The research so far points one way: AI that does the thinking for you can reduce how much you engage, learn and remember, while AI that makes you produce, explain and answer questions does not show that harm. The evidence is early, and most studies are small or show links rather than causes. How you use AI matters more than whether you use it.
In 2025, researchers at the MIT Media Lab (Kosmyna and colleagues) asked 54 students to write essays across several sessions while wearing EEG headsets: 18 with ChatGPT, 18 with a search engine and 18 with no tools. Brain connectivity was strongest in the no-tools group, lower with search, and lowest with ChatGPT, up to 55% lower than no tools on their main measure.
The most striking result was about memory of their own work. Shortly after the first session, 15 of the 18 ChatGPT users (83%) couldn't correctly quote a sentence from the essay they had just written, against 2 of 18 in each of the other groups. The authors call the build-up of this effect "cognitive debt".
Limits: it's a preprint that has not finished peer review, with a small sample from five Boston-area universities, and other researchers have published a critique of its methods. Lower connectivity doesn't automatically mean worse thinking either. Treat it as an early signal, not a verdict.
A peer-reviewed survey of 319 knowledge workers, covering 936 real examples of using AI at work (Lee and colleagues, CHI 2025), found that the more people trusted the AI to do a task, the less critical thinking they reported. The more they trusted their own ability, the more critical thinking they did. With AI, their effort moved from doing the work to checking and steering it.
Limits: it's based on what people reported about themselves, at one point in time, so it shows a link, not a cause.
The strongest evidence so far comes from a field experiment with nearly a thousand high-school maths students, published in PNAS (Bastani and colleagues, 2025). One group practised with a plain ChatGPT-style assistant, one with a tutor version designed to give hints rather than answers, and one with no AI.
During practice, the plain assistant raised scores by 48% and the tutor version by 127%. But when the AI was taken away for the exam, students who'd used the plain assistant scored 17% lower than students who never had it. The tutor version largely avoided that drop. Same technology, different design, opposite effect on learning.
A study of 666 people in the UK (Gerlich, 2025) found that heavier use of AI tools went with lower critical thinking scores, and that "cognitive offloading", handing mental work to the tool, explained much of that link. Younger participants relied on AI more and scored lower; more education went with better critical thinking whatever people's AI use.
Limits: it's correlational and self-reported, so it can't show that AI caused the difference.
Older and much better established research points the same way. People remember what they generate themselves far better than what they only read (the generation effect, Slamecka and Graf, 1978). Explaining material to yourself out loud improves understanding (self-explanation, Chi and colleagues, 1994). Testing yourself beats rereading.
TalkGraph was built on the side of this research that helps. You do the talking and the thinking; the AI organises what you said into a map instead of writing it for you. When you pause, it asks you one short question rather than handing you an answer. And your Brain answers from your own words, with links to where you said them.
To be clear: no study has measured TalkGraph itself. It's designed on these principles; the research above is about AI in general.
No study has shown that it lowers intelligence. Early research suggests that letting AI do the work can reduce how much you engage with and remember it, while using AI in ways that keep you producing and explaining does not show that harm.
In a 2025 MIT Media Lab study of 54 students writing essays, the group using ChatGPT showed the lowest brain connectivity on EEG, and 15 of 18 of them couldn't correctly quote their own essay shortly after writing it. It's a small, not yet peer-reviewed study and its methods have been criticised, so it's an early signal rather than proof.
It's the term the MIT researchers used for the idea that relying on AI to think for you saves effort now but builds up a cost later: less engagement, weaker memory of the work and less sense that it's yours.
Handing mental work to a tool, such as a calculator, a search engine or an AI. Some offloading is useful; research suggests heavy offloading to AI goes with weaker critical thinking.
It depends on how. In a large classroom experiment, students who used a plain AI assistant did worse once it was removed, while those using a tutor version that gave hints instead of answers largely avoided that. Use AI to question and test you, not to do the work.
Checked against the papers on 4 October 2026. Research moves fast; follow the links for the latest versions.
Why TalkGraph works: the research →Ask AI about your own thinking →Studying by explaining out loud →The problem with AI notes →