Do large language models affect cognitive capacity?

By Matthew Parish, Associate Editor

Sunday 26 April 2026

The arrival of large language models has revived an old anxiety in new dress: that tools which extend the mind may also atrophy it. From the invention of writing to the rise of the search engine, each generation has suspected that externalising thought weakens the inner faculty. Yet the evidence regarding contemporary systems is at present fragmentary, methodologically uneven and frequently overstated. What can be said with confidence is more modest — and more interesting.

The central difficulty lies in the poverty of the research base. Despite the ubiquity of systems such as ChatGPT, there are remarkably few controlled longitudinal studies examining their cognitive effects. What exists is often preliminary, speculative or philosophically framed rather than empirically grounded.

A frequently cited experimental result comes from a small study associated with the MIT Media Lab. Participants using a large language model to write essays exhibited lower neural engagement, as measured by EEG (a diagnostic test that records the electrical activity of the brain using electrodes attached to the scalp), and tended over time to rely increasingly on copying and pasting generated text. This has been interpreted as evidence that such systems diminish critical thinking. Yet even on its own terms, the study invites caution. Its sample size was small, its task narrow and its temporal scope limited. Writing essays under laboratory conditions is not the same as engaging in real-world reasoning — nor does reduced neural activation necessarily equate to diminished intellectual capacity. It may indicate efficiency, delegation or simply a change in cognitive strategy.

More sober academic commentary reflects this ambiguity. A short piece in Nature Human Behaviour suggests that the capacity of large language models to generate complex outputs with minimal input “has the potential to impoverish our own writing and thinking skills”. The language is telling: potential, not demonstrated effect. It is a warning grounded in plausibility rather than empirical confirmation.

Other strands of research complicate the picture further. Studies examining interaction with large language models indicate that they can alter patterns of problem-solving and decision-making, but do not uniformly degrade them. In some contexts, users become faster and more productive; in others, they exhibit reduced cognitive effort. This duality echoes earlier technological transitions — the calculator did not abolish arithmetic, but it did change the distribution of mental labour between human and machine.

A deeper question concerns what, precisely, is meant by “analytical capacity”. The human mind is not a monolithic faculty but a constellation of processes: memory, abstraction, inference, linguistic manipulation and judgement under uncertainty. Large language models interact unevenly with these components.

They appear to externalise certain forms of cognition. Drafting, summarisation and even elementary argument construction can be delegated. This resembles what philosophers of mind have called “extended cognition” — the notion that tools can become part of the thinking process itself. If a notebook may serve as an external memory, a language model may serve as an external interlocutor.

But the models themselves are not minds in any meaningful human sense. A growing body of research emphasises that they simulate patterns of language rather than instantiate understanding. They exhibit behaviours that resemble reasoning, yet lack the underlying semantic grounding characteristic of human thought. Indeed their susceptibility to cognitive biases — priming effects, numerical distortions and other artefacts — suggests not intelligence but statistical mimicry of human error.

This distinction matters because it reframes the concern. The risk is not that machines think better than humans and thereby render human reasoning obsolete. It is that humans may come to rely upon outputs whose apparent coherence masks underlying fragility. The so-called “AI trust paradox” captures this unease: as models become more fluent, their errors become harder to detect.

There is also evidence of a more subtle cultural effect. Some researchers have argued that widespread use of large language models may homogenise expression, amplifying dominant linguistic patterns and reducing diversity of thought. If this proves correct, the danger is not merely individual cognitive decline but collective convergence — a narrowing of intellectual variation.

Yet it would be a mistake to cast this development solely in negative terms. Large language models may also function as cognitive scaffolding. They can expose users to alternative formulations, prompt reflection and assist in structuring complex ideas. They resemble a particularly patient and infinitely available tutor. The critical variable is not the existence of the tool, but the manner of its use.

The historical analogy with writing is instructive. When writing first became widespread, critics feared the erosion of memory and the weakening of oral reasoning. In a narrow sense they were correct: literate societies do rely less on memorisation. Yet writing also enabled new forms of analysis — law, science, philosophy — that would have been impossible without external records. The mind did not diminish; it reorganised itself.

The present moment may represent a similar reorganisation. If large language models encourage passivity — the uncritical acceptance of generated text — then a diminution of analytical capacity is plausible. If however they are used as instruments of interrogation — tools to be questioned, challenged and refined — they may instead augment reasoning by shifting cognitive effort towards higher-order judgement.

The most honest conclusion therefore is one of uncertainty bounded by caution. The existing research suggests that large language models can reduce cognitive effort in specific tasks, and may, under certain conditions, discourage active reasoning. But the evidence is neither robust nor comprehensive enough to support sweeping claims of cognitive decline.

What can be said — and perhaps this is the more important observation — is that these systems alter the ecology of thought. They change what it means to know, to write and to reason. Whether this transformation ultimately weakens or strengthens the human mind will depend less upon the machines themselves than upon the intellectual discipline of those who use them.