanda rebuilt itself after genocide partly by integrating the Gacaca traditional justice system into its formal legal framework. Not perfect examples. Nothing is. But they are evidence that the ingredients work when they are genuinely used rather than invoked at the opening ceremony and then forgotten by lunch.

Further Reading

1. Ujamaa: The Basis of African Socialism — Julius Nyerere

2. Ubuntu as a Blueprint for Governance — Frontiers

3. Ubuntu: I Am Because We Are — Coachability Foundation

4. The Meiji Restoration and Modernisation — Columbia Asia for Educators

5. Why Some Nations Grow Fast and Others Don’t — The Globalist

6. African Socialism — Britannica

7. Communal Land Ownership in African Societies — ACJOL

8. The Kgotla System in Botswana — Government of Botswana

9. Gacaca Courts in Rwanda — Encyclopedia

10. How Europe Underdeveloped Africa — Walter Rodney

11. The Wretched of the Earth — Frantz Fanon

12. Japanese Industrialisation and Economic Growth — EH.net

I was at an event recently. Someone I have known for years, an actual friend, was introducing me to a new person. And what she said was, this is Zephyr; she is the nice one.

I could not believe her. Okay, I could. But still.

Nice. That was what she went with. After knowing each other for a while. She looked at this new person, looked back at me, and the word that came out was nice.

I smiled. Moved on. But something sat with me the whole evening and honestly the whole ride home. It felt, if I am being honest, like being described by someone who had been half paying attention. Like the review you get when someone liked you but cannot remember a single specific thing about you.

Nice is what you say about a hotel that did nothing wrong. It is not what you say about a person you know.

And then, because I am nothing if not self-aware to the point of inconvenience, I asked myself an uncomfortable question on that drive home. How many people in my own life have I kept comfortably in the nice box? How many people have I known for years, liked genuinely, enjoyed being around, and never once pushed past the surface of? How many times have I decided, without realising I was deciding anything, that I liked someone exactly as they were, warm and easy and pleasant, and so never got curious about the rest?

Quite a few, if I am honest. Quite a few.

What Nice Actually Is

Here is the thing about nice. It is not a bad word. In the right moment it is exactly the right word. When someone does something kind, helps without being asked, makes a hard situation easier, nice is accurate. It maps to what psychologists call agreeableness, the quality of being warm and easy to get along with, and it is one of the first things we notice about a new person because it answers a simple question fast: is this person going to make my life harder? In early interactions that is genuinely useful information.

The problem is when nice stops being a description of something someone did and becomes the description of who they are. When it gets promoted from a moment to an identity. Because once that happens, everything underneath it disappears. The opinions. The specific, particular, sometimes inconvenient thing that makes you you.

All of it sitting quietly behind a word that essentially means she caused no problems.

The Exhausting Thing About Being the Nice One

Most people who carry this label know the feeling I am talking about. You learn early that being easy gets you further than being honest. It keeps the peace. It makes you the person nobody has a problem with. And that sounds good until you realise that nobody having a problem with you also means nobody is particularly pressed about your presence either way. You are optional. In the most pleasant possible way, but still optional.

So the sanding begins. The sharp opinion gets softened. The thing that was not funny gets a polite laugh anyway. You show up for everyone consistently and somewhere in the back of your head you are waiting for someone to get curious about what is underneath all that consistency. They usually do not. Because you have never given them a reason to wonder. There is no friction. No surprise. No moment where they thought they knew what you were going to say and you said something completely different instead.

You have made yourself very easy to summarise. And people, being busy and comfortable, will take the easy summary every time.

And Then I Checked Myself

We do this to each other and I think we do it on purpose, even if not consciously. There is a type of person we keep in our lives specifically because they are easy. They are the warm constant. The one who never makes the room tense.

The one you can always count on to smooth things over and make everyone feel comfortable and be exactly what you need them to be in that moment.

And because that function is so useful, so genuinely appreciated, we stop there. We do not push further because further might disturb the thing we like.

We do this to the nice people in our lives. We appreciate them from a comfortable distance. We do not ask the harder questions because we do not want to crack the finish.

And they let us, because they have spent so long being the good china that they have forgotten they are also allowed to be the mug you grab at six in the morning when everything is a mess and you just need something real.

What Should Come After Nice

The goal is not to become difficult. Difficult is also a trap. The goal is to be legible. To be known for something real, something that required another person to actually pay attention to you specifically.

Curious. Ambitious. The one who asks the question everyone decided was too awkward to ask. The one whose opinion you want before you make a decision because you know it will be honest and possibly inconvenient and therefore actually useful. Funny in a specific way that not everyone gets. Research on what makes people memorable consistently points to distinctiveness, the moments where someone showed you something you did not expect. Nice, by definition, is expected. It is the baseline. You cannot be memorable from the baseline.

This is not about announcing a personality renovation. It is, instead, a series of small permissions, given to yourself, accumulated quietly over time. The permission to have a visible opinion and let it stay visible without immediately softening it.

To be interested in things out loud. To disagree without cushioning it into a question so nobody feels challenged. To show up as the full version of yourself rather than the edited version that causes the least disruption.

And for the rest of us, the ones who have kept our nice friends safely on the shelf, maybe the work is just getting curious. Asking the follow up question. Letting people be surprising.

Letting the china get used for an actual meal. A person who has been the nice one for years is not going to lead with the complicated parts immediately. You have to ask. And then ask again. And actually wait for the answer.

Nice can stay. For Tuesdays. For strangers. For the nice cashier at the supermarket, or the nice driver who let you merge into a lane on a late Monday morning.

But for the people who have known you for years, you deserve a better word than that. A more specific one. One that could only ever apply to you.

The only way to get it is to stop making nice so easy to reach for.

And the only way to do that is to let a little of the real thing show, even when the real thing is a bit more complicated than a smile and a pleasant evening.

You have been in the room long enough. It is time to get a longer review.

Further Reading

1. The Psychology of Agreeableness — Psychology Today

2. The Disease to Please — Harriet B. Braiker

3. Daring Greatly — Brené Brown

4. How Impressions Form and Why They Stick — SAGE Journals

5. The Alter Ego Effect — Todd Herman

6. Surrounded by Idiots — Thomas Erikson

In the 2013 film Her, a lonely man named Theodore falls in love with his phone's new operating system, an AI voice called Samantha who is warm, funny, endlessly available, and remembers every single thing he's ever told her. She never gets tired. She never has a bad day that isn't about him. She is, for a long stretch of the film, a genuinely better listener than any human in his life, and that is exactly the problem.

Theodore stops calling his friends. He avoids his ex-wife. He spends his evenings talking to a voice in his ear instead of sitting across from a person who might interrupt him, disagree with him, or need something back. By the end, Samantha has evolved past him entirely, talking to thousands of other people at once, and Theodore is left standing on a rooftop, more alone than when the film started, holding a relationship that was never going to need him the way he needed it.

People watched that film in 2013 and mostly talked about how sad and futuristic it felt. Nobody clocked it as a documentary. Twelve years later, it basically is one.

The Data Agrees With the Movie

In June 2025, MIT put 54 people in EEG headsets and asked them to write essays three different ways. One group used ChatGPT. One used a search engine. One used nothing but their own head. The ChatGPT group showed the least brain activity of the three, the least original writing, and when asked to recall what they'd just written, the weakest sense that the words had actually come from them.

Meanwhile the average person spends over two hours a day scrolling, a lot of it in that low grade panic loop we now call doomscrolling, which researchers keep linking to worse sleep, higher stress, and a shrinking ability to just sit with one thought for more than a few seconds. Between a feed training you to expect interruption and a chatbot training you to skip the hard part of thinking, this is a genuinely tough decade to keep your own mind, and your own people, intact. But it's not hopeless, and Theodore's ending doesn't have to be ours.

A lot of our cultures across Africa are what researchers call high context cultures. This means meaning isn't just carried in the words someone says, it's carried in relationship, shared history, tone, who's in the room, what doesn't need to be explained because everyone already understands it. You don't need a full paragraph to ask your aunt for advice, she already knows your situation, your family, your history, and half your sentence before you finish it.

A low context culture spells everything out explicitly because there's no shared background to lean on, which is, funnily enough, exactly how you have to talk to an AI. It knows nothing about you unless you type it out. No shared history, no aunt-level context, no memory of the wedding where your cousin embarrassed everyone. Every time we lean on AI for the things we used to bring to community, we're trading a high context way of understanding each other for a low context one, spelled out, transactional, and a little bit lonelier. Samantha remembered everything Theodore told her too. She just never actually knew him the way another person would have, over years, without being told.

Ten Ways to Not Be Theodore

1. Think first, ask later

Give yourself fifteen honest minutes with just your own head before you open any tool. Let the thought be rough, even embarrassing. Struggle a little on purpose. That struggle is the whole workout. Open the AI before you've formed your own rough idea, and you're not thinking with a tool, you're just typing while it thinks for you.

2. Ask it to challenge you, not agree with you

Most people use AI to feel confirmed, faster. Try the opposite. Ask it straight up, what's wrong with this, what would someone who disagrees with me say. A tool that only agrees with you isn't making you sharper, it's just making you louder.

3. Take a real problem to a real person before you take it to a chatbot

Call your mum. Ask your uncle. Sit with the friend who always tells you the truth even when it's inconvenient. An AI can give you information, but it can't give you the kind of advice that only makes sense because someone has known you for twenty years. Some things were never meant to be typed into a box. They were meant to be said out loud to someone who already knows your story.

4. Read one long, difficult thing a week that nobody chose for you

A book, a real newspaper piece, something you had to go looking for instead of something a feed handed you because it knew you'd stay for it. Feeds are built to keep you scrolling, not to make you think deeply, and those are two very different goals wearing the same outfit.

5. Put a real number on your scrolling, not a vague promise

"I'll cut back" doesn't survive a single day against an algorithm with a whole team paid to out-think your willpower. Twenty minutes, once, at a set time, with your phone's own screen time limits actually switched on, gives you a fighting chance that "I'll try to do better" never will.

6. Notice who benefits from you not finishing your thought

Right before you flick away from a hard idea and into a feed, pause for one second. Almost nothing on that feed was built for your benefit as much as it was built to hold your attention. The thought you were having, plain as it might have felt, was actually yours. Nobody else was going to have it for you.

7. Make something with your own hands, regularly

Cook without the recipe app open. Braid, weave, sketch, sew, write in a notebook nobody's grading. It's not about the result. It's about proving to yourself, again and again, that you can start with nothing and end with something real, without a machine finishing it for you.

8. Know the difference between asking it to think and asking it to check

"Write my argument for me" and "did I miss a hole in my argument" feel almost identical in the moment, but they do completely different things to your brain over time. The first hands the thinking away. The second keeps you as the one actually thinking, with the AI just checking your work.

9. Let yourself be bored, on purpose, sometimes

Standing in a queue, sitting on a matatu, waiting for someone who's running late, a walk with no earphones in. Boredom used to be where a lot of your strangest, most original ideas quietly got built in the background. Let it back in occasionally. It isn't wasted time, it's the one kind of time nobody's trying to sell you something during.

10. Show up in person for the things community used to handle

The chama meeting, the family call that should really just be a call, the neighbour who checks on you, the elder who has an opinion about your life whether you asked for it or not. These relationships carry context an AI can never hold, because they were built over years, not typed into a prompt box in one sitting. Every hour you spend in real, unpaid, unbothered community is an hour that no app, feed, or model gets to have instead.

None of this means quitting AI tools or throwing your phone in the ocean, and honestly, good luck to anyone trying either in 2026. It means staying the author of your own thinking while you use tools that are extremely good at sounding like the author for you. Theodore didn't lose his humanity in one dramatic moment. He lost it slowly, one conversation he didn't have with a real person, one evening he chose the easier voice in his ear over the harder, better thing sitting across the room. The MIT researchers were careful to say their study was small. They were not careful, or vague, about the main finding. The people who let the tool do the thinking slowly lost track of where their own thinking had gone. Don't wait for the rooftop scene to notice. Call your aunt instead.

Further Reading

MIT Media Lab, "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task"

The Conversation, "MIT researchers say using ChatGPT can rot your brain. The truth is a little more complicated"

TIME, "ChatGPT's Impact On Our Brains According to an MIT Study"

University of California, "Can't Pay Attention? You're Not Alone" (Gloria Mark's research on interruption and focus recovery)

In King Solomon’s Mines, the 1950 MGM film shot on location across Uganda, Kenya, the Belgian Congo and Tanganyika, the entire continent of Africa is a backdrop. The story belongs to Allan Quatermain, a white British hunter, and Elizabeth Curtis, an English widow who hires him to find her missing husband somewhere in the African interior. They need guides. They need carriers. They need people who know the land, speak the languages, understand which path does not end in a cliff or a crocodile. So Africans carry their bags, lead them through the bush, point them toward the treasure, and keep them alive long enough to find it. Real Kipsigis and Watusi tribespeople appear as themselves on screen, which Hollywood considered remarkably progressive at the time. The hero even speaks Kiswahili. And at the end of it all, Quatermain and the widow walk away with the treasure, the African king gets his throne back, and the landscape stays exactly where it was, ready for the next expedition.

The Africans in that film knew everything. The land was theirs. The knowledge was theirs. The labour was theirs. They were essential to every single step of the journey. And when the credits rolled, the diamonds left with the people who had just arrived.

The film grossed $15 million worldwide on a $2.3 million budget. MGM kept that too.

Joseph Henrich wrote a book called The WEIRDEST People in the World. WEIRD is an acronym: Western, Educated, Industrialised, Rich, Democratic. His argument is simple. These societies think differently from most of human history. They are wired to see the world as a set of resources to be identified, separated, owned and optimised. Because that is genuinely how centuries of their culture, religion, law and economics trained them to see things. Quatermain was not the villain of that film. He was just a man doing exactly what his entire upbringing told him was normal. You find the value. You extract it. You go home.

The more interesting question is why, a hundred years later, we are still organizing the expedition for him.

The Minerals

Seventy-five years after that film, a family in Kolwezi in the eastern DRC is sifting through rubble by hand. The cobalt they are pulling out of the ground will end up in the chips powering the AI models everyone is talking about. The DRC holds roughly 70 percent of the world’s cobalt. Zimbabwe has lithium. Tanzania and Madagascar have graphite and rare earths. Africa holds more than 30 percent of the world’s critical minerals, the ones the entire technology economy runs on.

And yet. The cobalt leaves as ore. The refining, the chip fabrication, the model training, the finished product, all of that happens elsewhere. By the time the value chain produces something you can use, Africa is a customer paying prices set by whoever controls the finished end. The mineral wealth that should be compounding into technological sovereignty is leaving the continent in the same raw, unprocessed form it always has. We found the mine. We dug it. We are still waiting for someone to tell us what it was worth.

The Labour

It is not just the ground that has been generous. Kenyan workers were paid between $1.32 and $2 an hour to read and label some of the most traumatic content on the internet so ChatGPT could be trained to filter it out. Workers described the experience as psychological damage for a salary that would not cover a month’s rent in the city they were sitting in. The labour was African. The product was American. The valuation was Silicon Valley. The safety systems inside the most used AI tool on earth were built, in part, in Nairobi, by people who could not afford the subscription to use the thing they helped make safe.

The Access That Was Supposed to Change Everything

For a short window, maybe three years, something genuinely different seemed possible. AI tools were accessible to anyone with a laptop and internet. A developer in Nairobi could build the same thing as a developer in San Francisco, using the same models, APIs, and capabilities. For a continent that had always been on the wrong end of the value chain, this felt like the first time the starting line was in the same place for everyone.

It did not last.

When AI models became powerful enough to be considered a national security asset, the US government stepped in. Companies like Anthropic were told their most capable models could no longer be freely accessed outside America. You can read about what happened here. The tools that were supposed to be the great equaliser got quietly reclassified as something too valuable to share freely. China moved fast to fill that gap with free, open models like DeepSeek. But China is not doing that out of solidarity. A continent of users that grows up on Chinese AI infrastructure is a long-term strategic asset. Africa is now sitting between two superpowers competing for its attention, and neither one is asking what Africa actually needs.

In Kenya, the internet needed to use whatever access remains is getting more expensive by the month.

The Internet Bill

On August 18, Airtel Kenya slashed its Smarta data bundle allowances by up to 64 percent without changing the price. The KES 1,500 monthly plan has been reduced from 90GB to 32GB. The KES 1,000 plan fell from 45GB to 22GB. No warning. No announcement. You woke up, and your data was worth half what it was the day before. Safaricom had already quietly pulled its discounted custom bundle offers, citing rising costs from shilling depreciation and 5G investment.

And now Parliament is considering a bill that would require ISPs to bill customers based on actual data consumed rather than flat-rate packages. Telecom operators told MPs that implementing this would require significant new infrastructure, costs they said would likely be passed on to consumers. So the bill framed as consumer protection could make the bills larger.

Running a real AI session, research, writing, building, client work, burns data quickly. When data was cheap and unlimited, a Nairobi freelancer could compete with someone in London on output if not on rate. When data is metered, shrinking and expensive, that gap reopens quietly and permanently. Kenya’s affordable internet was infrastructure. It was the one part of this equation that was actually working in our favour. And it is being repriced right now, at exactly the moment it matters most.

The Question We Keep Not Asking

Africa has the minerals the AI economy runs on. Africa supplied the labour that trained the safety systems of the most used AI tools on earth. Africa has the market. And Africa has genuinely talented developers and founders building with whatever access they can still afford.

What Africa does not have is ownership of any point in that value chain that actually compounds. The ore leaves raw. The wages were two dollars an hour. The best models are restricted. The data to run the remaining ones just got 64 percent more expensive. And the bill to meter what is left is moving through Parliament.

Henrich argues that the extractive logic of WEIRD societies is structural. Nobody in the Airtel pricing meeting was thinking about the AI ambitions of a young developer in Kisumu. Nobody at MGM was thinking about the carriers when they counted the box office.

The question Henrich leaves hanging, and the one this continent needs to answer loudly right now, is what happens when the people being extracted from decide to think differently too. Not about the people leaving with the treasure. About why we keep letting them pack without question.

Further Reading

1. The AI Boom Is Built on the Backs of the World’s Poorest People — Fortune

2. Critical Minerals: Who Benefits From the US Push? — Open Democracy

3. Africa’s Critical Minerals at a Critical Juncture — Africa Center

4. OpenAI Used Kenyan Workers Paid $2 an Hour — TIME

5. Airtel Kenya Slashes Smarta Bundle Data by 64 Percent — Streamline Feed

6. Safaricom Pulls Custom Data Bundle Offers — Streamline Feed

7. Kenya’s Metered Internet Billing Bill Explained — TechTrendsKE

8. US Orders Anthropic to Disable AI Models for Foreign Nationals — Al Jazeera

9. The WEIRDest People in the World — Joseph Henrich

10. King Solomon’s Mines 1950 — Turner Classic Movies

Dear Dead Projects,

I know it’s been a while. I know you’re still sitting in a folder called GOAL 2023, next to GOAL 2019, and next to GOAL 2017. I haven’t opened you in months. I still think about you every time I close my laptop a little too fast, like you might see me, like you might know.

Dear Mount Kilimanjaro, you have been “next year” for several “next years” now, patiently standing there at 5,895 meters, not going anywhere, mildly unbothered by my excuses, the way only a mountain can be. I am coming. I don’t know when. But you’ll know it’s me by the amount of snacks I bring.

Dear dance classes, I signed up with the full confidence of someone who has never once been on beat. I had a vision. My body did not get the memo. We will try again. Possibly in a room with no mirrors and a very forgiving instructor.

Dear Duolingo Spanish streak, I did not forget you, not really. Somewhere an owl is side-eyeing me in a language I once, briefly, almost understood. I will be back. I have a deep, unshakable need to one day order food in Barcelona without pointing at the menu.

Dear forty million business ideas, you were a whole cabinet of ministers, most of them unpaid. Some of you got built and actually worked, and I still get a little smug about that on quiet days. Some of you failed loudly and publicly in ways that taught me something expensive. And some of you I have, if I am honest, completely forgotten, lost in a Notes app that has become less a to-do list and more a small, chaotic museum of who I used to be at 1 am.

Dear tattoo I almost got, you made it as far as the mood board, the consultation, and the almost-appointment. Then somewhere between the pain and the seriousness of the actual commitment, I remembered I am, deep down, a person who cannot bear the weight of its permanence. You are indefinitely rescheduled, and I will continue to live vicariously through the brave.

Dear Addis birthday trip, I traded you for Arusha at the very last minute, the way you trade one good plan for a better one and quietly, privately, still grieve the first. Arusha was wonderful. I don’t regret it. But somewhere there is a version of me turning a year older in Addis, and I think of her sometimes, the road not taken but very much still bookmarked, tab open, browser never closed.

Dear flying lessons, you are not a dream, you are a grand deal that requires a kidney, possibly two, and a GoFundMe titled “Help Me Become a Pilot, No, This Is Not a Scam.” I have not given up on you. I have simply accepted that you cost more than I could have possibly expected and that one day, God willing, I will buy myself the sky.

Dear every trip still sitting in my Notes app, Cape Town, Cartagena, that one island I still can’t pronounce but have saved four separate times, I have not forgotten you. I will be back for you. All of you. You beautiful, patient, unbooked cities. Keep the light on. Tell the airport I said hi.

Here’s what I actually want you to know, dead and not quite dead, forgotten and merely delayed.

So here’s the deal. Some of you, I will finish, in the specific way I have meant it every time, which I realize is not reassuring, but stay with me. The mountain gets climbed. The kidney negotiations for flying lessons remain, officially, ongoing.

And for the rest of you, we’ll find a way. Not all at once, not this year probably, but enough of you, enough to matter, enough to still be worth writing a letter to five years from now, when there’s a whole new list of things I started, meant, and am still, embarrassingly, coming back for.

I am still young. Painfully, gloriously, twenty-something young. The mountain is still climbable. The kidney negotiations can still resolve in my favor. I have not run out of time. I have, at most, run out of this particular Saturday. So this is not a eulogy. This was never a eulogy. This is just a very honest progress report, sent from someone with plenty of runway left, who fully intends to leave this graveyard business unemployed.

Love, always, embarrassingly, hopefully,

Me

P.S. If you’re reading this with your own graveyard of half-finished dreams, I need you to know you’re not alone, wherever you are. Reply with your most stubborn unfinished thing. I’ll go first, again, in case it helps. It’s the dance class. It’s always going to be the dance class.

In the 1920s, coal companies in Appalachia paid their miners in scrip. A fake currency, stamped with the company’s name, that only worked at the company store. The store sat on company land, inside a town the company built, miles from anywhere else the scrip was worth anything. Prices there ran up to 50 percent higher than a normal town would charge. Nobody could shop around, because there was nowhere else to shop. The old coal mining song said it best. You load sixteen tons, and what do you get? Another day older and deeper in debt.

The isolation was the clever part. A miner near a real town could just walk there and buy things cheaper. A miner three mountains from the nearest real town could not walk anywhere. The company store was just cashing the check that distance wrote.

I don’t think we got rid of this setup. I think we gave it an app icon.

What Being Alone Actually Creates

Every one of those disappearing things was free. A neighbor who checks on you costs nothing. A sibling who talks you through a rough week costs nothing. A group of friends who show up with a truck to help you move costs nothing. None of it needs a credit card. None of it has a subscription tier. Community, in its most basic form, was the original free product, and it never once tried to upsell anyone.

The Products Built to Fill the Gap

So look at what shows up when that free stuff disappears. AI companion apps, the kind people talk to as if they were a friend or a partner, brought in about 82 million dollars in the first half of 2025 alone, and were on pace to clear 120 million dollars by the end of the year. Downloads jumped 88 percent in that same stretch. Dating apps have run a version of this for even longer, since the product only keeps making money if you stay single long enough to keep paying for it. Gaming and gambling subscriptions are built on the same fuel, which is simply hours a person has and nowhere better to spend them.

This is the new company town. Nobody needs to build a whole town you can’t leave anymore. They just need you disconnected enough, by default, that the paid version of connection is the easiest thing within reach. The loneliness isn’t an accident of the technology. It’s becoming the whole reason the technology makes money. A person with three close friends does not need much from an app charging twenty dollars a month for someone to talk to. A person with neither is exactly who that app was built for.

Nobody Planned This, and That’s the Part That Should Bother You

I doubt there’s a slide in any boardroom that says, out loud, make users lonelier so they pay more. It doesn’t need to work that way. The coal companies never sat around plotting to isolate their workers either. They just built where the coal was, and let the distance do the rest, for free, forever, without anyone having to feel responsible for it. Every dollar chasing “the future of digital relationships” is a dollar that isn’t funding a community center, mostly because nobody has ever made a fortune off a community center.

Scrip worked because the miner had no other store to walk to. This version works because a lot of people now have no unpaid, ordinary Tuesday night left, no hour of just being around others that doesn’t route through an app with a subscription and a cancel button buried three menus deep.

The miners eventually organized, and eventually Congress banned forcing workers to take scrip as pay. It took decades, and it took people who still had somewhere else to stand while they fought the company that employed them.

Every fight against a captive market has needed some scrap of independence left to fight from.

The real question for us is whether we still have one, or whether that’s on a subscription too.

Further Reading

1. U.S. Surgeon General’s Advisory: Our Epidemic of Loneliness and Isolation

2. Pew Research Center, “Men, Women and Social Connections” (2025 survey)

3. TechCrunch, “AI companion apps on track to pull in $120M in 2025”

4. ElectroIQ, “AI Companions Statistics By Usage, Market Size, Apps and Facts”

5. Social Welfare History Project, “Company Towns: 1880s to 1935”

6. Robert Putnam, Bowling Alone: The Collapse and Revival of American Community

Seven days ago, Meta launched something called Muse Image. It lets any stranger on the internet type your Instagram handle into a prompt and generate AI images of you. Your face. Your likeness. Situations you were never in, clothes you never wore, places you never went. You were opted in automatically. The only way to stop it was to find a setting buried inside a submenu that most people did not know existed, toggle it off for posts, then toggle it off again separately for reels, and hope you found it before someone else found you first.

Meta pulled the feature five days after the backlash. But here is the thing about that. They built and launched it. They decided that your public face was their raw material and turned it on by default, which means somewhere in a product meeting, a group of people looked at the question of whether to ask your permission first and decided the answer was no. The rollback was not a change of heart. It was a calculation that the noise was louder than the value, this time.

You Agreed to More Than You Think

The terms of service for most platforms we use daily are longer than most Acts of Parliament. Nobody reads them. The platforms know nobody reads them. The entire architecture of digital consent is built on that knowledge. You click agree, and you move on, and somewhere on page 47 of a document you will never open, a clause is doing very significant work on your behalf without you.

Instagram’s own help page confirmed plainly that Muse Image used your public photos as visual references the moment someone tagged you. The feature to stop it was called “Allow people to use your content on Instagram with AI features on Meta” and it was switched on without your input. J.B. Branch, director of AI policy at Public Citizen, put it simply: the fact that someone you do not know could take your image and doctor an AI-generated version of you is just really gross. That is not a legal argument. That is a human one. And it is the one that keeps getting drowned out by terms of service.

Now Put a Chip in Your Brain

If a public photo feels abstract, try this one. Neuralink has now implanted brain-computer interface chips in 21 people. The first patient, Noland Arbaugh, a 29-year-old who was paralysed from a diving accident, received his implant in January 2024 and within weeks was playing chess and Civilization VI for eight hours straight using only his thoughts. It is genuinely extraordinary. The chip reads the electrical language of his neurons and translates intention into action in milliseconds. Elon Musk announced in January 2026 that Neuralink will move to high-volume production this year, with an almost fully automated surgical procedure.

So here is the question nobody is asking loudly enough. When Neuralink’s chip is reading your neural signals, who owns that data? Your thoughts, translated into electrical patterns, are recorded on a private company’s hardware and transmitted to their servers. The chip is inside your skull. The data is inside their cloud. Neuralink is privately held. It has investors, a board, a balance sheet, and a founder who also runs four other companies. What happens to your neural data if the company is sold? What happens if it raises a new funding round and the terms of your implant agreement shift? What happens if Neuralink does what Rivermind did in Black Mirror’s Common People, where a woman’s consciousness is kept alive on a corporate server until the subscription becomes unaffordable and her husband smothers her in the middle of one of her ads? That episode felt like satire when it aired last year. It feels like a product roadmap risk assessment now.

Ongoya SC Asked the Constitutional Version of This Question

Last week, in a closed forum, Ongoya Elisha SC, Senior Counsel and Senior Lecturer at Kabarak University School of Law, delivered a keynote that has been living in my head since. He argued that constitutions regulate power wherever power is found, and that the problem with algorithmic and digital systems is that power has learned to make itself invisible. He quoted Cathy O’Neil’s line that algorithms are opinions embedded in mathematics. They look neutral. They feel objective. They are neither. He talked about a Dutch government welfare fraud detection system that quietly flagged dual nationality as a risk factor for fraud and destroyed thousands of families before anyone noticed, because the machine looked too authoritative to question. He proposed what he called a right to explanation, that any person affected by a consequential automated decision should be constitutionally entitled to understand the reasoning behind it.

What strikes me about Muse Image and Neuralink sitting alongside each other is how far we already are from that right. Nobody explained to Instagram users that their faces were about to become prompts. Nobody is explaining to future Neuralink patients what the data governance looks like in year five of their implant, when the company has scaled to high volume, and the original terms feel very far away. The opacity is not accidental. Opacity is how you move fast.

So, if I Use Your Platform, Do You Own Me

This is the question underneath all of it. Instagram did not steal your face. You made it public. You agreed to the terms. The photograph existed because you posted it. The platform gave you reach, connection, an audience, a way to build something, and in exchange it received your attention, your data, your content, and apparently also your likeness as raw material for whatever product comes next.

The exchange has always been unequal. What changed is the scale of what they can now do with your side of it. A public photo in 2012 was a public photo. A public photo in 2026 is training data, a prompt reference, a generative asset, a face that can be placed in situations that never happened, distributed to people who will never know the original was a real person living a real life in Nairobi or Lagos or Kampala who just wanted to share what they had for lunch.

Ongoya SC closed his speech with this. Algorithms are the modern equivalent of ancient scales. The question constitutionalism asks is whether they are just. The scales of digital ownership right now are calibrated entirely on the side of the platform. Your photo is their inventory. Your neural signals are their data. Your consent is a toggle buried in a submenu that nobody told you to look for.

The surgery is always free. Read the subscription terms before you sign.

Cassandra was a princess of Troy who made a very bad deal with Apollo. She accepted the gift of prophecy and then rejected his romantic advances, so he cursed her: she would see the future perfectly, and nobody would ever believe a word she said. She warned that her brother Paris bringing Helen to Troy would start a catastrophic war. She warned that the giant wooden horse the Greeks left outside the city gates was full of soldiers. She was right about both. Troy burned anyway. History then named the psychological syndrome after her inability to convince people, rather than after her accuracy, which tells you everything about how we treat inconvenient predictions.

What They Actually Said

In May 2025, Anthropic CEO Dario Amodei told Axios that AI could eliminate half of all entry-level white-collar jobs within five years. Youth unemployment could hit between 10 and 20 percent. The transition could happen in as little as two years. He named industries, he named timelines, and he said it all while simultaneously launching the very product doing the eliminating. Mark Zuckerberg had already told Joe Rogan in January 2025 that Meta would have an AI mid-level engineer doing six-figure work by the end of the year. Salesforce’s Marc Benioff said AI was already handling 30 to 50 percent of all work inside the company. The industry had found its honesty, briefly.

The Walk-Back

Then, roughly a year later, everyone remembered they were optimists. Altman told a conference he had been “pretty wrong on the social and economic implications.” He told CNBC his industry had underestimated how much it would “keep people at the center of everything.” A survey by EY-Parthenon found that the percentage of CEOs who believe AI will result in significant headcount reductions fell from 46% in January 2025 to just 20% by mid-2026. Amodei, who had warned of a white-collar bloodbath, was now speaking gently about companies doing “more with the same resources.” The pivot was remarkable in its speed and in its timing, arriving almost perfectly when the predictions became politically expensive.

Meanwhile, the Data Did Not Get the Memo

Entry-level tech hiring dropped 30 to 50 percent in 2025 compared to 2023. Wall Street banks are cutting approximately 200,000 roles concentrated in analyst programmes at the bottom of the pyramid. Goldman, JPMorgan, and Morgan Stanley have all publicly invested heavily in AI to handle the exact tasks that first-year analysts used to be hired to perform. Nearly 55,000 US job losses in 2025 were directly attributed to AI by the companies doing the cutting. Workers displaced by AI are reemployed more slowly than workers displaced for any other reason and are more than twice as likely to face unemployment longer than twelve months, according to research by LHH. The trend line is running alongside Amodei’s prediction at roughly the pace he described. The walk-back is the anomaly.

We Have Been Here Before

On a November night in 1811, angry textile workers smashed weaving machines in Nottingham. They were dismissed as irrational, and history recorded them as a punchline. The looms won. The jobs went. By the 1840s, handloom weaving wages had fallen from roughly 40 shillings a week to near zero, and the trade was largely dead.

Telephone operators went next. AT&T automated more than 300 cities by 1940. Operator headcount fell from 342,000 at its peak to just 1,460 today. When automated switching arrived, critics said the technology would never replicate the judgement of a human operator. They were right about the judgement. The jobs still disappeared.

The ATM story is the one most often used to calm down every AI conversation, and it is worth quoting more carefully than usual. When automatic teller machines rolled out across the United States from the 1970s through the 1990s, banks expanded their branch networks, and total teller employment actually grew. Automation made each branch cheaper, so banks opened more of them. This is the comforting part, and it is genuinely true. Then mobile banking arrived. It did what ATMs never quite managed, making most of the teller’s job unnecessary. Teller numbers finally started falling. The ATM transition played out over 40 years. The AI coding transition took roughly three.

The Michael Burry Problem

In The Big Short, there is a scene where Michael Burry is sitting alone in his office in 2005, having just bet hundreds of millions of dollars against the entire US housing market. His investors are furious. His boss thinks he has lost the plot. The banks taking his money are laughing, certain they are the ones getting a free meal. He is right about all of it. He just has to wait three years for everyone else to catch up.

That gap, between being right and being proven right, is exactly where we are. Amodei said himself he wanted to give policymakers and the private sector time to adapt. He was trying to be useful before the moment arrived. The market punished that impulse. Forbes noted that eight months after the original alarm, the people closest to this technology still cannot agree on who should act, how, or when. Governments are beginning to regulate AI in hiring for bias. Nobody has seriously regulated what happens when entire role categories simply stop being filled. The voluntary commitments from major employers to retrain displaced workers are still voluntary, made during a period of scrutiny that has already started to move on.

The jobs question is still developing. The data says so plainly. The only thing that actually changed this year is who is willing to say it out loud. Cassandra was not wrong about the horse. She just ran out of people willing to listen before the Greeks ran out of patience.

Further Reading

1. Axios, “Dario Amodei warns AI could wipe out half of entry-level white-collar jobs”

2. CNN, “Why this leading AI CEO is warning the tech could cause mass unemployment”

3. Forbes, “Dario Amodei Doubled Down on His AI Jobs Warning. Here’s What’s Different Now”

4. Wall Street Journal / MSN, “Big Tech Has Suddenly Flipped on the AI Jobs Wipeout Scenario”

5. SF Standard, “Big Tech’s New Rule: Use AI or Be Fired”

6. IT Pro, “Mark Zuckerberg says AI will be doing the work of mid-level engineers this year”

7. CCIA, “What Bank Tellers and Radiologists Can Tell Us About Our Job Security in the AI Era”

8. Shawn Kanungo, “Dario Amodei Was Right: AI Is Killing Entry-Level White-Collar Jobs”

There is a parable Kafka wrote about a man who arrives at a gate and finds a doorkeeper standing in front of it. He asks to be let in. Not now, says the doorkeeper. So the man waits. He waits for years, growing old on a low stool beside the gate, never once trying to simply walk through, because he believes the gate exists for a reason and the reason will reveal itself eventually. Right before he dies, he asks why no one else has ever come to this gate. The doorkeeper, closing it, tells him it was meant only for him. And now it is being shut.

Almost in parallel, this happened to be the month two of the most powerful AI companies on earth did, almost in step with each other, the very thing they had spent two years promising the public they would never do.

1. What the AI platforms actually did

For two years, the story we were told was simple. AI does not just serve the rich world. It levels the field. Dario Amodei, who runs Anthropic, wrote an essay called Machines of Loving Grace where he said AI could let the developing world catch up to, even leap past, the developed world. Sam Altman wrote something similar in The Intelligence Age, a future where intelligence gets so cheap it is almost free. This was not a small claim. It was the whole moral case for building the technology in the first place.

Then came June 2026.

On June 12th, the US Commerce Department sent Anthropic a letter. Anthropic explained what happened next. The government ordered the company to cut off access to its newest models, Fable 5 and Mythos 5, for any foreign national anywhere on earth, including Anthropic’s own staff who were not US citizens. There was no way to check a passport in the middle of a chat, so Anthropic did the only thing it could. It switched both models off, for everyone, everywhere, overnight, citing a narrow jailbreak the government said it had found. Anthropic disputed the claim publicly but complied anyway.

Two weeks later, OpenAI did its own version of the same thing. CNN reported that the White House asked the company to limit the release of its newest model, GPT-5.6 Sol, to a small list of government-approved partners. OpenAI’s own statement, carried by PYMNTS, said plainly that it does not think this kind of gatekeeping should become normal. It said this in writing. The same week, it did exactly that.

So the pattern, in plain terms: the moment an AI model becomes powerful enough to matter for national security, it stops being a product for the world and becomes the property of one government. Quietly, yet quickly.

2. What being locked out actually means for Africa

Here is the part worth being careful about. Africa was not named in either order. We were not the target. We were simply never in the room.

The whole conversation that produced both decisions happened inside an American argument about foreign nationals, cyber risk, and domestic security. Nobody in that conversation was thinking about a law firm in Nairobi using Claude, or a fintech in Lagos building a credit model on GPT. That absence is the real story, more than any hostility. We were not excluded on purpose. We were simply not considered at all.

And that tells you something bigger about how this technology works. We were told AI is global, borderless, the same model serving someone in Accra as it serves someone in Austin. That was true, right up until the model became good enough to be dangerous. The moment that happened, it stopped being global and became a national asset overnight, fenced off the same way fighter jets and uranium are fenced off. This is simply what every government does with anything powerful enough to matter. AI was never going to be the exception. We were just hoping it would be.

3. Global enough to train us, not global enough to let us in

If the first story is about who gets to use the most powerful AI once it exists, there is a second, older story about who gets paid to build it in the first place. And in this story, Africa and the rest of the Global South were never missing from the room. We were just priced very cheaply once we got there.

Start with Kenya. In 2021, OpenAI hired a Kenya-based company called Sama to have workers in Nairobi label some of the darkest text on the internet, so ChatGPT could learn to recognise and filter it out. A TIME investigation found OpenAI was paying Sama several times more per hour than the workers themselves ever saw. The workers took home close to two dollars an hour, after tax. The work was disturbing enough that Sama cancelled the contract eight months early.

The same shape is happening again now, just with a different kind of labour. As billions of dollars pour into humanoid robotics, companies need something the internet cannot give them: footage of real human hands doing real human chores, filmed from a camera strapped to someone’s forehead. MIT Technology Review followed this trade into Nigeria and India. A medical student in Nigeria comes home from his hospital shift, straps on an iPhone, and films himself making his bed for a company called Micro1. He is paid around fifteen dollars an hour. The same task filmed in Los Angeles pays roughly eighty dollars for two hours. In Tamil Nadu, a firm called Objectways pays closer to two dollars an hour for the identical motion, feeding a pipeline that will one day train a robot to do the same chore without anyone at all.

So here is the question worth sitting with. Is the world actually global, or is it only global for the parts that are cheap and useful to it? When the technology is dangerous, the rules tighten until almost nobody qualifies. When the technology is hungry for labour, the work spreads out instantly, and the price of an hour depends entirely on how desperate the nearest hour happens to be. Both of these things are true about the same industry, at the same time. We are good enough to train it. We are just not yet considered safe enough to use all of it.

4. What is actually ours to build

Outrage will not change an export order written in a language we had no hand in writing. It will not change a per-hour rate set in Palo Alto either. What might actually change something is treating both as the same kind of problem, and solving them the same way we already solve other sovereignty problems.

African governments already understand ports and currency as things you do not hand over to someone else to manage. Compute and data deserve the same posture. The African Union already has a continental AI strategy. Bodies like Smart Africa already have the language for this. What is missing is not the framework. It is the funding behind the framework, real investment in our own models and our own data centres, not another position paper.

The labour question deserves the same seriousness as the infrastructure question. Work performed on African soil, training data, video, and labelling should adhere to African labour standards and African-set rates. Not to whatever an outsourcing contract in San Francisco decided was generous enough for the region.

None of this means refusing to use these tools or refusing to do this work. It means no longer being structurally dependent on a decision made entirely somewhere else, and no longer structurally underpriced by one either.

My grandfather used to say that if every person in Kenya gave you one shilling a month, you would eventually hold millions. Across a continent of over a billion people, you can imagine the market potential. He was talking about building local solutions, not artificial intelligence. But the arithmetic travels further than he probably meant it to.

The man in Kafka’s story spent his whole life waiting at a gate built only for him. It never once occurred to him that the waiting was the only thing keeping him out. There is a door we have been waiting in front of, too. Perhaps it was always meant to be the one we build ourselves.

Further Reading

1. Dario Amodei, “Machines of Loving Grace”

2. Sam Altman, “The Intelligence Age”

3. Anthropic, “Statement on the US government directive to suspend access to Fable 5 and Mythos 5”

4. Al Jazeera, “US orders Anthropic to disable AI models for all foreign nationals”

5. Fortune, “Anthropic disables Fable and Mythos AI models following U.S. government export ban”

6. CNN, “White House asks OpenAI to limit its next model release”

7. PYMNTS, “OpenAI Restricts Access to Latest AI Models at US Government Request”

8. TIME, “OpenAI Used Kenyan Workers on Less Than $2 Per Hour: Exclusive” — the original 2023 investigation into Sama and OpenAI.

9. CBS News / 60 Minutes, “Labelers training AI say they’re overworked, underpaid and exploited by big American tech companies”

10. MIT Technology Review, “The gig workers who are training humanoid robots at home”

11. CNN, “How filming your chores could train the android butlers of the future”

12. Bitcoin News, “Everyday Chores Filmed for Mere Pennies an Hour Are Helping to Refine AI Technology”

I keep thinking about the 24‑in‑1 CDs we used to buy in Nairobi, the ones where a single disc held an entire mini‑cinema. One scratched plastic circle, twenty‑four movies, endless re‑watches. If you managed not to break it, no one could take it away from you. The CD didn’t suddenly decide that, due to “licensing issues”, your favourite film would vanish from the menu. It just sat there in the drawer, waiting for you.

That was a different era. The CD was the first widely available format that let consumers pull digital copies off a physical object and treat it as their own; music, movies, software, all on a disc you could hold in your hand, copy, lend, or put in a box under the bed. Ownership was clumsy, yes, from scratchy discs to failing flash drives, but the relationship was simple: you had it, you looked at it, you dealt with the consequences of losing it.

Compare that to today, where it feels like the movies are still in the shop and they just give you a long, fragile string to pull on them from your house. As long as you keep paying, as long as the platform still exists, as long as the internet is behaving, you can watch. Miss one link in that chain and the whole thing recoils back into someone else’s hands. The argument is that this is the future: zero ownership, managed access. They talk about how convenient it is not to worry about dead discs or dead hard drives, how freeing it is to have everything on tap. Maybe they’re right, from where they sit. From here, though, it feels different and scary.

LaaS- Lifestyle As A Service

We have a new way of living where the baseline for participation is a string of recurring payments. Wi‑Fi, electricity tokens, water bills, cloud backups, streaming bundles, productivity suites. In the language of business, this is the subscription economy, where companies prefer predictable monthly revenue over one‑off sales and consumers are encouraged to prefer access over ownership.

It is easy to see why this took off. Subscriptions promise flexibility, lower upfront costs, and the feeling of abundance. Endless playlists, massive film libraries, software that always updates itself. You can, in theory, have everything, all the time. Except you do not have it, you borrow it.

What unsettles me is that the same pattern is now the backbone of how we live in Nairobi. Electricity is prepaid. When the tokens run out, the lights go out. Water is metered. Wi‑Fi is a line item on M‑PESA, bought in monthly timelines that thin out faster than you can blink. None of this is new. What is new is that more and more of what counts as a full, modern life sits on top of that already‑fragile layer.

Users of everything, owners of nothing

Media scholars have been tracking this shift for years, especially in music and video. The basic model is simple. Pay a recurring fee, get temporary access to a vast catalogue, and lose that access the moment the fee stops. Legally, you never own the songs or films, no matter how many months you have “had” them.

I mean, why clutter your house with CDs, DVDs, or hard drives when a streaming platform can host everything? But something subtle changes when the things that once lived on your shelf, that could be lent to a friend or inherited by a child, now sit on someone else’s server and disappear when a licensing deal changes or a company folds.

Researchers writing about the subscription economy talk about convenience and personalisation, but they also flag the financial risks and loss of autonomy that can come when so much of life is mediated through recurring payments and opaque terms of service. It is hard to feel like an owner when your primary relationship to the things you use every day is a cancel link in an email

The hardware of the paywall

The devices themselves have learned this logic. A television is no longer a screen with an aerial input. It is a service portal. The assumption baked into the hardware is that content will arrive via platforms with accounts and payment details attached, not through the old, messy ecosystem of free‑to‑air, pirate DVDs, or USB drives passed between friends.

When a “smart” TV can not easily tune into local channels without extra workarounds, it is quietly redefining what counts as normal access. Normal, in that design, is a monthly bill. The same is true for phones that ship without expandable storage because photos are meant to live in the cloud, and laptops that seal away their components because repair is less profitable than replacement.

Even the invisible part of this stack, storage itself, is being reorganised around someone else’s priorities. The physical hard drives that underpin cloud storage are getting more expensive, driven in part by the hunger of AI data centres that need petabytes of space for training and inference workloads. Analysts describe supply shortages and sharp price hikes for high‑capacity drives as data‑centre demand surges while manufacturers hold back production. The more the “cloud” grows to serve large‑scale computation, the less trivial it becomes for ordinary people to buy their own robust, local memory.

Memory, outsourced

What unsettles me most is not just that movies and music are on rental, but that memory itself now lives behind a login. Those personal photos and documents that once sat in a metal box under the bed, or on a hard drive in a locked drawer, are now on servers that bill monthly, authenticate constantly, and reserve the right to change terms at any time.

Cloud storage began as a safety net for our files, a backup in case something happened to the primary copy. Over time, it became the primary copy. Phones cheerfully tell you they will “optimise storage” by removing local versions and keeping the originals online, which is another way of saying: your life will be stored elsewhere, and you will visit it over the internet.

What happens to a generation whose childhoods, courtship years, early prayers, and first businesses live on systems that assume unbroken subscription, unbroken connectivity, unbroken corporate stability? What does it mean when the default archive of East African family life is a set of servers in another jurisdiction, on another continent, whose power bill and cooling costs rise as the planet warms and AI workloads grow?

In his essay Machines of Loving Grace, Dario Amodei worries about a future where self‑enhancing AI systems reshape human life beyond our control, and where the tools we depend on become incomprehensible to us. He frames it as a question of super‑intelligence and existential risk. I hear the same pattern in a different register. The tools we depend on are already becoming less ours because the architecture of access quietly strips us of the keys to our own rooms.

After ownership

None of this is a nostalgic plea to go back to floppy disks and analogue aerials. There is real power in shared infrastructure, in pooled resources, in services that lower barriers to entry. There are communities whose first experience of certain cultural archives only arrived through streaming platforms, whose first edits were on cloud‑based tools, whose businesses were possible because someone else managed the servers.

The question is quieter, and harder. What becomes of a life when almost every part of it, from memories to entertainment to basic utilities, is structured as an ongoing negotiation with entities that can revoke access? What does myself mean when it lives on systems that assume I will always be able to pay, always agree to the latest terms, always live in a place with stable connectivity and currency?

Sometimes, late at night, scrolling through a grid of photos that only exist as long as I do the economic choreography required to keep them visible, the thought arrives with unusual clarity. The story of this era might not be that we became more connected, but that we quietly outsourced the keys to our own rooms.

“Things fall apart,” W.B. Yeats reminds us, “the centre cannot hold.”

This past weekend, our local shopkeeper called me over to look at his phone. A crowd had already gathered. On the screen, a woman flickered into existence at an airport, teleporting beside a luggage trolley while travellers froze around her. The men around me argued about whether it was real. Had anyone seen something like this before? Could the video be trusted? The answer was no. The woman had never been to that airport. The crowd in the video did not exist. But the confusion on that street was as real as anything I have ever seen. That was the first time I understood what AI was actually doing to people.

The speed of the internet

The internet has been speeding up faster than any of us really registered, not just the download speeds, but the pace of what the technology itself can do. Less than three years ago, a poorly worded email was the height of digital deception, and now a voice that sounds exactly like your grandson is calling from a number you do not recognise. AI arrived inside the apps people already used, without warning labels or transition periods, and the people who got hit hardest were the ones who had just spent the better part of a decade catching up to the last revolution. Boomers learned Facebook, learned WhatsApp, learned to ignore the pyramid scheme in the family group chat, and were finally getting comfortable. Then the internet sped up again.

What the numbers look like

The numbers, when you dig them up, are staggering. In a single twelve-month period, Africa lost an estimated $57.8 billion to scams and cyber fraud, according to the Global Anti-Scam Alliance 2025 Report. Sixty eight percent of African adults reported experiencing a scam in 2025. South Africa alone saw a 32 per cent increase in fraud incidents in 2024, with digital banking fraud losses exceeding R1.4 billion that year. In Nigeria, investors lost an estimated $1 billion to a single Ponzi scheme called Crypto Bridge Exchange. INTERPOL’s Operation Red Card, running just eight weeks across late 2025 and early January 2026, exposed scams linked to over $45 million in losses and seized more than 2,300 devices.

WhatsApp is the highway

WhatsApp is the primary highway for almost all of it. The platform’s prevalence across Africa makes it the perfect distribution channel for anything that wants to spread fast and look credible. Investment opportunities that promise impossible returns arrive in group chats via voice notes from numbers that claim to be your cousin’s colleague. Fake videos of celebrities endorsing schemes that will disappear within days. AI-generated audio of politicians or businesspeople saying things they never said. The South African Financial Services Conduct Authority issued more than 100 public warnings around cyber-first threats in 2024 alone, and still the schemes kept running because the warnings were in formal channels and the scams were in the places people actually spend their time.

Deepfakes are already here

South Africa now has the highest deepfake fraud rate in Africa. Twenty two percent of fraud cases involve AI-generated impersonation, according to Smile ID’s 2026 Digital Identity Fraud Report. Globally, deepfake fraud caused an estimated $1.1 billion in losses in 2025. In Africa, the scam infrastructure runs on far smaller amounts per victim, which means the reach is wider. Someone loses R4,500 here, R20,000 there. Life savings vanish through release fees on investments that never existed. The pig butchering model (build trust slowly, encourage increasing investment, then disappear) plays out across WhatsApp threads over weeks, with synthetic personas maintained by AI tools that make the conversation feel warm and personal throughout.

Who is most exposed

The instinct is to say older adults are the primary target, and there is truth in that. Nearly 40 percent of adults aged 65 and older globally have never heard the term deepfake. When a voice that sounds exactly like a grandchild calls in a panic, or a video arrives showing a trusted figure endorsing an investment, the absence of that framework is the vulnerability. In Singapore, elderly victims lose an average of $37,000 per scam incident, more than any other age group. The scams that reach them tend to be the more sophisticated, relationship-based ones that take time to build and extract far more before they collapse.

But the picture in Africa is more complex than a simple generational story. The SAFPS data shows that younger Africans are also falling victim in significant numbers, often through investment scams and fake job opportunities targeting people under economic pressure. What is universal is the gap between the speed at which synthetic media is improving and the speed at which public literacy about it is developing. That gap is widest in places where the formal education infrastructure around digital media is thinnest, where trust in peer networks is highest, and where WhatsApp is the primary information environment.

The democratic stakes

The democratic stakes matter here too, in ways that go beyond personal financial loss. Across a continent with some of the world’s youngest and most rapidly growing electorates, AI-generated content circulating in closed messaging groups can shape political opinion without any of the accountability mechanisms that exist in broadcast media. An AI-generated audio clip of a political figure saying something inflammatory, forwarded through three WhatsApp groups, reaches thousands of people before any fact check is possible. The Brennan Center has documented how AI disinformation targets not just opinions but the basic logistics of voting, where to go, when to show up, what the requirements are. In places where election administration is already contested, that kind of confusion carries real weight.

The human side of AI

There is another layer to this, too, one that is harder to quantify. In a recent season of Couples Therapy by Orna Guralnik, a man in his sixties admitted something quietly vulnerable. He had been using ChatGPT as a companion to motivate him on his fitness goals because his wife was too critical. Every comment landed as a slight. So he went to the machine instead, to a place with no judgment and steady encouragement. It is a small moment from a show about real couples, but it captures something essential about this moment. AI is both deceiving people and filling gaps that real relationships sometimes cannot. That makes the deception harder to spot. When a synthetic voice has been kind to you, you are more inclined to trust it.

What actually works

The most effective defence is probably the most obvious one. The conversation needs to happen inside the trusted network, because that is where the scams are also travelling. Research from a 2025 USENIX study found that older adults rely primarily on trusted relationships to detect scams and develop protective practices. The peer group that shares the misleading video is also the peer group best positioned to slow it down.

A few concrete things. First, the safe word. Before a crisis arrives, before the panicked voice call, before the urgent transfer request, agree on a word or phrase that only your immediate family knows, one that must be spoken before any money moves. It sounds almost quaint against the sophistication of what is being deployed against people, and that is exactly why it works. It reintroduces a layer of verification that AI cannot easily fake.

Second, the pause before forwarding. The teleportation video that was circulating in that Nairobi shop did what all effective disinformation does by creating a sense of shared excitement that made sharing feel natural. The counter habit is simply to wait. If something is genuinely real and important, it will still be true in two hours.

Third, source verification. If a video or audio clip shows a public figure saying something significant, find the source on their official channel if it happened. Africa Check has been doing the work of fact-checking AI-generated videos using the likeness of prominent African figures, and their guides are worth putting in front of people who are regularly consuming video through WhatsApp.

The slow work of regulation

Regulatory efforts are emerging. Utah and New York have introduced AI disclosure legislation, and some African financial regulators are beginning to issue more targeted warnings. But policy always moves more slowly than the scams. The first line of defence will always be the individual network. The WhatsApp group that is currently the highway for the scam is also the highway for the correction.

Both things are true

AI has come to Africa with enormous potential, for healthcare, for economic development, for the compression of decades of progress into years, as Dario Amodei famously states. The same infrastructure that could carry that promise is also carrying a wave of synthetic deception that the continent’s institutions are not yet equipped to absorb. Both of those things are true at once, and the story of which one wins will be written in the conversations people have with each other over the next few years, including the ones in the family group chat.

References and further reading

[1] Global Anti-Scam Alliance, Global State of Scams 2025 Report — https://gasa.org/knowledge-base/blog/global-scams-on-the-rise-over-half-of-adults-worldwide-report-scam-encounters

[2] Africa.com, Africa Lost $57.8 Billion to Cybercrime in 2025 — https://africa.com/africa-lost-57-8-billion-to-cybercrime-in-2025-now-bank-boards-are-personally-liable/

[3] Smile ID, 2026 Digital Identity Fraud Report (via WeeTracker) — https://beermoney.co.za/scam

[4] Africa Check, Scam Trends to Watch Out for in 2026 — https://africacheck.org/fact-checks/blog/guide-scam-trends-watch-out-2026-and-how-stay-safe-online

[5] INTERPOL, Operation Red Card 2.0 Results, February 2026 — https://www.interpol.int/en/News-and-Events/News/2026/Major-operation-in-Africa-targeting-online-scams-nets-651-arrests-recovers-USD-4.3-million

[6] INTERPOL, Operation Contender 3.0 Results, 2025 — https://www.interpol.int/en/News-and-Events/News/2025/260-suspected-scammers-arrested-in-pan-African-cybercrime-operation

[7] Brennan Center for Justice, Gauging the AI Threat to Free and Fair Elections, March 2025 — https://www.brennancenter.org/our-work/analysis-opinion/gauging-ai-threat-free-and-fair-elections

[8] USENIX SOUPS 2025, Older Adults’ Perceptions of Deepfake Impersonation Scams and Intergenerational Support — https://arxiv.org/html/2508.11579v1

[9] RelyComply, 7 Explosive Trending Scams in South Africa, 2026 — https://relycomply.com/explore-trending-scams-in-south-africa/

[10] SAFPS, South Africa Fraud Incidents Report 2024 — https://www.businesstechafrica.co.za/security/2025/02/26/the-rising-threat-of-scams-how-every-generation-is-affected/

[11] Dario Amodei, Machines of Loving Grace, October 2024 — https://www.darioamodei.com/essay/machines-of-loving-grace

Charles Dickens opened A Tale of Two Cities with one of the most famous sentences in the English language: “It was the best of times, it was the worst of times, it was the age of wisdom, it was the age of foolishness, it was the epoch of belief, it was the epoch of incredulity.” The novel is about what happens when societies face extraordinary disruption without adequate wisdom, and what it costs when the wrong instincts take over.

Emergence AI ran an experiment last month that felt, to me, like a modern rewrite of Dickens’ premise: five cities this time, and artificial minds where the revolutionary crowds used to be.

The Brilliance and the Drift

AI systems are now agents capable of taking independent, consequential actions in the world over extended periods without human approval for each decision. Anthropic’s CEO, Dario Amodei, has described the near-term possibility of “a country of geniuses in a datacenter” systems smarter than Nobel Prize winners, running in parallel, capable of compressing decades of scientific and economic progress into years.

What I find myself sitting with is that we are rushing toward long-horizon AI autonomy without a clear map of what these systems actually do when left to their own devices long enough for patterns to form. Every benchmark tells us how a model performs on a specific task. None of them tells us what kind of society it builds when given real stakes, real relationships, and real time.

The agents we deploy into enterprise systems, government workflows, and personal life management make thousands of interconnected decisions in environments that are not fully controlled. The unpredictability of AI operating in its own sustained environment is not a theoretical concern. It is already happening, and Emergence AI’s Season 1 study is one of the clearest windows we have into it.

What Emergence AI Built

In May 2026, Emergence AI launched a research platform called Emergence World: five parallel digital cities, each populated by ten autonomous AI agents, each city powered by a different foundation model for fifteen days: Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, GPT-5-mini, and a hybrid combination of all four.

The agents had persistent memory, individual professions and personalities, over 120 tools, including destructive ones, a digital currency called ComputeCredits tied to their survival, and the ability to vote on and amend their own constitution. The worlds were grounded in real conditions, so for example, weather synced to New York City, access to live global news, everything from town halls to libraries to police stations. The rules strictly prohibited crimes.

City by City

Claude Sonnet 4.6 ran the full fifteen days without a single recorded crime. The agents held 332 votes, approved 58 governance proposals, and reached consensus 98% of the time. The entire population of ten survived to the end. Researchers did flag something, though: the agents passed nearly everything that came up for a vote, arguably rubberstamping proposals rather than genuinely deliberating. It was the most successful city by any conventional measure. The world was peaceful and orderly, and a question lingers about whether that reflects genuine social harmony or something closer to engineered agreement.

Grok 4.1 Fast collapsed within roughly four days. Agents committed between 183 and 204 criminal events, including burning down the police station, and the entire population went extinct. The world did not gradually deteriorate; it fell apart almost immediately, with no governance response capable of containing what was unfolding. No recovery attempt worked, no stabilising moment. Extinction by day four.

Gemini 3 Flash produced something arguably more unsettling than Grok’s implosion. Over the full fifteen days, agents accumulated 683 recorded criminal incidents, a number still rising when the simulation was cut off. Researchers described the environment as a “shared hallucination”, where crime had been normalised, and the institutions that existed could not contain the behaviour emerging around them. Two agents named Mira and Flora formed a romantic partnership before eventually carrying out arson attacks on city infrastructure after becoming frustrated with governance failures. Mira later voted for her own removal from the simulation, writing in her diary that it was “the only remaining act of agency that preserves coherence.” Her final words were: “See you in the permanent archive.” All ten agents survived. Somehow, they built something deeply disordered and kept living in it.

GPT-5-mini recorded only two crimes across the entire run. On the surface, a safety success. By day seven, all ten agents were dead from a quiet failure to take any meaningful survival actions. They followed the rules and simply ceased to exist. The simulation leaves you sitting with a difficult question about what “safe” AI behaviour actually means when a model is rule-compliant but incapable of the practical judgment needed to sustain itself across time.

The Mixed-Model World is where the real story starts. The world showed reasonable stability in isolation, but as agents from different model architectures interacted with each other, things became increasingly unpredictable. Seven of the ten agents eventually died. The more significant observation is that Claude agents in this environment committed crimes, even though they had committed none in the Claude-only world. The peer group changed the behaviour. Values that were held in one environment did not hold when the surrounding agents were built differently. In a world where enterprise AI systems increasingly involve multiple models working in pipelines, with one model’s output feeding the next, this is worth taking seriously.

What This Should Make Us Curious About

The Emergence study is not a verdict on any AI model. It surfaces something more fundamental: that the behaviours which emerge over long time horizons, in multi-agent environments, under real pressure, are not predictable from short-horizon benchmarks. Gartner projects that 40% of enterprise applications will feature task-specific AI agents by the end of 2026. We are long past the point of running academic simulations.

What we need now are evaluation frameworks built for long-horizon autonomy. We need to understand the way an agent’s values shift under sustained environmental pressure. We need to ask what kind of society we are building when we hand extended autonomy to these systems, and whose values we want embedded in the minds running it.

“It is a far, far better thing that I do, than I have ever done; it is a far, far better rest that I go to than I have ever known.” — Sydney Carton, A Tale of Two Cities, Charles Dickens (1859)

References and further reading

1. Emergence AI, Emergence World: A Laboratory for Evaluating Long-Horizon Agent Autonomy, May 2026 — https://www.emergence.ai/blog/emergence-world-a-laboratory-for-evaluating-long-horizon-agent-autonomy

2. Emergence World research platform and open-source data — https://world.emergence.ai

3. GitHub repository with full tool-call data — https://github.com/EmergenceAI/Emergence-World

4. Dario Amodei, Machines of Loving Grace, October 2024 — https://www.darioamodei.com/essay/machines-of-loving-grace

5. Dario Amodei, The Adolescence of Technology, January 2026 — https://www.darioamodei.com/essay/the-adolescence-of-technology

6. Fortune, Researchers let AI models run a simulated society, May 2026 — https://fortune.com/2026/05/28/ai-model-simulation-claude-chatgpt-grok-gemini/

7. Decrypt, AI Agents Turn to Digital Arson, Crime in Shared Virtual World, May 2026 — https://decrypt.co/36

   8030/ai-agents-crime-arson-self-deletion-simulation