
Does Using AI Make You a Worse Researcher? What Studies Say
By Samuel de Bruyn 10 min read
Using AI does not automatically make you a worse researcher, but the evidence so far says it does when the tool does the thinking for you. In MIT Media Lab’s 2025 “Your Brain on ChatGPT” preprint, 83.3% of the ChatGPT group (15 of 18) could not correctly quote the essay they had just written, and in a field experiment published in PNAS in 2025, students who practised maths with unrestricted GPT-4 scored 17% lower than students who never had access once the tool was taken away.
The same studies point to the fix. Students who used a version of GPT-4 built to give hints instead of answers avoided most of that drop, and MIT participants who wrote on their own first and only then used ChatGPT showed higher memory recall in the fourth session. The line that matters is not “AI or no AI”. It is whether you still do the reading, the judging and the writing.
What is cognitive debt? Cognitive debt is the term the MIT Media Lab team used for the cost that builds up when you hand a thinking task to an AI assistant: the task gets done, but you remember less of it, feel less ownership of it, and engage less of your own reasoning in doing it. The researchers measured it in a 20-minute essay task, not in long-term research skills.
What do the main studies on AI and thinking skills actually show?
Four studies get cited most often in the “is AI making students worse?” debate, and they measure different things. Only one is a field experiment on learning; the others are a small lab study with EEG and two surveys.
| Study | Who took part | Design | Main finding | Main caveat |
|---|---|---|---|---|
| Kosmyna et al., “Your Brain on ChatGPT” (MIT Media Lab, arXiv preprint, June 2025) | 54 people aged 18 to 39 from five Boston-area universities | Three groups (ChatGPT, search engine, no tools) wrote essays over four sessions while EEG recorded brain activity | 83.3% of the ChatGPT group (15 of 18) could not correctly quote their own essay in session 1, against 11.1% (2 of 18) in each of the other groups | Preprint, small sample from one area, one model (GPT-4o), one task |
| Bastani et al., “Generative AI without guardrails can harm learning” (PNAS, vol. 122, 2025) | Nearly a thousand high school maths students | Field experiment: unrestricted GPT-4 (“GPT Base”), a hint-giving tutor version (“GPT Tutor”), or no access | Practice grades rose 48% with GPT Base and 127% with GPT Tutor; once access was removed, GPT Base students scored 17% lower than students who never had it, while GPT Tutor largely avoided the drop | School maths, not postgraduate research |
| Lee et al., “The Impact of Generative AI on Critical Thinking” (Microsoft Research, CHI 2025) | 319 knowledge workers, 936 real examples | Survey of self-reported critical thinking | Higher confidence in GenAI went with less critical thinking; higher self-confidence went with more | Self-reported, correlational |
| Gerlich, “AI Tools in Society” (Societies, vol. 15, January 2025) | 666 UK participants, plus 50 interviews | Survey and interviews | Frequent AI tool use correlated with lower critical thinking scores (r = -0.68), with cognitive offloading as the mediator | Correlational, so it cannot show that AI caused the lower scores |
Read together, the four studies agree on one pattern: the risk sits in offloading the thinking, not in using the tool. None of them tested postgraduate students doing a literature review or a thesis, so any claim that AI “damages researchers” goes further than the evidence.
What did the MIT “Your Brain on ChatGPT” study find?
The MIT Media Lab study found that people writing essays with ChatGPT showed the weakest brain connectivity of three groups, remembered their own essays worst, and felt the least ownership of them. Brain connectivity “systematically scaled down with the amount of external support”, in the authors’ words: strongest with no tools, moderate with a search engine, weakest with ChatGPT.
The design is worth knowing before you quote the MIT study. Nataliya Kosmyna, Pattie Maes and colleagues split 54 participants (mean age 22.9) from MIT, Wellesley, Harvard, Tufts and Northeastern into three groups of 18. Each person wrote SAT-style essays in 20 minutes per session, across three sessions over four months. The ChatGPT group used OpenAI’s GPT-4o as its only source.
The most useful finding for researchers is in the fourth session, which 18 participants completed. People who had used ChatGPT and then wrote without it showed weaker alpha and beta connectivity. People who had written without tools and then used ChatGPT showed higher memory recall. The MIT results suggest the order matters: thinking first and bringing AI in second looked better than starting with AI.
What are the limits of the MIT cognitive debt study?
The MIT cognitive debt study is a preprint, and the authors themselves call the findings preliminary. The project page still describes the paper as a preprint (as of November 2026), and the authors ask readers to treat the conclusions “with caution”.
The authors list the limits plainly on the MIT Media Lab project page: a small group of participants from one geographic area, only ChatGPT tested, only essay writing, and an essay task that was not split into stages such as idea generation and writing. The MIT team also asks journalists not to describe the study with words like “brain rot”, “dumb” or “damage”, because the paper uses none of them.
So the MIT study supports a narrow claim: writing a short essay with ChatGPT, from the first word, led to weaker recall and ownership of that essay. The MIT study does not show that AI use causes lasting harm, and the MIT study says nothing about reading papers, searching databases or checking citations.
Does AI make students worse at learning, or only at the task?
The strongest evidence on learning comes from Bastani and colleagues’ PNAS field experiment, and the answer depends on how the AI tool was set up. Students with unrestricted GPT-4 did far better during practice (48% higher grades) and then worse on their own (17% lower than students who never had access). Students whose GPT-4 was prompted to act as a tutor did even better in practice (127% higher) and largely avoided the later drop.
The PNAS authors describe the unrestricted group as using GPT-4 as a “crutch”. For a research student, the crutch is accepting an AI tool’s account of what your papers conclude; the tutor version is asking where a claim is made, then reading that passage.
The Microsoft Research survey adds a second point. Lee and colleagues found that generative AI shifts critical thinking toward “information verification, response integration, and task stewardship”. In plain terms, the job changes from producing answers to checking them, and people who trust the AI more check less. A research student who uses AI and stops verifying is the case the Microsoft survey describes.
Which research tasks can you hand to AI, and which should stay with you?
Universities that publish thesis-level AI rules mostly agree on where the line sits: spelling and grammar help is fine, search and reading support is a grey area that usually needs your department’s permission and a declaration, and AI-written text or AI-made analysis in the submitted thesis is not allowed. The rules differ by country and by faculty, so the table below names each one.
| Task | University of Bristol (UK, research degrees) | UCT Faculty of Commerce (South Africa, dissertations, Dec 2025) | University of Melbourne (Australia, graduate research) |
|---|---|---|---|
| Spelling and grammar checks | Allowed | Always allowed | Allowed (grammar, clarity and spelling) |
| Search terms and reading plans | Not addressed on the thesis page | Allowed with disclosure, if the department authorises it | Not addressed on the thesis page |
| AI summaries of sources | Not as thesis text | Summary labels as starting points, with disclosure; AI summaries or citations pasted into the dissertation are not allowed | Not as thesis text |
| Drafting or redrafting thesis text | Not allowed: AI may not write any text used in the thesis | Not allowed for sections that add claims or synthesis | Not allowed: no initial drafting or substantial redrafting |
| Data analysis | Code help allowed for non-examined technical parts, with acknowledgement | Exploratory summaries with disclosure, if authorised; outsourcing analysis or interpretation is not allowed | Not addressed on the thesis page |
| Declaring use | Acknowledge translation, code and diagram help | Declare all AI use in front matter and keep a record of prompts and outputs | Acknowledge use and keep a record of tools, dates, prompts and outputs |
In Canada, the University of Toronto’s School of Graduate Studies requires prior approval from the supervisor and supervisory committee for any generative AI use in a doctoral thesis, and expects the student to be able to describe and defend any use of generative AI at the final oral exam. Toronto’s School of Graduate Studies does not regulate master’s theses, so master’s students there follow their own department. For a stage-by-stage view across more universities, including US graduate schools, see Can I Use AI for My Dissertation? What’s Allowed by Stage.
How do you use AI for research without losing your skills?
To use AI for research without losing your skills, keep three jobs for yourself: deciding what the papers say, judging whether they are any good, and writing the argument. Hand AI the work of finding, sorting and checking, and verify every answer against the source. That is the pattern the MIT and PNAS results favour, and it is the line Bristol, UCT and Melbourne draw in their rules.
Six habits make the line concrete for a thesis or a literature review.
- Think first, then ask. Write down what you expect a paper or a set of papers to show before you ask an AI tool anything. In the MIT study, participants who wrote on their own for three sessions before switching to ChatGPT showed higher memory recall.
- Ask for locations, not conclusions. “Where do these papers report sample sizes?” keeps you reading. “What do these papers conclude?” invites the crutch effect the PNAS authors saw.
- Open every passage you cite. If you have not read the sentence a citation points to, you do not know whether it supports your claim. This also catches invented references, which the University of Toronto names as a known AI risk.
- Write your own prose. Bristol and Melbourne both rule out AI-written thesis text, and UCT Commerce rules out AI drafting of any section that adds claims or synthesis.
- Run the viva test. For each paper you cite, could you explain its method and its main limitation to an examiner without notes? If not, read it again.
- Keep a log as you go. UCT Commerce and Melbourne both require a record of prompts and outputs, and a log makes your declaration easy to write. The AI disclosure statement generator turns that record into a declaration in your university’s format.
How do I use an AI assistant for writing backed by sources without handing over the thinking?
Litrevu is an AI assistant for writing backed by sources. It reads every paper, report and PDF you add, answers your questions with a citation to the exact passage behind each claim, compares sources side by side, and checks your own draft against them.
The Litrevu workflow keeps the same split between finding and thinking: Litrevu finds and checks the evidence, and the student reads, judges and writes. You add your own papers as PDFs, PDF links, DOI or arXiv links, or pasted text, or pick open-access papers Litrevu finds on your topic. You ask questions of those papers, and every answer cites only them; when your papers do not cover a question, Litrevu says so. You open any citation to see the highlighted passage on the page of the PDF, and you read it. You line up studies side by side in a cited table to see where they agree and differ. Then you paste a paragraph you wrote, and each claim comes back labelled Supported, Partly supported, Contradicted or Not in your papers.
Here is the shape of a claim check, shown as a format illustration with placeholder sources rather than a real output:
Your sentence: “Most studies of the intervention report a positive effect on retention.”
Supported (Author, Year): passage on page X, highlighted.
Contradicted (Author, Year): passage on page Y reports no effect.
Not in your papers: no paper in your library measures retention at follow-up.
You still decide what that result means. “Not in your papers” means your library does not cover the claim, not that the literature has a gap; it is a cue to search wider. The reading, the judgement and the sentence stay yours, which is the point of using Litrevu this way.
Frequently asked questions
Does ChatGPT cause brain rot?
No study has shown that ChatGPT causes “brain rot”, and the MIT Media Lab authors ask readers not to use that phrase about their work. The MIT study measured brain connectivity, recall and ownership during a 20-minute essay task. It found weaker engagement with ChatGPT, not lasting damage.
Was the MIT cognitive debt study peer reviewed?
The MIT “Your Brain on ChatGPT” study was posted on arXiv as a preprint in June 2025 and revised in December 2025. The MIT Media Lab project page still describes it as a preprint (as of November 2026), and the authors call the findings preliminary.
Does using AI hurt critical thinking in students?
Survey studies link heavy AI use with lower critical thinking, but they cannot show cause. Gerlich’s 2025 study of 666 UK participants found a correlation of r = -0.68 between frequent AI tool use and critical thinking scores, and younger participants aged 17 to 25 relied on AI tools most.
How do I use AI without losing my research skills?
Use AI to find, sort and check evidence, and keep the reading, judging and writing for yourself. Form your own view of a paper before asking an AI tool about it, open every passage you cite, and write your own prose. Bastani and colleagues’ 2025 PNAS study found that a hint-giving GPT-4 tutor largely avoided the learning loss that unrestricted GPT-4 caused.
Is it cheating to use AI for a thesis or dissertation?
Using AI for a thesis is not cheating where your university allows the specific use and you declare it, and it is misconduct where your university forbids it. The University of Bristol allows spelling and grammar checks but forbids AI-written thesis text. UCT’s Faculty of Commerce treats undeclared or inappropriate AI use in a dissertation as academic or research misconduct.
Do I have to declare AI use if I only used it to read papers?
Many universities require you to declare any AI use, including reading support. UCT’s Faculty of Commerce requires a front-matter disclosure of all AI use in a dissertation, and the University of Melbourne requires graduate researchers to acknowledge AI use and keep a record of it. Check your own faculty’s rules, because some set stricter conditions than the university.
Where should you start?
Start by deciding which jobs stay yours: the reading, the judgement and the writing. Then check your faculty’s AI rules and keep a log of anything you hand to a tool. If you want an assistant that sends you back to the exact passage in your own papers instead of replacing the reading, try Litrevu with your own papers. Free to start, no credit card required.