Now, let's start with the research because the research is where the nuance lives. As always, academic papers can be pretty dense, but my job here is to translate it so we can all have the same conversation irrespective of our background knowledge. I won't use any jargon. I promise. If I accidentally say back propagation or stochastic gradient descent, feel free to unsubscribe, but please don't. Just judge me in the comments instead. Anyway, the study that generated the most attention is from MIT's media lab, a preprint titled Your Brain on ChatGPT by Cosmina at all 2025. The researchers fitted 54 participants aged 18 to 39 with EEG headsets that measured brain activity across 32 neural regions while they wrote SAT style essays. One group used ChatGPT, one group used Google search, one group used nothing, just their own brains, which in 2026 feels less like a control group and more like a dystopian episode of survivor. The results were striking. Participants who used ChatGPT showed the weakest neural connectivity of all three groups. The lowest brain engagement across areas associated with memory, attention, critical thinking, and creative reasoning. Search engine users showed moderate engagement, brain only writers, however, showed the strongest, most distributed neural networks. 83% of the AI users could not quote a single line from the essays they had just written. They had total cognitive amnesia. They were basically living out the plot of The Hangover, but instead of waking up with a tiger in a hotel room, they woke up with a 2,000-word essay on The Great Gatsby that they had absolutely no memory of creating. Their brains basically went into screen saver mode the moment they hit enter. Now, important caveats here. This is a preprint. It has not been peer-reviewed yet. The sample size is 54, which is very small. A 2026 critique by Stanovich at all flags methodological concerns. These are preliminary findings that need replication at scale, not settled clinical results. But the directional signal is consistent with every other study in this space. And the finding I want to highlight is this. In a fourth session, the ChatGPT users were reassigned to write without AI. Their brain activity did not return to normal. The cognitive debt persisted even after the tool was removed. So your brain essentially goes, "Oh, we're not doing the thinking thing anymore. Cool. I'm just going to take a nap in the corner" and then it just refuses to wake up.

But, and this is the part that gets lost in the alarming headlines, a 2024 study published in PNAS found that students who critically engage with AI, so asking questions, editing ideas, revising thoughtfully rather than just copying outputs, actually performed better and reported less mental fatigue. So again, the tool is not the problem, the process is. How you use AI determines whether it builds your thinking or replaces it. But ultimately, this is not a story just about students. I want to be very clear about that because I think framing it as a youth problem lets everyone else off the hook. The MIT study was conducted on adults aged 18 to 39. The Carnegie Melon study surveyed adult knowledge workers. The cognitive offloading research measures working professionals. This is essentially a use it or lose it phenomenon that applies to everyone regardless of age. Thinking requires energy. And I want to be honest about the world we're all operating in. The economic environment is demanding, exhausting, and relentless. There are a bazillion things competing for your effort and attention every single minute of every single day. Reaching for a tool that gives you a cognitive break is not some sort of a moral failure. It is profoundly human. Nobody should be shamed for using AI to lighten a mental load that was already crushing them before the technology existed. If you use ChatGPT to write an email that says per my last email in a way that won't get you sent to HR, you're just surviving. It's cool. This is a safe space.

And here's a counterpoint that I think is very important to engage with honestly. Education Next published a piece with a very provocative title. AI didn't destroy critical thinking. We did. Their argument is that there was no golden age of critical thinking before AI came along. NAEP scores have been stagnant for decades. The collegiate learning assessment found college students were academically adrift and learning almost nothing across their years of college. And that research was published well before ChatGPT existed. AI in this framing didn't erode critical thinking. It instead exposed how poorly educational systems have been teaching it all along. I think there is a genuine truth in that. But I also think it doesn't let anyone off the hook. If anything, it makes the situation more urgent. If critical thinking was already fragile, adding a tool that makes it even easier to avoid thinking is essentially pouring water on a drowning man. The pre-existing weakness is not an argument for complacency. It is an argument for acting faster. By the way, I did actually write an ebook on developing critical thinking, and I'll leave a link to that below if you want to have a look. It was written entirely by my own human brain, which I can prove because it contains at least three typos and a deeply questionable metaphor about trees. But anyway, the research is also clear that when you stop exercising cognitive abilities, they atrophy. The MIT finding that brain activity didn't bounce back after AI was removed should give everyone a pause. Not because it means you're broken if you use ChatGPT, but because it suggests there are real trade-offs to be aware of. A teacher in Minneapolis named Maureen Mulvehill banned all technology from her AP literature class at Washburn High School. phones gone, laptops gone, everything done with pencil and paper. In September, 46% of her students felt confident about their reading ability. By February, that number was 95%. The students, after, of course, initial resistance, described falling in love with the analog approach. A Wharton study on phone bans in college found that struggling students saw the biggest grade improvements, roughly double the average effect. First year students benefited most precisely the group still forming their study habits. Sometimes removing the tool is what rebuilds the muscle, but removing it permanently is not the answer either.

A few days ago, the UK announced a ban on social media for under 16s. Snapchat, Tik Tok, YouTube, Instagram, X the works. This is the furthest any country has gone, by the way, going even beyond Australia's landmark legislation from last year. 20 US states now ban phones during the school day. France, Spain, Greece, Slovenia, and South Korea have all introduced or are working on similar restrictions. It took a decade of evidence about what unregulated social media access does to developing minds before regulation caught up. The research on AI and cognitive development is following the same trajectory only faster because the cognitive impact of AI is more direct than social media. Social media captured attention. AI captures thinking itself. A Rand Corporation survey from March 2026 found that 70% of students themselves are worried that AI is eroding their critical thinking skills. That number grew from 48% to 65% among high schoolers in just one year. When the people being affected are telling you they are concerned and the concern is growing, that is data. And the UK technology minister, Liz Scendle, specifically flagged AI chatbots as a separate concern, noting that children were forming one-to-one relationships with AI systems that were not designed with child safety in mind. The social media ban is one regulatory response. The AI dimension hasn't been addressed yet, but the pressure is building.

I think this is where the field needs to go. And I want to say this carefully because it's a vision rather than a prescription. Instead of putting guard rails on general chat bots, filtering out smut, restricting mental health conversations, adding safety warnings exposed. What if the investment went into developing highly specialized purpose-built AI applications? An educational AI designed around the Socratic method, for example. I'm personally a huge fan of the Socratic method as a framework for learning that develops cognitive abilities by questioning, challenging, and genuinely refusing to just hand over the answer. Just imagine a tool that when a student asks, "What caused the French Revolution?" doesn't give them a summary and instead it asks well what do you know about the conditions in France before 1789 what was happening economically what was happening politically basically the AI acts like a tough love personal trainer or Gordon Ramsay in a kitchen refusing to serve the dish until you actually chop the audience yourself and then builds on whatever the student provides correcting misconceptions filling gaps but always making the student do the thinking that is a fundamentally different interaction from here's your answer. Would you like it in bullet points? Then there's healthcare AI built specifically for clinical context with appropriate safeguards. Research AI designed for synthesis and verification rather than generation. These things do exist in early forms. Khan Academy's Kmigo is a notable example of this, but they are nowhere near as prevalent as the general purpose tools. The investment, the research energy, the venture capital, the deployment urgency, it overwhelmingly flows towards general LLMs. And that balance I think needs to shift. And critically different people learn differently. Visual learners, kinesthetic learners, verbal processors, social learners. An educational AI built with this diversity in mind would be a fundamentally different product from a chatbot that gives the same kind of response to everyone. The research says that critical engagement with AI, questioning it, editing its output, arguing with it, actually improves learning outcomes. So build a tool that forces that engagement. Let's build a tool that makes us think harder, not less.

The technology to do this exists. The research base to inform it exists. What's missing is the priority. Because we have reached the ultimate holy grail of this channel, my friends, incentives. And seriously, I talk about incentives so much on this channel. I should probably just tattoo the word directly onto my forehead. That way, if I ever pass out from reading too many academic papers, the paramedics will at least know exactly what my core personality trait was. But this requires studies, real studies with adequate sample sizes across diverse age groups and populations with longitudinal follow-ups, not 54 participants over four months. Important as that MIT study is as a starting point, we need hundreds, ideally thousands followed over years. We need studies that track cognitive development in children who use AI from an early age versus those who don't. We need studies that differentiate between types of AI use. passive consumption versus active engagement. We need studies that account for the diversity of learning styles and neurological profiles. If we frame this as a further research is urgently needed rather than ban this whole thing, we arrive at a much better destination for everyone irrespective of who makes money or who pays how much tax.

I started this video with a headline about college students losing the ability to read and the research tells a story that is more complicated and more interesting than that headline suggests. AI is measurably changing how human brains engage with information. The cognitive cost of indiscriminate use are real documented and very concerning especially for developing minds that may never build the capacities they're outsourcing. But AI used critically actively as a tool for challenge rather than a shortcut to answers can actually make people better thinkers. The difference between those two outcomes comes down to design, deployment, and intention. The field of computer science, my field, needs to take this seriously, arguably in an interdisciplinary effort with education, not as a PR problem to manage with safety filters, but as a genuine research and development challenge. How do you build AI that makes humans smarter rather than lazier? How do you deploy it in education in a way that respects the diversity of how people learn and the vulnerability of developing minds? How do you do that while still allowing the technology to do what it genuinely does well? These are not rhetorical questions. They are research questions and they deserve the same intensity of funding and attention that currently goes towards making models faster, cheaper, and more capable. I think about this a lot and I keep coming back to the same thought. The technology is extraordinary. It is awe inspiring. The potential is real. But potential without intentionality is just capability waiting to be misused. We saw it with social media, a technology with genuine capacity to connect people, deployed without guard rails, optimized for engagement rather than well-being. And it took a decade of documented harm before the UK banned for children. We are at an earlier point on the same curve with AI and we have the advantage of seeing the pattern this time. The question now is whether we act on it or we wait for the damage to accumulate and the headlines to get even worse. I don't have all of the answers, but I think if we don't start building AI that makes us smarter, we're going to wind up in a world where the machines are doing the art and the philosophy while humans are relegated to doing the dishes and filling out captchas to prove we still exist. And frankly, I really don't want to live in a world where a robot laughs at my bad jokes or do the dishes for that matter. If you want the full context on how AI is changing education assessment and what a degree even means in 2026, I made a dedicated video about Princeton ending its 133-year honor code. And it connects directly to everything we discussed today. That's the video that I would watch next. Thank you so much for watching this one. Subscribe and I'll see you all in the next.