How to use AI without losing your own thinking

The Risk Is Not Wrong Answers

The usual worry about AI at work is that it will be wrong. The more common problem is that it is right enough, and that you stop doing the part of the thinking that was yours. The change is small at each step: you ask for a first draft instead of making one, you accept a summary instead of reading the source, you pick the best of three options the tool gave you instead of the one you would have come up with. The result looks better than what you would have produced alone, so nothing tells you that something has changed.

The research I found points the same way on three things. When people trust the tool more, they report thinking critically less.^[1] People who practised with an unrestricted AI did better while they had it and worse once it was taken away.^[2] And in one large experiment, AI ideas made each person's story better and the stories more alike.^[9] None of this says AI is bad for thinking. It says the effect depends on how you use it.

This article covers what I ran into, what the studies show and how strong each one is, and instructions built from those studies to help the tool extend your thinking and not replace it.

What I Ran Into

I use AI every day and I have an M.Sc in AI. Two things were hard, and neither was about facts.

It was hard to make it learn my voice. The output kept coming back smooth, balanced and generic, and not in my voice.

It pushed back too much when I tried something innovative. It steered me back to the common answer before the idea had room to develop.

I want to be careful about what the research says on this second point, because it is easy to overstate. I found no study showing that AI models reject unusual ideas. What the studies show is two separate things. Models tend to agree with the user, which is called sycophancy.^[14][15] And models trained with human feedback produce a narrower range of answers than models fine-tuned without it.^[13] My experience was the second one, and the pull toward the typical answer is what felt like pushback. That is a hypothesis about my own case, not a finding. It suggests giving the tool a defined job at each stage.

Thinking Less When the Tool Is There

Using something outside your head to save mental effort is normal. Notes, calendars and calculators do it, and researchers call it cognitive offloading.^[5] The question is what you give up when the thing you offload is the thinking itself.

The strongest evidence is a field experiment with nearly 1,000 high school maths students.^[2] Students who practised with a plain version of ChatGPT solved 48% more problems correctly during practice. On an exam without AI they scored 17% lower than students who never had it. A tutor version, built to give hints and not answers, largely removed that drop. The tool made practice feel productive and the learning thinner, and the version that held back was the one that protected the learning.

A survey of 319 knowledge workers found a similar link.^[1] The more confident people were in the AI, the less critical thinking they reported, and the more confident they were in their own skill, the more they reported. It is self-reported and shows an association, not a cause. Two other studies get more attention than their strength supports: a survey of 666 people linking AI use to weaker critical thinking, which is correlational and carries a published table correction,^[3] and an MIT brain-scan study, which is a preprint with 54 people in its first three sessions and 18 in its fourth.^[4] I would read both as signals worth watching, not as findings.

Trusting the Tool When It Is Wrong

Decades of research on automation, from aviation to medicine, found the same pattern: people given a good but imperfect aid miss what it does not flag and follow it when it is wrong. A major review concluded that this happens to experts and teams alike and is not prevented by training or instructions.^[6] That last point matters: telling your team to "double-check the AI" is not a control.

The AI-era version comes from a preregistered experiment with 758 consultants at BCG.^[7] On tasks inside what the AI did well, consultants using it completed 12.2% more tasks and finished 25.1% faster. On a task designed to sit just outside its ability, they were 19% less likely to reach a correct answer than consultants without it. The loss came on the one task designed to sit outside what the AI did well.

A working paper on recruiters, not yet peer reviewed, points the same way.^[8] Recruiters given a 75% accurate AI judged CVs better than those given an 85% accurate one, apparently because the weaker tool kept them alert. A related finding from two studies on logic problems: people using AI scored about 3 points higher than a norm group but overestimated their own score by about 4.^[21] In that study, the people who knew the most about AI were the most overconfident.

Everyone's Ideas Start to Look Alike

The most useful study for anyone who needs original ideas used 293 writers asked to write short stories.^[9] Writers who could see AI-suggested ideas wrote stories rated more novel, by 5.4% with one idea and 8.1% with five, and the least creative writers gained the most. But the stories also became more similar to each other, by about 10.7% with one idea and 8.9% with five. Each person gained, and the group lost variety.

Two more studies add detail. A small one found ChatGPT users produced more ideas, but ideas less distinct from one another's, and felt less responsible for them.^[11] A larger study of over 800 people found AI ideas did not make any one person more creative, and summarised it as "AI made ideas different, not better".^[12] Whether AI widens or narrows a group's range depends on the design, which the next two sections cover.

The effect also reaches past the moment of use. In two preregistered experiments with 1,100 people, some kinds of AI help left people worse off once it was removed: a helper that supplied strategies lowered later originality, and a coaching style of guidance left people scoring 10.6 percentage points lower than people with no AI on a reasoning task. Direct answers did not.^[10] The format of the help changes what it does to you. On the model side, training a language model with human feedback was shown to substantially reduce the diversity of its outputs.^[13] Both the model and the people using it contribute to the sameness.

There is an older effect that may explain part of it. Give people an arbitrary number first and their estimates move toward it, even when they know the number came from a spinning wheel: groups anchored on 10 and 65 gave median estimates of 25 and 45.^[16] The first answer an AI gives you can act as an anchor for what you think next. I am applying that finding to AI by inference, and it has not been tested there directly.

What Protects Your Thinking

The research on what helps points to friction placed in the right spot.

• Produce first. Information you generate yourself is remembered better than information you read, across five classic experiments.^[17] Practice that feels easy can hide how little was learned, and some difficulty is what makes learning last.^[18]
• Commit before you see the answer. In an experiment with 199 people, designs that made people decide first and then see the AI's advice reduced over-reliance on wrong advice. People also liked those designs the least.^[19] In that study the designs that worked were the least liked, which is a reason to build them into the process and not leave them to choice.
• Make checking cheap. Explanations reduced over-reliance only when they made it cheaper to check the AI's answer.^[20] A long explanation that is hard to verify does not help.
• Be accountable for how you used it. In an older study with students and commercial pilots, people told they would answer for how they used the automation verified it more and made fewer errors.^[22] It is a conference abstract, so I treat it as a pointer.
• Judge the work done without the tool. The maths study and one creativity study show the cost after the tool is removed.^[2][10] If you only measure output with the tool, you will not see it.

One gap. I found no peer-reviewed study comparing "write your own ideas first, then ask AI to extend them" with "ask AI first". The recommendation below is built from the studies above. It is not a tested finding.

Instructions That Keep You Yourself

My view is that many of the problems above come from the tool being given the wrong job at the wrong time, and that the job can be written into the instructions you give it, once, and reused. These instructions are built from the studies above. They are not tested as a set, and the wording is an example to adapt to your own work.

Instruction	What it is meant to counter
Voice file. Give it three or four things you wrote and approved, and say: "Match the sentence length, word choice and structure of these. If you are not sure it sounds like me, say so."	The competent stranger. It also gives the tool something to be wrong against, which makes the drift visible.
My draft is the base. "When I give you a draft, extend it in my voice. Do not summarise it, reorder it or replace it."	Your thinking being swapped for the tool's version while you stay the author on paper.
Two modes, named. "Build mode: assume my idea is good and develop it. Challenge mode: find its weaknesses. Stay in the mode I name."	Criticism arriving before the idea has had room to grow.
Ask me first. "Before you answer, ask me what I think and why. Do not show your answer until I reply."	Anchoring, which has not been tested on AI, and over-reliance.[16][19]
Steelman, then attack. "State the strongest version of my idea in two sentences. Then give the three best objections, each with what would have to be true for it to be wrong."	Sycophancy on one side and a reflex to the average on the other. You get an honest case for and against.
Novelty guard. "Name what is common about this answer. Then give one option that a typical answer would not include, and say what it risks."	Sameness. It makes the tool point out when it is giving you the most typical thing.
Show the check. "For any claim I may act on, give the source and the one thing I can check in under a minute."	Automation bias. It makes checking cheap, which is when checking happens.[20]
Say what you do not know. "If the question is outside what you handle well, say so before answering."	The 19% drop: errors on the hard case, delivered with confidence.[7]

Two of these matter more than the rest. The first is naming the mode. The tool criticised my ideas before they had room to develop, and naming the mode addresses that directly. The second is asking first, which the one experiment on it supports.^[19]

For a team, the same principle scales: put the voice file, the modes and the checks in the shared instructions, not in each person's habits. People will not choose friction every time, so it has to be the default.

A Five-Minute Self-Test

Use these on your own work this week.

1. Pick something you finished with AI last week. Can you explain every choice in it without opening the chat?
2. Did you write down what you thought before you asked? If not, how much of the result is yours?
3. Would the same prompt from a competitor produce the same output? If yes, what in it is only you?
4. When did the tool last tell you it was unsure, or that you were wrong, and you changed your mind for a good reason?
5. Could you do this task without it tomorrow, at the same standard? If you are not sure, do one such task a week without it.

Where This Does Not Fit

• Most of the studies use short tasks and online samples, and several test models that are now a generation old. Sycophancy in particular is being trained against, so current models may behave differently.
• Routine work does not need this. For a standard report, a translation or a summary of a document you will read anyway, AI help is a fair trade. The risk is in the work where your judgement or your difference is the product.
• Several of the findings are self-reported or correlational. They show a pattern worth testing in your own team, not proof of harm.
• The instructions above are built from the studies, within their stated limits. I have not run a controlled test of them.

Sources

1. Lee, Sarkar, Tankelevitch, Drosos, Rintel, Banks, Wilson, The Impact of Generative AI on Critical Thinking, CHI 2025 (Microsoft Research and Carnegie Mellon). Source - A survey of 319 knowledge workers: higher confidence in the AI went with less critical thinking, higher confidence in their own skill went with more. Self-reported, an association.
2. Bastani et al., Generative AI without guardrails can harm learning: Evidence from high school mathematics, PNAS 122(26), 2025. Source - A randomised field experiment with nearly 1,000 students. Plain ChatGPT: 48% better in practice, 17% worse on an exam without AI. A tutor version with guardrails largely removed the drop.
3. Gerlich, AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking, Societies 15(1), 2025. Source - A survey of 666 people found a negative link between frequent AI use and critical thinking. Correlation only; a table correction was issued in September 2025, so no coefficient is quoted here.
4. Kosmyna et al., Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task, arXiv 2506.08872, MIT Media Lab, 2025. Source - A preprint, not peer reviewed. 54 people in the first three sessions and 18 in the fourth. The ChatGPT group had the weakest brain connectivity and found it hardest to quote its own essays.
5. Risko and Gilbert, Cognitive Offloading, Trends in Cognitive Sciences 20(9), 2016. Source - A review that defines cognitive offloading: using actions in the world, such as notes or tools, to reduce the demand on your own thinking.
6. Parasuraman and Manzey, Complacency and Bias in Human Use of Automation: An Attentional Integration, Human Factors 52(3), 2010. Source - A review of automation bias: people miss what an imperfect aid does not flag and follow it when it is wrong. It occurs in experts and teams and is not prevented by training or instructions.
7. Dell'Acqua et al., Navigating the Jagged Technological Frontier, Organization Science, 2026 (first published as HBS Working Paper 24-013). Source - A preregistered field experiment with 758 BCG consultants. Inside AI's ability: 12.2% more tasks, 25.1% faster. On a task just outside it: 19% less likely to give a correct answer.
8. Dell'Acqua, Falling Asleep at the Wheel: Human/AI Collaboration in a Field Experiment on HR Recruiters, working paper, 2022. Source - A working paper with 181 recruiters. Those given a 75% accurate AI did better than those given an 85% accurate one, and spent more time and verified more. Not peer reviewed.
9. Doshi and Hauser, Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances 10(28), 2024. Source - 293 writers. One AI idea raised rated novelty by 5.4% and five by 8.1%. The stories also became more similar to each other, by about 10.7% (one idea) and 8.9% (five).
10. Kumar, Vincentius, Jordan, Anderson, Human Creativity in the Age of LLMs: Randomized Experiments on Divergent and Convergent Thinking, CHI 2025. Source - 1,100 people in two preregistered experiments. After some kinds of AI help, people did worse on their own: a strategy-list helper lowered later originality, and coach-style guidance cut a reasoning score by 10.6 points. Direct answers did not.
11. Anderson, Shah, Kreminski, Homogenization Effects of Large Language Models on Human Creative Ideation, ACM Creativity and Cognition 2024. Source - A small study (36 people). ChatGPT users produced more ideas, but ideas less distinct from one another's, and felt less responsible for them.
12. Ashkinaze, Mendelsohn, Qiwei, Budak, Gilbert, How AI Ideas Affect the Creativity, Diversity, and Evolution of Human Ideas, ACM Collective Intelligence 2025. Source - Over 800 people. Exposure to AI ideas did not change individual creativity; high exposure increased the diversity of the group's ideas. The authors' summary: AI made ideas different, not better.
13. Kirk et al., Understanding the Effects of RLHF on LLM Generalisation and Diversity, ICLR 2024. Source - Training a model with human feedback (RLHF) helps it handle new inputs but substantially reduces the diversity of its outputs compared with models fine-tuned without it. A result about models, not people.
14. Sharma et al., Towards Understanding Sycophancy in Language Models, ICLR 2024 (Anthropic). Source - Five assistants showed sycophancy: they tended to match the user's views. Human raters and preference models sometimes preferred a convincing agreeable answer to a correct one. Models tested in 2023.
15. Perez et al., Discovering Language Model Behaviors with Model-Written Evaluations, Findings of ACL 2023. Source - Larger models more often repeat back the answer a user seems to prefer, and more RLHF training sometimes made this worse. 2022 models.
16. Tversky and Kahneman, Judgment under Uncertainty: Heuristics and Biases, Science 185(4157), 1974. Source - Groups given a starting number of 10 or 65, from a spinning wheel, gave median estimates of 25 and 45. Paying for accuracy did not remove the pull.
17. Slamecka and Graf, The generation effect: Delineation of a phenomenon, Journal of Experimental Psychology: Human Learning and Memory 4(6), 1978. Source - Five experiments: information you generate yourself is remembered better than information you read. Word lists, not complex work.
18. Bjork and Bjork, Making Things Hard on Yourself, But in a Good Way: Creating Desirable Difficulties to Enhance Learning, 2011. Source - Performance during practice can be an unreliable index of learning; some difficulties help long-term learning. Learning research, not AI research.
19. Bucinca, Malaya, Gajos, To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making, CSCW 2021. Source - 199 people. Designs that made people commit to their own judgement before seeing the AI answer reduced over-reliance on wrong advice, and were liked the least.
20. Vasconcelos et al., Explanations Can Reduce Overreliance on AI Systems During Decision-Making, CSCW 2023. Source - Five studies, 731 people, a toy task. Explanations reduced over-reliance only when they made checking the AI's answer cheaper.
21. Fernandes et al., AI makes you smarter but none the wiser: The disconnect between performance and metacognition, Computers in Human Behavior 175, 2026. Source - Two studies (246 and 452 people) on logic problems. With AI people scored about 3 points higher but overestimated their score by about 4, and those who knew more about AI were the most overconfident.
22. Mosier, Skitka, Burdick, Heers, Rosekind, Decision Making in a High-Tech World: Automation Bias and Countermeasures, 1996 (NASA record). Source - Students and commercial pilots. People told they would be accountable for how they used the automation verified it more and made fewer automation-related errors. A conference abstract.

Related Reading

• The AI Tax - what AI adds to the cost of the work, not only what it removes.
• AI and the Fractional Operator - how AI changes the work of interim and fractional roles.
• The Deliberate Slow - which processes should stay slow when AI makes everything faster.