Does Using AI Make You Worse at Thinking?

Using an AI assistant does not make you dumber in any measurable general sense. What the evidence shows is narrower and more useful: when you delegate a task to AI, you do not practise the skill that task would have exercised, and that shows up on tests of the specific skill. (This is the same mechanism behind why your confidence can outrun your code when an assistant writes it.) The best-known study on this, the MIT “cognitive debt” experiment, had 54 participants, is not peer-reviewed, and has drawn a published critique. The best-known counter-evidence, a meta-analysis finding a large learning benefit from ChatGPT, was retracted in April 2026.

The honest picture in late 2026 is therefore two-sided and quite specific, and the specifics matter more than the headline.

The study everyone quoted

The MIT Media Lab preprint divided participants into three groups for an essay-writing task: one using a large language model, one using a search engine, and one using no tools. Each group wrote essays over three sessions under the same condition. A fourth session swapped some participants across conditions. A total of 54 people took part in the first three sessions, and 18 completed the fourth.

The authors used EEG, a scalp recording of electrical brain activity, to estimate cognitive load while writing. Their central finding was a gradient: the no-tools group showed the strongest and most distributed connectivity, search engine users showed moderate engagement, and the LLM group showed the weakest. In the authors’ summary, cognitive activity scaled down in relation to external tool use.

The behavioural findings were the ones that travelled. Self-reported ownership of essays was lowest in the LLM group, and LLM users struggled to accurately quote their own work minutes after writing it. Essays within each group were also more alike in named entities and phrasing, which is the same convergence pattern this site examined in AI ideas converge on sameness.

The authors’ own conclusion is more cautious than the coverage: the findings highlight potential cognitive costs and raise concerns about long-term educational implications, with a call for deeper inquiry.

What the critique says

In December 2025 a group of researchers posted a formal comment on the preprint. They congratulate the authors for initiating the research and collecting a valuable dataset, and then set out concerns intended to improve the manuscript’s readiness for peer-reviewed publication. Their primary concerns cover study design including the limited sample size, the reproducibility of the analyses, methodological issues in the EEG analysis, inconsistencies in reporting, and limited transparency in procedures. Their framing is that some results could be interpreted more conservatively.

That is not a rebuttal. It is a description of a preprint that has not yet been through review, with a sample of 54 split three ways and a fourth session of 18. A reader should weight the EEG gradient accordingly. The quote-your-own-work finding is simpler and more robust to those objections, and it is the one most relevant to a working adult.

The larger surveys

A 2025 study in the journal Societies surveyed and interviewed 666 people across age groups and education levels. It found a significant negative correlation between frequent AI tool use and critical thinking scores, with cognitive offloading as the mediating factor. Cognitive offloading is the term for handing a mental task to something outside your head, whether a calculator, a search engine, or a chatbot. Younger participants showed higher dependence and lower critical thinking scores.

This is correlational. People with weaker critical thinking may lean on AI more rather than the reverse, and the paper cannot separate the two. It also carries a correction published in September 2025: a table was a duplicate of another table, which the author states does not affect the conclusions. That is a minor correction, reported here because this site reports corrections.

The 2026 International AI Safety Report summarises the field the same way. It cites a clinical study in which, three months after AI support was introduced, clinicians’ ability to detect tumours without the AI had dropped by 6 per cent, and the 666-person survey above. It then states that research into AI use, cognitive offloading and critical thinking is nascent and further studies are warranted.

The experiment closest to a developer’s day

The most directly relevant evidence for this site’s readers is a randomised experiment on how developers learn a new library, published as a preprint in January 2026. Developers were assigned to learn an unfamiliar asynchronous Python library with or without an AI assistant. The finding: AI use impaired conceptual understanding, code reading and debugging ability, without delivering significant efficiency gains on average. Participants who fully delegated the coding showed some productivity improvement, but at the cost of learning the library.

The same study identified six distinct patterns of AI use, three of which involved staying cognitively engaged and preserved learning even with AI assistance. That is the actionable part: the outcome tracked how people used the tool, not whether they used it.

Two disclosures. One author is at Anthropic, a company that sells AI assistants, and the protocol was reviewed internally at Anthropic rather than by an external board. The direction of the finding runs against the company’s commercial interest, which is worth noting, but a reader should still know who ran it. Second, the sample was 52 people, and the company’s own summary states the setup differs from agentic coding products, where it expects the impact on skill development to be more pronounced.

The counter-evidence, and the retraction

Until this spring, anyone arguing the other side reached for a 2025 meta-analysis of 51 studies reporting that ChatGPT had a large positive effect on learning performance, at an effect size of 0.867. It was cited over 400 times.

It was retracted on 22 April 2026. The editor’s stated reason is concerns relating to discrepancies in the meta-analysis, raised by two named readers, which undermined the editor’s confidence in the validity of the analysis. The authors did not respond to correspondence about the retraction. Any article or product page still quoting that 0.867 figure is quoting a retracted paper.

The surviving counter-evidence is a 2026 meta-analysis of 35 experimental studies with 4,193 participants, in the same journal. It found a moderately positive effect of ChatGPT on student learning outcomes, g = 0.670, across both cognitive and non-cognitive skills, moderated by subject, duration and instructional mode.

So the two bodies of evidence are not in as much tension as they look. Studies where AI is used as a structured learning aid inside a course tend to find better test outcomes. Studies where AI does the task for you tend to find you learn less about that task. Both can be true.

What this means at your desk

The practical variable is whether your own reasoning stays in the loop. The developer study’s three engaged patterns, and the handwriting versus typing evidence on this site, point at the same mechanism: reformulating something in your own words is what builds the skill, and a tool that removes the reformulation removes the practice. The same holds for reading on a screen, where the cost showed up under time pressure rather than from the device itself.

None of this is a medical claim. If you are worried about memory or concentration in a way that is affecting your life, that is a question for a doctor, not a question about software.

Common questions

Does using ChatGPT lower your intelligence?

No study has shown that. The MIT preprint found weaker EEG connectivity and worse recall of their own essays among 54 participants in the LLM condition, and a 666-person survey found a correlation between heavy AI use and lower critical thinking scores. Neither measures general intelligence, neither is peer-reviewed in a way that has survived critique unchallenged, and the survey cannot tell cause from effect. What the evidence supports is skill-specific: the thing you delegate is the thing you stop practising.

Is the MIT cognitive debt study reliable?

It is a preprint with 54 participants, 18 in the crucial fourth session, and a published comment raising concerns about sample size, EEG methodology, reproducibility and reporting consistency. The authors themselves call for deeper inquiry. Its behavioural finding, that LLM users could not accurately quote work they had just produced, is the most robust part. The EEG gradient should be read as preliminary.

Does AI help or hurt learning?

Both, depending on the design. A 2026 meta-analysis of 35 experiments found a moderately positive effect (g = 0.670) when ChatGPT was used as a learning aid inside courses. A randomised experiment on developers found AI use impaired understanding and debugging of a new library without speeding people up on average, except for those who stayed engaged rather than delegating. The earlier meta-analysis claiming a large effect (0.867) was retracted in April 2026 and should not be cited.

How do I use AI without losing skills?

The developer experiment identified three usage patterns that preserved learning: broadly, patterns where the person kept generating, checking or explaining rather than accepting output. The mechanism matches evidence elsewhere on this site that reformulating in your own words is what builds retention. This article does not prescribe a workflow, because no source here tests one against another in working adults over time.

A skill, not a brain

The question in the search box is whether AI makes you dumber, and the evidence does not answer that question because it does not measure that thing. It measures specific skills on specific tasks, and there it is consistent: delegate the task, lose the practice.

That is neither alarming nor reassuring. It is the same trade a calculator offers, made at a much wider scale and much faster, and the studies that matter are the ones that look at how people use the tool rather than whether they do. Those are only now being run.