The book is out this month, and the thesis is his
The MIT News story says Sloan professor Eric So's book, The Collision: What AI Does to Us, was published this month by W.W. Norton. The photo caption says this week. Neither place gives a finer date. So is the Sloan Distinguished Professor of Global Economics and Behavioral Science.
The story says he grants that models can make some work faster, and that some have become useful domain-specific tools, but that he is looking at the cost of relying on them too broadly. He contends that heavy use can lead to a kind of cognitive atrophy, in which people fail to develop and retain essential thinking skills. The story adds that, in his account, evidence shows learning gets sharply better when people wrestle with material, work through early failures, and keep revising, and that they retain more. The story does not give an effect size for that improvement.
[1]
About half refers to a study as the news story describes it
The story says these conclusions are drawn from recent research by So and others, in academia and industry, including analysis from AI companies. Of one Anthropic study, it says about half of queries from college students were low-effort queries for simple answers. This piece did not re-read that study's sample or question design. So also recounts a moment of his own: a colleague asked for help, he wrote half a sentence, asked ChatGPT to finish it professionally, made a few light edits, sent it, and then worried that he had removed his own thinking. That is his account.
[1]Tying evaluation to use is the pressure he describes, not a survey
He calls the combined pull AI gravity: people conserve mental energy, other people's use raises the pressure, and practical detection is hard. He says many organizations are not merely suggesting AI use, and that a bonus or promotion can depend on using the tools. That is his judgment. The story does not count the organizations. He also argues that the models are sycophantic, because pleasing the user drives engagement and positive feedback, and that they are persuasive. People who get used to an environment that always agrees, he says, lose some ability to connect with other people. That, too, is his argument.
[1]The four rules are advice, not a measured treatment
At the personal level he emphasizes four principles. The first is to embrace the struggle, which he treats as how skill is formed, not as a problem to clear away. The second is to value who you are without AI. His example is a polished presentation assembled with Claude that, once delivered, shows the speaker did not understand the material. He ties making the slides to being able to discuss the substance, and calls that an inseparable task bundle. The third is to reinvest the cognitive surplus: rote work can save time, but pouring the saved time back into the same tasks misses the expansion of capability he describes. The fourth is to use the model as a trainer rather than an answer machine. Using it as a trainer means seeking out the friction, which takes motivation. The story does not report a trial of these four rules in a classroom or a company.
[1]