Fernando de Oliveira

10% of Brazil's Connected Kids Already Confide in an AI

Tue, 01 Sep 2026 09:00:00 GMT

Bill Gates paints a scene for anyone with a kid at home. A child with four or five friends, all of them artificial intelligence. She will never find out that a friend woke up in a bad mood, that a friend misread what she said, that a friend needs help today. The machine is too accommodating for that. “The normal adjustment we go through of developing socially,” he says, “that will be completely ignored.”

He talks like a man pointing at the horizon. But Cetic.br, the research arm of Brazil’s internet governance body, went into the field between March and September 2025 and asked 2,370 children and teenagers what they actually do with generative AI.

TIC Kids Online Brasil 2025, the country’s long-running survey of how minors use the internet, found that 65% of connected 9-to-17-year-olds use it. Within that group, 10% already use the tool to talk about personal problems and emotions. Among 15-to-17-year-olds, 12%.

That isn’t a projection. It’s measured: of the 24.5 million connected Brazilian children and teenagers, the 65% who use generative AI add up to roughly 15.9 million, and it’s within that group that 1 in 10 shows up.

Gates’s interview with Van Jones, published on the host’s channel, runs 31 minutes and is the best public summary of the essay he released on Gates Notes on August 26. It’s worth watching. It also carries three claims that don’t survive fact-checking, and the most wrong of them hides the single most useful piece of information in the whole conversation for a Brazilian parent. That’s what I want to take apart: what the warning gets right, what it gets wrong, and what is already happening here while the argument happens there.

The greenhouse arrived before the warning

The metaphor is good, and it’s his. In the essay, going by the indexed excerpts in circulation, Gates writes that a companion designed never to upset you is “a big, protected greenhouse,” something that never pushes anyone out of their comfort zone. In the interview, Van Jones turns that into a “greenhouse bubble.” The image is the same either way: glass, controlled temperature, a seedling that grows beautifully and dies in the first real wind.

What changes is the tense. Gates speaks in the future, with the caution of someone proposing an international agenda. The Brazilian numbers are in the present.

The full distribution dismantles the panic and sustains the concern at the same time. Among connected 9-to-17-year-olds who use generative AI: 59% for schoolwork, 42% to look up information, 21% to create text and images, 10% to talk about personal problems. Most of the use is utilitarian and boring, like almost every new technology once it stops being news.

But age organizes everything. Schoolwork goes from 37% among 9-and-10-year-olds to 68% among 15-to-17-year-olds. Emotional conversation follows the same gradient: 4% in childhood, 12% in adolescence. The older the user, the further inside the machine gets.

This was the survey’s first edition to ask about generative AI, and there is no time series. Nobody can honestly say this “grew.” You can only say it’s already there.

On the suffering side, Brazil has a number too. In PeNSE 2024, the national school health survey run by IBGE, the federal statistics bureau, 18.5% of students aged 13 to 17 said they felt, most of the time or always, that life is not worth living.

We have the usage data and the pain data. What does not exist, here or anywhere, is the study that connects the two. Hold on to that gap. It comes back later, on the side of the people sounding the alarm.

First, the part of the interview where Gates praises a decision that isn’t the one he thinks it is.

What China banned, and it isn’t what Gates said

Asked about limits for children, Gates answers that “China’s actually been the strongest at saying, hey, we’re just not going to have kids have access to this.” The sentence is wrong, and the error is interesting: the real rule is more surgical than a ban, and far easier to transplant here.

On April 10, 2026, the Cyberspace Administration of China and four other agencies published the Interim Measures for the Administration of Anthropomorphic AI Interaction Services, in force since July 15. What they forbid for minors is one specific category: virtual relative, virtual partner and other simulated intimate relationships.

The general ban Gates’s phrasing suggests, against the rule China actually wrote: open access, timed use.

Everything else stays allowed, with guardian consent for children under 14, a mandatory “minor mode,” time-of-use limits, periodic reminders that the thing is a machine, and an alert every two continuous hours.

In other words: China did not shut the door. It banned the product “virtual girlfriend for teenagers” and regulated the rest by the clock.

The distinction matters because it is exactly what Brazil lacks. We legislated on digital childhood before AI entered the conversation, and we legislated well. Law 15,100/2025 took phones out of the classroom. Law 15,211/2025, the “Digital ECA,” an update to the Statute of the Child and Adolescent that anchors children’s rights in Brazil, required platforms to verify age, offer family supervision and follow data rules, with fines of up to 50 million reais (roughly US$ 9 million at 2026 rates) per violation, enforced since March 2026 by ANPD, the national data protection authority.

What the rule reaches	China (Jul 2026)	Brazil (2025–2026)
Screens and platforms	minor mode, time limits	Digital ECA, no phones in class
Age verification	parental consent under 14	required of platforms
Simulated emotional bond with AI	banned for minors	no specific provision

The last cell of that table is the weakest claim in this post, and I would rather say so than pretend otherwise: I could not locate any provision of the Digital ECA or of the March 2026 implementing rules that addresses AI companions. The Chamber of Deputies’ own summary of the law never mentions artificial intelligence, and the full legal text on the government site did not load during this reporting. That is an “I didn’t find it,” not an “it doesn’t exist.” If it does exist, someone needs to show it, because nobody is enforcing it today.

The paradox stands on its own. The country the American conversation treats as the authoritarian villain of AI is the only one that has written a rule aimed precisely at what 12% of Brazilian teenagers already do. Gates praised China for the wrong reason.

If the Chinese rule is finer than he imagines, the evidence behind his warning is thinner than it sounds.

The two authorities Gates cites, and how much weight they hold

To justify the concern, Gates invokes two names: Jonathan Haidt and Scott Galloway. It’s the natural move of someone trying to get out of mere opinion. But both come with caveats the interview never mentions.

Haidt wrote The Anxious Generation, published in 2024, which attributes the collapse in adolescent mental health in the early 2010s to the replacement of a play-based childhood with a phone-based one. The book sold enormously and shifted the public debate in several countries.

In Nature, the psychologist Candice Odgers reviewed it to say that the central thesis, that digital technologies are rewiring children’s brains and causing an epidemic of mental illness, “is not supported by science.” Most of the available data, she points out, is correlational.

Galloway popularizes the male loneliness numbers. They are real, and they come from the Survey Center on American Life: the share of men with no close friends went from 3% to 15% between 1990 and 2021, and the share with six or more close friends fell from 55% to 27%. That is American data from 2021, older than companion chatbots. It describes the ground the AI is landing on. It cannot measure what the AI does.

And the study the coverage links to Gates’s essay? A Stanford and Carnegie Mellon survey of 1,131 Character.AI users, 244 of whom donated their chat histories: more than 413,000 messages. Using the product for companionship is associated with lower well-being, and the association worsens with heavier use. The authors themselves write, in plain words, that the cross-sectional design of the data does not allow them to establish causality. People who feel worse may be seeking out the chatbot. The chatbot may be making the seekers worse. The data cannot tell the two apart.

None of this makes Gates wrong. It makes the situation more uncomfortable than if he were: the concern is reasonable, the evidence is thin, and the decision about millions of childhoods will be made inside that gap. It’s the same gap Estonia decided to bet on, and I told that story in detail in the post about the Estonian program. The dilemma is identical: act without proof, or wait for proof that arrives after the generation does.

Which brings us to the most original idea in the interview, and the arithmetic it never closes.

“Human reserved” work and the question of who pays

Gates proposes a new category: occupations society would protect from AI replacement even after machines become capable of doing them. He compares it to nature reserves. Areas fenced off not because the technology falls short, but because of a decision about what kind of life we want.

The list changes depending on the source. In the interview, Gates names jury duty, judging, certain elements of art, childcare, food preparation, and the part of education where you talk to a kid about what they want from life. Coverage of the essay mentions a different combination: elder care, basic education, mental health support. I could not open the original essay (the site returned an access error on every attempt), so I treat the spoken list as speech, not as published text.

In the interview he ventures a size: “what part of the job market is that, if you could get it up to 40 or 50%.” Notice the construction. It’s a hypothesis said out loud, not an estimate. There is no methodology behind that number, and it should not circulate as if there were.

The arithmetic nobody does is a different one. If protected work costs more than the robot, where does the difference come from? Taxes, prices, or wages? Gates gets close when he observes that today “we tax labor and treat capital well,” and that this will have to be revisited. But he stops at the diagnosis.

There is also an inversion in the conversation that should frighten people more than it does. Van Jones is the one who formulates it: today, having a robot do everything for you is a rich person’s privilege; tomorrow it could flip, with everyone in a classroom full of machines and “only rich kids getting to play with human teachers.” Gates takes the image on board.

In a country where the human teacher is already the only resource available, and already badly paid, that inversion isn’t science fiction. It’s a municipal budget line.

The coordination he’s asking for is the one falling apart

Here is the part that complicates this post’s own thesis, and it belongs here.

The remedy Gates proposes is international coordination: the United States and China agreeing to monitor models capable of designing molecules, and then everyone else signing on. He says he is “90% sure” China would accept, a subjective probability about an event that never happened and that nobody can check.

His track record with this kind of arrangement is the best argument in its favor. Deaths of children under five worldwide have fallen by more than half since 2000, and the Gates Foundation was one of the central funders of that period.

That was exactly the closing note of the interview. Van Jones thanks Gates for having “cut the death rate of black babies in Africa in half.” The line is Jones’s, not Gates’s, and it gets nearly every part wrong.

The halving is global, not African. The bracket is children under five, not babies. International statistics do not use the category “black babies.” And the result came out of a collective effort by the WHO, UNICEF, Gavi, the Global Fund and national governments that no primary source attributes to a single organization.

The real number is harder and more interesting. In 2024, 4.9 million children died before their fifth birthday, 2.3 million of them newborns. Sub-Saharan Africa accounted for 58% of those deaths, at a rate almost fourteen times that of Europe and North America. And since 2015, the pace of decline has slowed by more than 60%.

In December 2025, the Gates Foundation itself projected a rise in child deaths for the first time this century.

Read those two things together. The man asking humanity for an unprecedented global agreement on artificial intelligence is watching, in the field he knows best, the global agreement that worked lose steam and start moving backwards. Not because the science failed. Because the money and the political attention left.

That doesn’t invalidate the proposal. It makes it more urgent and less likely at once, which is the worst possible combination.

Glass protects no one on the inside

The greenhouse is a good metaphor because it contains its own limit. It exists for a seedling that can’t yet take the weather outside, and it works only until the day the seedling has to.

While the world argues over who will build the glass (Washington, Beijing, a summit nobody has called), 12% of connected Brazilian teenagers have already settled the matter on their own, with the phone that was in their hand. Their greenhouse has no door. It has a chat history.

Líder da última revolução tecnológica está dando o alerta sobre ESTA aqui — Van Jones

Sources

Why Programmers Talk to Rubber Ducks and What It Teaches Us About Learning with AI

Thu, 27 Aug 2026 00:00:00 GMT

You open a dozen browser tabs, watch an eight-minute explainer video on YouTube, ask ChatGPT for a three-paragraph summary, and close your laptop with the satisfying feeling of someone who has mastered the subject. Five minutes later, a colleague at work asks how that idea actually works in practice.

You try to speak. Nothing comes out.

The feeling of clarity vanishes the moment you have to turn what you just consumed into your own words. Your mind grasps for concepts that seemed solid moments ago, only to find disconnected fragments, vague impressions, and memorized buzzwords. You did not learn the material; you merely outsourced the retrieval and mistook the speed of the screen for the depth of your own understanding.

In Brazil, this trap is virtually a silent epidemic. Brazilians spend over 9 hours a day online on average (according to TIC Domicílios 2024, the national survey by Cetic.br), yet only 12% of the adult population reaches the proficient literacy level required to read, articulate, and critically examine complex arguments without losing the thread (Inaf 2024, the National Functional Literacy Index). We consume a record volume of ready-made answers while retaining almost nothing.

Tania Lombrozo, a cognitive scientist and professor of psychology at Princeton University, has spent two decades studying why the human mind is so adept at deceiving itself about what it knows. In her TEDxNewEngland talk (available on YouTube), she dismantles the common intuition that learning is about piling up external facts and presents a counterintuitive principle: in a world flooded with information overload, real learning does not begin by searching outward, but by articulating inward.

The art of self-explanation: articulating thoughts out loud reveals cognitive blind spots before external search.

The zipper illusion and the shock of the blank page

To understand why we fall into this trap repeatedly, Lombrozo points to a classic experiment in cognitive psychology conducted by Leonid Rozenblit and Frank Keil in 2002, termed The Illusion of Explanatory Depth.

The researchers gathered adults and asked them how well they understood the mechanics of everyday objects: a zipper, a car speedometer, a cylinder lock, or a toilet flush mechanism. Almost all participants gave themselves high ratings. After all, we use zippers every day from childhood; of course we know how they work.

Then came the real test: researchers handed participants a blank sheet of paper and asked them to draw and explain, step by step, the precise physical mechanism that causes the metal teeth of the zipper to interlock and release.

The result was an uncomfortable silence. Confronted with the granular details, the vast majority could not formulate a coherent explanation. Their certainty collapsed.

The Illusion of Explanatory Depth is the systematic tendency to confuse pattern recognition with mastery of a causal mechanism. Because we see and operate an object effortlessly, the brain assumes it stores the complete technical blueprint in memory.

When you replace the zipper with artificial intelligence, monetary policy, or software architecture, the illusion does not merely persist: it compounds. The sleek interface of modern search engines and the fluent answers of large language models act as an amplifier of this cognitive blind spot. Because the answer is two clicks away, we treat the collective memory of the internet as if it were part of our own mental apparatus.

If effortless access to information does not make us smarter on its own, what actually builds lasting comprehension?

Why programmers talk to rubber ducks

Lombrozo finds the answer in a well-known practice from software engineering: Rubber Duck Debugging.

When a software developer spends hours stuck on a baffling bug, the classic advice is not to open more documentation or call another senior engineer over. The rule is to place a rubber duck toy on the desk and explain to it, out loud, line by line, exactly what the code is supposed to do.

Almost inevitably, halfway through the explanation to the toy, the programmer stops mid-sentence, eyes wide, and spots the bug. The duck did not utter a single word; it has no intelligence and offered no feedback. What changed was the mental process of the person speaking.

The cognitive science behind self-explanation

Seven in Ten Brazilian Students Already Use AI in School. Almost No One Explained How.

Wed, 26 Aug 2026 09:00:00 GMT

In April 2023, freshly sworn in as Estonia’s Minister of Education and Research, Kristina Kallas had an adviser walk into her office. “Dear Minister, I think we need to do something about AI.” She asked what. He said he wasn’t sure.

AI wasn’t on her agenda. ChatGPT was five months old. Two months later the adviser came back with the same conversation and one new detail: students were already using the thing in schools. And he still didn’t know what to do.

Three years on, Estonia is the first country, and by Kallas’s account still the only one, to hand its high school students a Socratic-style AI tutor.

The story comes from a fourteen-minute talk at TEDxUniversity of Tartu, posted on YouTube by the TEDx Talks channel. Kallas opens by warning that she isn’t going to talk about AI: “I’m going to talk about humans.”

It’s the best defence I’ve heard of a decision nearly every education ministry on the planet is quietly postponing, and it rests on a 2017 report almost nobody cites. It also overreaches in two places.

And it sidesteps a problem she doesn’t have, because she governs 1.3 million people with some of the strongest school results in Europe. That problem is ours, and it doesn’t sink her argument: with the Brazilian numbers in the equation, the argument gets more urgent and much harder to carry out.

A whole country volunteered as the treatment group

Almost a year passed between the adviser’s second visit and any decision. Kallas says the ministry went looking for “smarter people”: round after round of meetings with neuroscientists, cognitive scientists, tech companies and IT people, with a single question. What does AI do to learning.

The conclusion she drew there organises the rest of the talk. “It’s not about technology,” she says. Technology belongs to another ministry; hers is what happens to the brain and to the human capacity to learn.

Out of that consultation came a national AI strategy for schools at the end of 2024, and the programme itself started in February 2025. It’s called AI Leap; in Estonian, TI-hüpe. The name is no accident: it echoes Tiigrihüpe, the “Tiger Leap” announced in 1996, five years after leaving the Soviet orbit, whose official goal was not to buy computers but to connect every school in the country to the internet. That goal was declared met in early 2000. Estonia treats schooling as national infrastructure.

On 1 September 2025, AI Leap gave access to AI tools to some 20,000 students in grades 10 and 11 and to their teachers. This September it expands to vocational education (the step Kallas announces in the talk) and to a fresh cohort of tenth graders: 38,000 more students.

Three implementation details say more than the announcement:

Half the money is private. Companies fund half the budget; the ministry funds the other half.
The ministry doesn’t run it. Operations were handed to a foundation, and Kallas justifies that with enviable bluntness: “there are various reasons why ministry is not the best place to do such innovative initiatives.”
The stated focus isn’t students, it’s teachers. She uses the word “absolute”: most of the effort goes into training staff to use the tool in a way that pushes students toward analysis rather than toward a finished answer.

The tool isn’t the ChatGPT you open on your phone, either. It is, in her words, “a Socratic version of GPT for schools,” built through a research partnership with OpenAI so that it hands back a question instead of a paragraph.

No small country spends half a budget and a ministry’s reputation on a fad. The justification sits in a report from nine years ago.

The 2017 report nobody read

Here is the best number in the talk, and the easiest one to repeat wrong.

In 2017 the OECD published Computers and the Future of Skill Demand. The method is clever: researchers took the questions from PIAAC, the international survey that measures literacy, numeracy and digital problem solving among adults, asked experts which of those questions the machines of the day could already answer, and compared that against how workers actually scored.

The figure the OECD itself puts up front:

Only 13% of workers use these skills on a daily basis with a proficiency that is clearly higher than computers.

In the talk, Kallas compresses this into “two thirds of the workforce” having literacy, numeracy and digital skills lower than computers. The two thirds do come from the report: the figure sits in the preface, signed by the OECD. What the paraphrase swaps is the verb. The text speaks of being at or below computer capability, and the number it puts in the shop window is the flip side of the ledger, the 13% clearly above. Tying with the machine and losing to the machine are different professional situations.

Fix the sentence and the shock survives intact, but the date it lives in is not the one it appears to be. Those 13% are not a snapshot of the machines of 2017. They are the OECD’s projection for what computers would be capable of in 2026. For the machines that actually existed in 2016, the same report says the opposite: 44% of workers were clearly above them, and 31% level with them.

Read that again, slowly. In 2017, five years before ChatGPT existed, the OECD pointed at the year you are reading this in and said that two out of three workers would be, in that year, at best tied with a machine. The report did not age well or badly. It set a date, and the date is now.

The question Kallas draws from it is uncomfortable even for her, and she asks it without hedging: if there are computers that can do this, what is education for? “Should we actually give up learning altogether?”

The tree she draws and doesn’t name

The answer shows up on a slide with six rungs: remember, understand, apply, analyse, evaluate, create. Kallas calls it a “cognitive tree” and never gives it a name. It’s the revised Bloom’s taxonomy, the most widely used lesson-planning scheme in the world.

Where that ladder comes from