Steven Bartlett wanted to add a new role to his company, so he sat down and wrote a two-page memo about it. He did the whole thing himself: the background context, the kind of person he wanted, how the organization would be structured, what impact the role would have, who they would report to, and how he would measure whether it had worked. It took him four hours.
Then he pasted it into three AI assistants he uses — Gemini, ChatGPT and Grok — with one instruction: critique my work and tell me how I could have written this better, pretending that you're a top consultant from Boston Consulting Group. The models came back with a long analysis. One told him the memo needed financial forecasts about the role's impact, and that he had to be clearer about who would report to whom. He went back into the document and added those things.
He is well aware of the shortcut he declined. "I could have, within 30 seconds, said, hey, can you write me this job description?" he says on the Diary of a CEO segment published on 11 September 2026, a replay of the most-listened-to stretch of an earlier conversation with brain experts. ChatGPT already remembers his company, he points out. "And I could have saved myself three and a half hours."
His guest's answer: "But that's not why you're the CEO of your company."
The reason he took the slow route
Bartlett says he wrote the memo himself partly because he is scared, and partly because of what writing did for him at 23. For five years he wrote every single day and simplified what he thought into 140 characters so he could tweet it. Without that, he says, he would never have become "religiously attached" to the ideas he was working out.
The guest gave that practice a name. It draws on the basal ganglia, he said — a part of the brain that needs "practice, practice, practice" — and once that foundation is in, "then you become much better cognitively." The cognitive side, in his account, involves two big learning systems that have to work together. He extended the point to schools, complaining that rote learning is now treated as if it were something bad: "No, that's practice. You need to have a foundation. You have to memorize things." Reading and mathematics have to become fluent, he said, before anything else rests on them.
The half-sentence that keeps winning
Against that sits a friend the guest declines to name. He has known the man for ten years, and the essays now appearing daily on his LinkedIn profile do not sound like him. "That's not my friend. That's not how he speaks." The man is deferring all of his thinking to the tool, the guest said — and it is working. He is getting more likes and more reach than he ever got in his life. "And so why would he go back? Why would he go back to harder?"
Bartlett had a version of the same story. A friend of his, a heavy ChatGPT user, posted an article the group knew he had not written in his own voice, so they asked him in their WhatsApp group to show the prompt. He posted it. The prompt was half a sentence long — write something about this issue and that issue, and include this — and it had produced a two- or three-page article, now on the man's LinkedIn. "That's exactly the wrong way to use it," the guest said. "You're not going to improve yourself, your brain at all, if you do that."
Bartlett expects most people to do what is easiest, fastest and cheapest, and what gives them a short-term advantage; on that basis he thinks the ability to think critically will erode to some degree. He also notices the surface of it everywhere: everything he reads now has em dashes in it that he never saw two years ago.
The temptation came out as a thought experiment about two versions of Bartlett sitting side by side — the ordinary one, and one with a PhD in everything. Of course you would get the second one to do your work. Even if it was bad for you?
Acquiring information versus understanding it
Bartlett draws a distinction he says he has not resolved. Since using ChatGPT he thinks he is learning more information, and probably losing some ability to think critically. "And I think they're two very different things."
His example is school German. He memorized the words he needed to pass the exam and cannot speak the language now, because he never understood it. He suspects something similar is happening with AI-assisted work: he might be able to regurgitate things, but whether he understands them is a question mark. That matters to him specifically, he says, because his businesses were built on thinking critically about a problem and arriving at a slightly novel solution from first principles — and he is concerned his own ChatGPT use will make him less effective at exactly that.
What the MIT study does and does not show
During the conversation Bartlett read out figures from MIT's "Your Brain on ChatGPT" work: that the ChatGPT group had nearly two times less activity in the part of the brain linked to memory than the group that wrote unaided, and that 83% of ChatGPT users could not remember what they had just written or correctly quote their own finished essay. The guest's explanation was that you are not part of the experience of writing it, so the material never gets encoded.
The study is narrower than those numbers suggest. MIT's project overview describes a four-month experiment with 54 participants writing essays across three sessions using ChatGPT, a search engine or no external tools; 18 returned for a fourth session in which the groups swapped. The researchers combined EEG recordings with linguistic analysis, interviews and scoring by teachers and an AI judge. They call the June 2025 findings preliminary, and note that a small, geographically concentrated academic sample and a single writing task limit how far the results generalize. The EEG measured connectivity between brain regions rather than pinpointing activity in memory structures such as the hippocampus.
The swapped session is the part that looks most like Bartlett's routine: participants who first wrote unaided and then used ChatGPT showed higher recall and renewed widespread connectivity, while those moving the other way showed weaker connectivity. The experiment did not separately label brainstorming, drafting and revision, so it cannot say which way of using assistance produces which effect. It does not establish that writing first and asking for criticism afterwards protects anyone's thinking.
Calculators, a clone and a five-year guess
The guest's own reference point is not alarming. Electronic calculators arrived at least 30 or 40 years ago, and the results are in: there is probably less brain activity when someone punches the numbers in, but people became more accurate and more productive. "It freed up cognitive space for them to do other things." You have to look at the risks and the benefits, he said. He uses AI heavily himself, including a clone built from his uploaded books, research papers, public television specials and scripts, which he asks to answer things for him. That can be very helpful — but not, he said, if he is not interacting with it, not thinking with it.
For people who never interact, he offered a speculation rather than a finding: it will decrease their cognitive load, "which is going to potentially increase their risk of dementia." Asked how long unintended consequences would take to appear, he guessed five years rather than 20 or 30, because everything is accelerated. Earlier in the same conversation he had said that nobody can predict where this ends up, and that long-term use is the real issue.
His practical rules were short. Use it to amplify, not replace, thinking. Alternate AI-assisted tasks with brain-only tasks. Keep doing deep learning, problem-solving and memorization — so that you can then ask the AI to test you on them. Question what comes back: is that really true, can you explain that better, the way you would push a teacher. Asked what a good relationship with AI looks like, he said: "That you don't use it to do your work, you interact with it to get better work."
Bartlett, for his part, said he was wondering whether to put some rules in place for himself — self-regulation, as the guest called it — the way he has with social media, where his notifications are off and TikTok is not on his phone at all.