Does AI Weaken Your Thinking? What Studies Show and a Plan

You open a blank document to write an email or a report, and before you type a word your hand has already opened the AI window. You did not decide to; your hand got there first. Then you notice that the first paragraph, the one you used to wrestle with for ten minutes before it started to flow, no longer comes out of you alone.

So you wonder: am I losing something? And are the headlines saying AI "rots your brain" true? I opened the studies behind those headlines rather than the headlines themselves, so I can tell you what has been shown and what has not, and then give you a small plan that does not depend on willpower. Everything here is current as of October 2026, and each study is listed at the end.

What did the studies actually find? Five pictures, no final verdict

No single study settles this. Here are five views from different angles, each with its own limits.

1. The brain-and-essay study (MIT, 2025). Researchers at the MIT Media Lab had 54 adults write essays over three sessions: one group used a chatbot, one a search engine, one no tools at all, while EEG recorded their brain activity. Connectivity between brain regions was weakest in the chatbot group, those writers felt the least ownership of their essays, and they struggled to quote a line from what they had written. A fourth session involved only 18 of them, with the chatbot group and the no-tool group swapping conditions.

Do not build more on it than it can carry. The sample is small and drawn from one geographic area, and the paper is still a preprint on arXiv; I found no peer-reviewed version as of October 2026. The authors themselves call it preliminary and ask that it not be summed up with phrases like "brain rot." A scientific comment posted on arXiv also takes issue with the sample size, the EEG analysis and how consistently the results are reported.

2. The knowledge-worker survey (Microsoft and Carnegie Mellon, 2025). Researchers surveyed 319 people who use AI at work and collected 936 real examples. The more someone trusted the tool's ability, the less critical thinking they applied to its output; the more they trusted themselves on the task, the more they applied. People said they used some critical thinking in about 60% of the examples, but its nature had changed: their effort moved to checking and integrating what the tool produced, instead of gathering information and solving the problem themselves. The limits: everything is self-reported rather than measured, and the sample skews younger and more tech-savvy.

3. The correlation study (Gerlich, Societies, 2025). A survey and interviews with 666 people found a negative relationship between frequent use of AI tools and critical-thinking scores, running through "cognitive offloading," meaning handing mental tasks to the tool. But this is correlation, not causation: the tool may be the cause, people who think less may reach for it more, or something else may drive both.

4. The randomized experiments (several universities, 2026). This is the strongest evidence I found, because it separates cause from correlation. Researchers from Carnegie Mellon, MIT, Oxford, UCLA and Berkeley randomly assigned people in three online experiments, two on fraction problems and one on reading comprehension, each session lasting roughly 13 to 15 minutes. In all, 1,222 people were recruited and about 1,060 were analysed. One group had an AI assistant and one worked alone, and the assistant was removed before the last three test questions. Those who lost it solved fewer test questions in all three experiments: 57% versus 73% in the first, 71% versus 77% in the second, 76% versus 89% in the third. Skipping questions, the researchers' proxy for giving up, was significantly higher in two of the three.

This is the study behind the headlines saying ten minutes with AI erodes your persistence and your unassisted performance. The headlines are not wrong, but they leave out that the tasks were fractions and reading comprehension, that participants were US residents recruited through an online platform, and that the authors acknowledge they did not test daily use over months or years. Berkeley News reported that an updated, peer-reviewed version was presented at the Conference on Language Modeling in October 2026.

In the second experiment, people in the assistant group were asked how they had used it. 61% said they asked for the answer outright, and 27% asked for a hint or a clarification. The first group solved 65% of the test questions, the second 76%, and people who worked alone 77%. But that compares people's own choices, not randomly assigned groups, so we cannot tell whether the gap comes from how they used the tool or from who chose which way.

5. The student field experiment (PNAS, 2025). Nearly a thousand high-school students in Turkey practised maths in one of three ways: textbook and notes only, an unrestricted ChatGPT-style chatbot, or an AI "tutor" that gave hints instead of answers. During practice the chatbot group did about 48% better than the textbook group, but when the chatbot was taken away at the exam they scored roughly 17% below that group. The tutor group did about 127% better in practice and finished close to the textbook group on the exam. The measurement was short-term and in a single subject.

How to read these results without panicking

Ask four questions of any headline that scares you:

• Is it an experiment or a survey? A randomized experiment can say "cause"; a survey can only say "linked."
• How big is the sample? 54 people are not 1,222.
• Has it been peer-reviewed? A preprint's results can still change.
• How long did it last? Most of what was measured is minutes or weeks, not years.

With that filter, three results recur across more than one study. First, the harm shows up when you hand the tool the whole problem and take the answer ready-made, not when it acts like a tutor and gives a hint. Second, more trust in the tool goes with less critical thinking, at least by people's own account. Third, your performance with the tool really does improve, which tempts you to believe that your own ability improved.

What has not been shown is that years of daily use damage your ability to think. Nobody has measured that yet, and the authors of more than one of these studies say so. So there is no need to panic, and no need to flatter yourself either.

Cognitive offloading: when it serves you and when it eats your skill

"Cognitive offloading" is a name for something you do every day: you put an appointment in a calendar instead of memorising it, set an alarm, use a calculator. Researchers Evan Risko and Sam Gilbert (2016) define it as using a physical action in the outside world to lighten the mental load of a task. It is useful by default, because it frees your attention.

But they also note that people offload when they believe it is more efficient than relying on themselves, and that this belief is not always right. They point to possible costs too: skills that may fade when handed wholesale to GPS or autopilot, and weaker memory for things you photographed instead of looked at.

So the question is not "should I offload?" but "what should I offload?" Here is a rule that is my own design, not a finding: offload what you do not want to master, and keep what is your skill. Summarising a forty-page report to decide whether it is worth reading, formatting a table, translating a passing sentence: offload without hesitation. But the paragraph that carries your opinion, the argument you will defend in a meeting, the decision you will be held to: that is your craft, and if you hand it over every time, do not expect it to stay yours.

There is another trap: you will not feel the decline. In a study with researchers from Aalto University, 246 people solved logical-reasoning problems with AI help. Their scores rose by three points against a reference group, but they overestimated their own performance by four points. A second experiment (452 people) repeated the result. Your sense that you are "fine" is not a measure you can rely on, which is why you need a log.

Willpower will not protect your thinking: you need an anchor, not a decision

Why do we skip the thinking even when we mean to do it? In the Microsoft survey, participants repeatedly gave the same reasons for skipping critical thinking: the task seemed trivial, time was short, and their trust in the tool was high after earlier successes. Notice how they meet in one familiar moment, four in the afternoon with a deadline close, which is exactly when morning resolutions do not help.

That is the idea the book Istiqala min Nafsik ("Resigning from Yourself") builds in its second chapter: willpower is not the hero of the story. People who deliver consistently usually do not have superhuman resolve; they have a fixed moment in the day where the action begins without negotiation. The book sums it up in one line: "You don't need more enthusiasm. You need a smaller action and a fixed cue."

The wider research points the same way. A 2006 review by Gollwitzer and Sheeran pooled 94 independent tests and found that "if X happens, I will do Y" plans raise goal attainment with a medium-to-large effect (d = 0.65). That is not a claim that your success will rise by 65%; it is an average standardised difference, and it was not tested on AI specifically. But its logic applies: instead of deciding each time whether to think first, make the decision in advance and tie it to a specific moment.

The two-minute habit: think first, then ask

The fixed moment is clear: the instant you reach for the tool on a task that matters. Make that your anchor. This plan is my own design, applying the ideas above (a fixed cue, starting from yourself, a hint instead of a solution), and it has not been tested as a package in any study.

Why only two minutes? Because a small action is easier to repeat, and because the goal is not a good draft but for your mind to start before the tool does. I see a second benefit: when the reply arrives after you have written your position, you have a reference to compare it with, instead of accepting the first wording you read.

A one-week log: is your independent thinking shrinking?

How do you know the plan is working? Do not rely on your gut; you have just seen that feeling competent can run ahead of being competent. What is not measured slips away quietly, as the book says in its fourth chapter on the notebook we never open. And in a 2016 review by Harkin and colleagues covering 138 experimental studies and nearly 20,000 participants, goal attainment rose (d+ = 0.40) among people prompted to monitor their progress, with a larger effect when progress was recorded in writing or reported to others.

Take just one week, and each evening write three numbers on a single line, in a notebook or a note:

1. How many mental or writing tasks did I use the tool for today?
2. In how many of them did I start with my own two minutes?
3. How many times did I change or reject something in its reply, whether a number, an argument or a sentence?

The second number divided by the first is the share of times you started yourself. The third tells you whether you read the reply or just copy it.

At the end of the week, add a small test inspired by the quoting measure in the MIT study; it is a self-observation tool, not a validated test. Close everything, pick a text you sent this week with the tool's help, and recall its three main points and one sentence from it. If you cannot, the text is not yet yours.

Read the numbers without beating yourself up. If you started on your own in 3 tasks out of 7, the anchor is still weak; look at where it sits in your day and prepare the paper and the place for it, rather than blaming your willpower. If the share shrinks week after week, that is a signal worth stopping for. And show the page to a friend or colleague at the end of the week; as you saw, the effect was larger when someone else knew about the progress.

The takeaway: do not drop the tool, and do not hand it the part that is you

AI does not weaken your thinking just by existing. The current evidence suggests the effect depends on what you hand over: give it the whole solution every time and you lose the practice, or use it as a tutor and reviewer and your mind stays in the ring. What has not been shown is the effect of years of use, so it is wise to be careful without panicking.

Start tonight with something very small: choose your anchor, try the two minutes on one task, and write it down. And if you would like to practise building habits like this away from the screen, Istiqala min Nafsik has stories and exercises that build the same idea step by step.