{"text":[[{"start":5,"text":"The writer is a theoretical neuroscientist and author of ‘Robot-Proof’ "}],[{"start":9.45,"text":"Generative AI is, by design, a machine for producing the feeling of competence without the substance of it. To see what sort of impact this is having on our ability to think, I asked a classroom of students to wear EEG headsets while they worked on the same assignment with the same AI agent. From the front of the room they looked indistinguishable: heads down, screens glowing, fingers tapping. Inside their skulls, however, two very different stories were unfolding."}],[{"start":39.95,"text":"In most of the students, the EEG measuring their brain activity revealed that the high-frequency “gamma” oscillations that mark cognitive effort collapsed within minutes of using the AI agent. Their neural state drifted towards something closer to watching TV than solving maths problems. "}],[{"start":58.85,"text":"In a small few, however, the gamma lit up. These students, it seemed, were arguing with the machine, pushing back on its confident answers and forcing the AI to critique their own thinking in turn. The final essays produced might have looked broadly similar but over the course of a term it is this handful of students who would get measurably more intellectual benefit. "}],[{"start":81,"text":"I have been documenting the divergence in responses to technology use for the better part of two decades and have seen the same pattern show up everywhere: two people use the same tool for the same task. One’s brain gets sharper. One’s brain goes quiet. "}],[{"start":96.1,"text":"You can sort technology users into types — the “automators” who copy and paste information, the “validators” who seek confirmation bias and the “cyborgs” who spar with machines. But the taxonomy matters less than the mechanism producing it."}],[{"start":112.39999999999999,"text":"A Harvard study published in 2019 captured this in a single, counterintuitive finding. Students challenged to wrestle with problems learnt significantly more than those in traditional lectures, yet they reported feeling as if they had learnt less. Our brains mistake the smooth, fluent sensation of being told something — whether in a lecture hall or by AI — for the harder, messier process of actually learning. And generative AI is the most fluent thing that humans have ever built."}],[{"start":141.29999999999998,"text":"To resist that requires choosing productive discomfort: being wrong, or doing the unglamorous work of interrogating an answer that looks correct. "}],[{"start":150.1,"text":"Most of us do not do that. Between 2012 and 2016, the education researcher (and my wife) Norma Ming and I worked through the discussion-forum activity of roughly 60,000 students across undergraduate and MBA programmes. The finding that stayed with me is that the students who earned top marks were those who were “wrong” most often in the forums. They proposed, they explored, they were frequently and visibly incorrect. Those who passed safely — nine out of ten students — were far less likely to venture a claim they couldn’t defend. (The failing students, if you are curious, talked mostly about weekend plans and the funny thing their dog had once done.) "}],[{"start":192.75,"text":"This is what the AI benchmarks miss when frontier labs announce near-perfect scores on coding, law or medical exams. These exams test the models in isolation, but in the field, AI works alongside a person — a radiologist weighing up its read, a junior lawyer checking its brief. "}],[{"start":211.1,"text":"Policy is making the same mistake at scale. Article 14 of the EU’s AI Act mandates “human oversight” for high-risk AI deployments and treats the requirement as a defensive checkbox: put a human in the loop and the loop is safe. But placing a human operator in a loop with a fluent, persuasive AI almost guarantees automation bias. The operator becomes a bored clerk, rubber-stamping authoritative-sounding hallucinations. For most people, the alternative — active engagement with every output — is cognitively exhausting. "}],[{"start":244.2,"text":"The fix is not more oversight. Britain’s AI Security Institute, America’s CAISI and the EU’s AI Office test AI models the way labs do: in isolation. They should shift to evaluating AI’s effect on the humans who use it. My proposal is a Hybrid Intelligence Index that measures whether a human-machine collaboration leaves the human user sharper or duller over time. "}],[{"start":269.95,"text":"That means designing for friction. When I ran a version of my experiments in which the AI was instructed to respond with questions and context instead of answers, the proportion of high-gamma, actively engaged students more than doubled. "}],[{"start":285.05,"text":"In a labour market increasingly defined by AI products, the gap between those able to argue with confident machines and those who cannot will compound. The economies that optimise for fluency risk automating away the very minds they need."}],[{"start":305.6,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1780982700_8571.mp3"}