What AI is doing to your thinking (and what to do about it)

ChatGPT, released to the public in November 2022, reached a hundred million monthly users two months later.^[1] This was the fastest adoption of any consumer product in history. The previous record, TikTok, had taken nine months.

By 2024 the Microsoft and LinkedIn Work Trend Index was reporting that 75% of knowledge workers globally were using generative AI in their work.^[2] By 2025, virtually every commercial software product had an AI feature. By 2026, where we now are, AI tooling has been quietly integrated into the daily cognitive routines of more than a billion people.

This piece is not an argument against AI. The argument here is more specific. It is that the speed of integration has outpaced our ability to notice what we are giving up, and that some of what we are giving up is worth knowing about.

What we are actually trading

The most obvious thing AI tools do is eliminate small frictions. The friction of remembering a half-forgotten fact. The friction of writing a difficult sentence. The friction of summarising a long document. The friction of finding the right opening for an email. The friction of having to decide between three options when a tool will, given enough context, just decide.

All of these frictions, until very recently, were the conditions under which much of our actual thinking happened.

When you used to try to remember a half-forgotten fact, you would search through your own mind in a way that woke up related memories, prompted associations, generated unintended ideas. You would stop in the middle of typing an email and look out the window because the right phrasing had not arrived yet, and in the window-looking some thought about the person you were writing to would, occasionally, arrive. You would read a long document and form your own structure of it because nobody had structured it for you.

Now: ask the chatbot. Answer in two seconds. No rummaging. No window-looking. No structuring of your own.

The transaction looks like a clean gain in efficiency. The hidden cost is the cognition that used to happen in the gap. This is the same argument Nicholas Carr made in 2010 about search engines in The Shallows, and it has aged uncomfortably well.^[4] AI is the same trend, accelerated.

The cognitive offloading literature

There is a small but growing research literature on what psychologists call cognitive offloading — the practice of externalising cognitive tasks to tools. The literature is older than ChatGPT; it began with studies of how reliance on GPS affects spatial memory, how reliance on search engines affects fact retention, how reliance on calculators affects mental arithmetic.

The general finding, summarised in a 2023 review by Liu and colleagues in the Annual Review of Psychology, is that the offloaded capacity tends to weaken in proportion to the offloading.^[3] This is not surprising — the same principle applies to physical training. What is more interesting in the literature is what happens to the related cognitive capacities. Offloading navigation to GPS, for instance, has been associated with measurable decline in spatial reasoning more generally, not just in the specific navigation task.

The cognitive offloading literature on generative AI is in its early days, and the longitudinal data does not yet exist. But the existing framework, applied to current AI use, suggests a useful question: when you ask a chatbot to draft an email, are you offloading just the typing, or are you also offloading the part where you would have thought about what you wanted to say?

For many people, much of the time, the answer is: both.

The “closer-in” feeling

There is a subtler thing that AI is doing that is harder to put into measurable terms, but worth naming.

Many people who use conversational AI heavily report, after a period of use, a vague feeling that information has moved closer to the inside of their head than it used to be. The interface no longer feels like a tool over there, separate from them. It feels like a way of thinking. The chatbot's output starts to feel, slightly, like one's own thoughts.

This is partly an interface effect. Search engines, with their list of links and their search box, kept a small physical distance between the user and the information. The chatbot mimics conversation, which is the most intimate form of human exchange. The distance closes.

Jenny Odell's 2019 book How to Do Nothing identifies, in a different context, why the closing of this distance matters.^[5] The gap between you and information was where your own thinking lived. When the gap closes, the thinking does not move to you; it moves to the platform. You become a fluent passer-on of conclusions that the platform formed.

In small doses this is fine, and useful, and the same charge can be levelled at any tool, including books. In large doses, applied to most of your daily cognition, it is a slow shift in where the thinking happens, and where the thinking happens is more or less what we mean when we say “you.”

What to do

Three small practices, supported by the same broader literature this site keeps coming back to.

Restore one or two cognitive gaps a day. When you reach for the chatbot, sometimes do the thing yourself first. Write the bad version of the sentence. Try to remember the fact. Form your own opinion before consulting any model's. The gap restored is not all the gaps, but it is more than zero.

Use AI tools for what they are good at. Genuine factual lookups in two seconds, summarising a long technical document, drafting a starting structure you will then rewrite, translating between languages. These are net gains. Use them.

Keep the cognitive frictions you actually need. Writing the difficult email yourself, working out what you actually want to say to a friend, sitting with not knowing something for an unhurried minute. These are not friction. These are where your own thinking lives. Letting AI smooth them away is a different transaction from letting AI find the capital of Burkina Faso for you.

Install one daily session in which you don't reach. A morning hour, a walk, an evening read. Not phone-free, not anti-AI, just a window in which the tools are not what you are using. The default mode network, which we have already discussed elsewhere on this site, does not get the conditions it needs unless some of the day is set aside for it.

A small note on the AI debate

Public conversation about AI tends to oscillate between two equally unhelpful positions: AI is going to destroy us, or AI is the best thing that has ever happened. Both positions tend to be held with more conviction than the underlying evidence supports.

The position taken on this site is neither. The position is the one Jenny Odell, Nicholas Carr, and a number of careful thinkers have articulated in slightly different forms over the last fifteen years: tools change the people who use them. Sometimes the change is good. Sometimes the change is invisible until you notice the small thing it has removed. The honest response is not to abstain from the tools, and not to throw oneself into them, but to use them while keeping an eye on what they are doing.

The book the studio publishes — How to Be Bored Again — has a chapter on this specifically, and is the longer form of this argument. The chapter is titled “The AI Acceleration.” It is the chapter the writer of the book reports being the most worried about getting wrong, which seems, in 2026, like the right amount of worry to have.

What you are doing right now — reading a long, structured argument about a subject in your own time, without an AI summary of it — is itself an act of the kind of cognition this article is about. The point of the article is not that you should stop doing other things. The point is that this kind of cognition is worth keeping in your week.

If you do that, on most of the available evidence, you stay recognisably yourself in 2026. That is, on the whole, the win worth aiming for.