Is AI making us dumber? What the research on AI and critical thinking means for your doctorate
You've probably seen the headlines: an MIT study says ChatGPT is bad for your brain. If you've used AI most days during your doctorate, you might be wondering what it has done to your thinking, and whether your supervisor or examiners will notice.
The research gives a clearer answer than the headlines do. Here's what it says, where it's weak, and what it means for how you use AI in your research.
The short answer
It depends on how you use AI. In the studies so far, people who let AI do their thinking tend to do worse once the tool is taken away. People who use AI to question, explain and check their own thinking tend to keep their skills, and some studies find they improve them.
One caveat before we go further. Most of this research is on school pupils, undergraduates and office workers. Very few studies look at doctoral researchers, so treat the findings as a strong signal and expect the picture to sharpen as more studies come out.
What the MIT "Your brain on ChatGPT" study found
This is the study behind most of the headlines. Nataliya Kosmyna and colleagues at MIT asked 54 people to write essays over three sessions. One group used ChatGPT, one used a search engine, and one used no tools at all. While they wrote, each person wore an EEG cap that recorded their brain activity (Kosmyna et al., 2025).
The group that wrote with no tools showed the strongest and most widespread brain connectivity. The ChatGPT group showed the weakest. The ChatGPT group also felt the least ownership of their essays, and they struggled to quote accurately from essays they had just written. In a fourth session, 18 of the participants swapped groups. People who moved from ChatGPT to writing on their own showed lower brain engagement than the group that had written unaided from the start.
Before you panic, there are three things to know about this study.
- It hasn't been peer reviewed. At the time of writing (September 2026), it's a preprint, a paper shared publicly before journal review.
- It's small. 54 people took part, and only 18 completed the fourth session.
- Other researchers have raised concerns. A published comment questions the sample size, the EEG analysis and how some of the results were reported (Stanković et al., 2025).
The study also tested one way of using AI: asking it to help write the essay. It didn't test using AI to question your argument, check your reasoning or quiz you on what you've read. So it shows what can happen when you hand over the writing, and it says nothing about those other ways of working with AI.
The key finding is that people couldn't quote their own essays. If you can't quote what you wrote, you'll struggle to defend it, and defending your work is exactly what your viva asks of you.
What other studies on AI and critical thinking say
The MIT study isn't the only one. Here are five more, and what each found.
A maths experiment with nearly 1,000 secondary school pupils (Bastani et al., 2025). Pupils practised with one of two AI tutors: one that worked like standard ChatGPT, and one set up to give hints and guidance. With either tutor, their practice grades went up. When the AI was taken away, pupils who had used the standard version scored 17% lower than pupils who had never had AI at all. The authors describe those pupils as using the AI as a "crutch". The tutor designed to give hints largely avoided that drop.
A survey of 319 knowledge workers at Microsoft Research (Lee et al., 2025). People described 936 real examples of using AI at work. The more confidence they had in the AI, the less critical thinking they reported. The more confidence they had in their own ability, the more critical thinking they reported. Their critical thinking also changed shape. It moved away from doing the task and towards checking information, combining AI responses and overseeing the work. These are self-reports, so they tell us how people saw their own thinking.
A survey of 66 people in the UK (Gerlich, 2025). People who used AI tools more often had noticeably lower critical thinking scores, and handing mental work to the tools explained part of that link. This is a correlation. It can't tell us if AI was the cause, i.e., whether AI use lowered their critical thinking, or whether people who think less critically lean on AI more.
An eight-week study of 168 undergraduates writing in English (Chen, 2026). One group wrote without AI, one used AI within set limits and had to reflect on how they used it, and one used AI freely. The free-use group produced the best work while they had the tool. At the end of the eight weeks, everyone wrote without AI. The limited-use group then scored a little higher than the free-use group on writing quality, higher-order thinking and depth of argument. The differences were small and only just at the edge of statistical significance, and the author says they show an association and can't prove cause.
A survey of 383 university students in Italy (Zuin and Donadel, 2026). Students with a stronger habit of critical thinking were less likely to use AI output without changing it. When they were suffering from decision fatigue, that protection weakened. The effect was clearest when students used AI to revise their writing or to give them feedback. If you've ever accepted an AI suggestion at 10pm because you were too tired to argue with it, this finding will feel familiar.
The wider picture. A 2026 systematic review looked at 63 studies of AI and students' critical thinking (Balart et al., 2026). Almost half of them (47%) found a significant improvement in critical thinking, mostly where AI use was structured with guided prompts, reflection and assessment of the student's own performance. The risks showed up mainly where use was unguided: accepting AI output without question, and less intellectual independence.
Put together, the pattern is consistent. When AI does the thinking, people do worse once the tool is gone. When AI is used to push people's own thinking, most of that loss disappears, and some studies find gains.
What is cognitive offloading?
Cognitive offloading means handing mental work to something outside your head. You do it every time you write a shopping list, set a reminder or let reference software format your bibliography. Most of it is sensible. Nobody needs to memorise a reference style.
What decides whether offloading helps you is what you hand over. A 2026 study of 589 students and early-career professionals, some of them doctoral students, separated two kinds of offloading to AI (Zhu et al., 2026). In "dependent" offloading, people accept AI answers with little checking and let AI structure their reasoning. In "autonomous" offloading, people use AI to spark their own ideas and compare its output with their own reasoning.
Here's the part to remember. At the time, both kinds felt equally helpful. The difference only showed up later, when people who relied on dependent offloading reported weaker thinking, creativity and independent judgement. All of this was self-reported, so the study measures how people saw their own thinking.
The authors concluded: "These findings highlight the manner of AI engagement—not merely its frequency—as a key factor in understanding its associations with perceived cognitive functioning, and point to practical strategies for educators, learners, and AI tool designers seeking to harness AI without undermining cognitive autonomy."
Why it's a bigger deal in a doctorate
A doctorate is an apprenticeship in judgement. You're learning to decide which questions are worth asking, which methods fit, and which results are real. In the book, we call this taste, and we separate it from technique, the practical skills of searching, coding and drafting. AI is getting very good at technique. It can't build your judgement for you, and your viva is where your judgement gets tested.
In our book, we tell the story of Nadia, a doctoral researcher who used AI sensibly for two years to write the code for her analysis. Two months before submission, a journal asked for one extra analysis. She expected it to take an afternoon. It took a week, because she could no longer follow her own code well enough to change it safely. She hadn't done anything wrong. She had just stopped checking, one reasonable decision at a time.
How to use AI without weakening your thinking
These five habits come from the book, and the research above supports each one.
1. Have a go first, then ask
Try the problem, draft the paragraph or explain the concept yourself before you ask AI. In a large pre-registered experiment, people who attempted maths problems themselves before reading an AI explanation learned the most (Kumar et al., 2023). It's a working paper, not yet peer reviewed, but it fits the other studies. You can go a step further: explain the idea to the AI in your own words and ask it to question you until it finds the gap in your understanding.
2. Ask AI to argue with you
Ask it to make the strongest case against your interpretation, or to find the weakest point in your argument. Then do the rethinking and the rewriting yourself.
3. Check it against something outside the answer
Reading an AI answer and deciding it looks right only catches the mistakes you already know how to spot. Check the answer against something outside it instead: open the source it cites, or run your code on a test case where you already know the answer.
4. Work without AI regularly
Write a section, rebuild an analysis or reconstruct a paper's argument with the tool closed. Then ask yourself: could I explain, use or defend this idea without AI? If you can, using AI is a choice. If you can't, it has become a dependency.
5. Set your rules in advance
Decide with your supervisor what AI will and won't do at your stage of development, and write it down. That way, the decision is already made on the tired afternoon when accepting the output without critically thinking about it looks reasonable. The Italian study above shows why this helps: tiredness weakens even strong critical thinking.
Frequently asked questions
Does AI reduce critical thinking?
It can, depending on how you use it. Studies find that people who let AI do their thinking perform worse without it, while people who use AI to question and check their own thinking tend to keep their skills. Most studies so far involve pupils, undergraduates and office workers.
What is cognitive offloading?
Cognitive offloading means handing mental work to a tool, such as a calendar, a calculator or an AI. It becomes a problem when you hand over the thinking itself, for example by accepting AI answers without checking them.
How reliable is the MIT "Your brain on ChatGPT" study?
The MIT study is a useful early signal with real limits. At the time of writing, it hasn't been peer-reviewed; it involved 54 people, and other researchers have questioned its methods. It also tested only one way of using AI: help with writing the essay.
Can AI improve critical thinking?
Yes, in the right conditions. A 2026 review of 63 studies found that almost half reported significant gains in critical thinking, mostly where AI use was structured, with guided prompts and reflection.
Want to go further?
This blog post sheds light on some useful habits. Our book, Still Yours: A Doctoral Researcher's Guide to AI, gives you the decision-making behind them. Michael Rowe and I set out four questions to ask before you hand any task to AI, show you how to write rules for your stage with your supervisor, and explain how to use AI as a thinking partner that questions you.
If you're using AI in your research right now, these posts may also help: How to use AI for a literature review without outsourcing your thinking, How to declare AI use in your thesis and How to cite ChatGPT in your thesis.
The book is coming soon from Springer Nature.
The views in this post are my own and don't necessarily reflect those of my employers.
0 comments
Leave a comment
Please log in or register to post a comment