Is AI making us dumber?

One study put electrodes on people's heads while they wrote essays with and without an assistant. What it found is narrower than the headlines and more uncomfortable.

Analysis by Inzonex · published 2026-09-25 · counted from published open data · how we work

Key figures

Participants writing with an LLM showed the weakest brain connectivity of the three groups. Writing with no tools produced the strongest.

LLM users struggled to quote their own essays minutes after submitting them.

They also reported the lowest sense of ownership over what they had written.

What was actually done

Researchers at the MIT Media Lab split participants into three groups for an essay writing task: one used a large language model, one used a search engine, one used neither. Brain activity was recorded by EEG during writing. 54 participants completed the first three sessions over four months, and 18 came back for a fourth in which the groups were swapped onto the opposite condition.

That swap is the cleverest part of the design. It is what lets you ask whether the effect follows the tool or the person.

What it found

GroupBrain connectivity during the taskRecall of own essay
No toolsStrongest and most distributed networksHighest ownership
Search engineModerate engagementBetween the two
Language modelWeakest connectivityStruggled to quote own work

There is no chart here because the study reports patterns of neural connectivity and behaviour, not a single quantity that can honestly be put on an axis.

The authors summarise it as cognitive activity scaling down in proportion to external tool use. The striking behavioural result is the quoting one: people who had just submitted an essay could not reliably reproduce a line from it.

What the authors mean by cognitive debt

The phrase in the title is cognitive debt, and it is a precise metaphor rather than a rhetorical one. Debt is a benefit taken now against a cost paid later. The essay gets written faster. What is skipped is the effortful construction that would have left the material in your head.

That framing also explains why this is not a claim that people become stupid. Nothing was measured about general intelligence. What was measured is how much of the work you did yourself, and what you retained as a result.

What the study cannot tell you

Fifty-four people, one task type, four months, and a preprint. That is a real study with real instruments, and it is not a verdict on a civilisation. Anyone citing it as proof that AI is destroying human cognition is overreaching, in the same way that dismissing it because the sample is small is underreaching.

Two specific limits matter. Essay writing is a task where the effort is the learning, so it is close to the worst case for delegation; using a model to draft a routine email is not the same activity. And the study measures what happens during and just after writing, not what a person can do a year later.

The practical version

A second and much larger study points the same way by a different route. Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real uses of AI at work and found that the more someone trusted the tool, the less critical thinking they reported doing; the more they trusted their own judgement, the more they did. The effort did not vanish so much as move, from producing an answer to verifying one.

The usable rule from this is about sequence, not abstinence. If the point of the task is the output, delegating it costs you nothing you wanted. If the point of the task is that you end up understanding something, doing it yourself first and using the model to check is the version that leaves the understanding with you.

The uncomfortable part is that the two feel identical while you are doing them. The participants who retained least were not aware they had retained less.

Sources and method