🔒 Free Privacy App IT Contact Us
← Observatory

AI and the brain: what science really says about the cognitive risk

Is generative artificial intelligence making us less able to think? The question is everywhere, often framed in alarmist tones. In this article from the Retoria Observatory we set aside the headlines and look at the data, because over the past two years scientific research has begun to measure, seriously, what happens to the mind when it delegates thinking to a machine.

The findings don't say that AI makes us stupid. They say something more precise and more useful: the risk is real but conditional, it depends entirely on how we use it, and it concentrates on the ground we care about most — the logical-linguistic one, that is, the ability to reason with words, argue, distinguish, choose. It is the skill with which, before governing any algorithm, we govern ourselves.

The brain has always offloaded

Let's start with a fact well known to cognitive science. Offloading mental effort onto external supports is not an AI invention. We do it with writing, with the calculator, with our calendars. Researchers call it cognitive offloading, and it is often a smart strategy, because it frees up resources for more important tasks.

The problem arises when delegation becomes a total habit. A study published in Nature Scientific Reports showed that habitual use of satellite navigation reduces spatial memory — the ability to find our own way — and that the decline is proportional to how much GPS was used in previous years. Cognitive functions follow an ancient rule: what is not exercised grows weaker.

With generative AI the stakes rise, because for the first time the tool does not replace an accessory task such as remembering a number or a route, but the very heart of intellectual work: writing, synthesising, arguing, deciding.

Cognitive debt, measured in the lab

The most cited study is Your Brain on ChatGPT, from the MIT Media Lab. Researchers followed 54 university students over four months as they wrote essays, split into three groups — those writing with their own mind only, those who could use a search engine, and those using ChatGPT — all monitored with electroencephalography.

Bar chart: brain connectivity drops from writing unaided to using ChatGPT.

Brain connectivity while writing, by group. Source: MIT Media Lab (EEG, 54 students).

The results are clear. Those writing without tools showed the most active and extended neural networks. Those using ChatGPT showed the weakest brain connectivity of the three groups, with some indicators of cognitive engagement lower by up to 55 per cent. But the most interesting finding concerns memory and language: participants in the ChatGPT group struggled to correctly quote the texts they had just handed in and reported the lowest sense of ownership over their own work. The authors call this phenomenon cognitive debt: every time we delegate the effort, we save energy today and pay interest tomorrow, in the form of shallow learning and memories that don't consolidate.

Scientific honesty requires a caveat: the sample is small, the study concerns a specific task and the conclusions will need confirmation from larger research. But the signal is consistent with the rest of the literature.

More trust in the machine, less critical thinking

Two large-scale studies complete the picture. The first, published in the journal Societies by researcher Michael Gerlich on 666 participants, found a significant negative correlation between frequent use of AI tools and critical thinking ability, mediated precisely by cognitive offloading. The most relevant finding for anyone working in education is generational: younger participants showed the highest dependence on the tools and the lowest critical thinking scores. As a correlation, it cannot automatically prove cause, but the direction of the signal deserves attention.

The second is research by Microsoft and Carnegie Mellon presented at the CHI 2025 conference, conducted on 319 professionals who use generative AI at work. The central finding is almost a paradox: the more a person trusts AI, the less they exercise critical thinking on its outputs; the more a person trusts their own skills, the more they check, verify and correct the machine. Critical thinking doesn't disappear, it shifts — from producing the content to verifying it. But it shifts only for those who have the skills and the habit to do so.

The most underestimated front: logical-linguistic decay

Here we reach the heart of the matter, the part public debate overlooks. The greatest risk of generative AI is not that we forget facts: it is that the very instrument we think with — language — is impoverished. And on this, the evidence is becoming solid.

Language flattens out

Research comparing 2,200 texts measured a clear effect: human writing generates far more variety of ideas than writing produced with AI. Each essay written by a person added two to eight times more semantic diversity to the group than an essay generated by GPT-4. In parallel, several studies document that writing together with a language model significantly reduces the lexical, syntactic and semantic diversity of the texts produced.

Bar chart: semantic diversity of texts collapses when writing with GPT-4 versus human writing.

Semantic diversity of texts: human writing vs writing with GPT-4 (study on 2,200 essays).

This is not an aesthetic detail. A review published in 2026 in Trends in Cognitive Sciences — one of the leading journals in cognitive science — warns that language models, as they spread, risk standardising not only language but also thought, reinforcing dominant styles and marginalising alternative voices. When everyone writes with the same assistant, everyone tends to write — and therefore to think — in the same way.

It even changes how we speak

The effect doesn't stop at the written page. A team at the Max Planck Institute for Human Development analysed nearly 280,000 academic videos, comparing the vocabulary used before and after ChatGPT's release. In the eighteen months after launch, people — speaking off the cuff, not reading a script — used the "GPT-words", the terms the model favours, such as delve, meticulous, realm, adept, up to 51 per cent more often.

Bar chart: ChatGPT's signature words in spontaneous speech rise after the model's release.

Use of ‘GPT-words’ in spontaneous speech, before and after ChatGPT. Source: Max Planck Institute (280,000 videos).

It is a striking phenomenon: AI is not just changing how we write, but how we speak spontaneously. Its lexicon enters ours, and with the lexicon comes a certain way of building sentences and arguments.

Why impoverishing language means impoverishing thought

Cognitive science has been saying it for decades: we don't think and then put it into words. Often we think through words. The vocabulary we have defines the distinctions we are able to make; the syntax we master defines the lines of reasoning we can hold together. A poorer language is not just less beautiful: it is less able to catch nuance, to sustain a complex argument, to expose a flawed line of reasoning.

This is where the threads tie together. The studies on critical thinking, on brain connectivity and on linguistic flattening tell the same story from three different angles. If I delegate the wording, I train my judgment less; if I train my judgment less, my language standardises; and a standardised language leaves me less equipped to exercise judgment. It is a loop that can spiral downward — or, if reversed, upward.

A decline that predates AI

There is a reason this risk should be taken seriously: it doesn't start from zero. The OECD's PIAAC survey — which in 2023 measured the skills of around 160,000 adults across 31 countries — captured a decline or stagnation in literacy, the ability to understand and use written texts, across most advanced countries. Between 2012 and 2023 adult literacy fell in the United States and 18 other OECD countries; only Finland and Denmark improved.

Bar chart: adult literacy is declining in far more OECD countries than those where it improves.

Adult literacy trend 2012→2023 across OECD countries. Source: OECD, PIAAC 2023.

There is more, and it is the part that directly concerns anyone in education: the gap is widening. The weakest segment of the population is getting worse, while the top segment is improving. It is the same "cognitive bifurcation" other researchers observe with AI: those who start ahead use the tools to multiply their abilities, those who start behind use them passively and fall further behind.

A note of caution: this data predates the mass diffusion of generative AI and is not its effect. It does describe an already fragile ground, on which passive AI use risks acting as an accelerant, not an original cause.

The proof in the classroom

That this is not just theory is confirmed by an authoritative Italian source. The EDUNext 2026 report, by the Look4ward Observatory (Luiss Guido Carli and Intesa Sanpaolo), ran an experiment on 800 students, assigned to three conditions: working alone, with AI support, or with AI plus a human tutor.

Bar chart: learning is highest in the no-support group compared with the AI group.

Learning by support condition (1–7 scale, higher = better). Source: EDUNext 2026, Luiss · Intesa Sanpaolo.

The result is counterintuitive and valuable. On manageable tasks, the group working without support achieved the highest levels of learning, engagement and motivation. AI became genuinely useful only when the task was complex. In other words: when a problem is within our reach, delegating it to AI makes us learn less; when it is genuinely hard, AI can act as scaffolding. The difference is not the tool, it is the situation and the way we use it.

The risk is not the tool, it's the posture

Put side by side, the studies tell a clear story. Generative AI does not damage the brain like a toxic substance. But it creates the perfect conditions to stop exercising it, because it makes the result immediate and the effort optional. And effort, in cognitive terms, is not a flaw in the process: it is the process. It is precisely the struggle to formulate, remember and structure that builds lasting neural connections and keeps language in shape.

The difference is made by the posture of the person using the tool. Used passively — asking for the ready-made answer and copying it — AI produces cognitive debt and flattened language. Used actively, as an interlocutor that tests our ideas, it can do the opposite: force us to ask better questions, to defend our claims, to spot the errors in its answers. This is what at Retoria we call using AI as a cognitive sparring partner, not as a substitute.

What to do, in practice

The literature offers concrete guidance, valid for school as well as work.

  • Think first, then ask. Drafting your own answer before querying the machine protects memory and the autonomy of judgment.
  • Every answer is a draft, not a truth. It must be read, verified, rewritten in your own words — the act in which critical thinking shifts to checking.
  • Keep your basic skills in training. Writing, calculating, navigating a text: practise them even when the tool could do them for you.
  • Mind your language. Rephrasing in your own words, finding the exact term, varying your syntax: it is the most direct way not to delegate thought along with words.
  • In education, sequence matters most. The student who produces alone first and then compares with AI learns; the one who starts from AI accumulates debt.

The Retoria method

This is why the Retoria method puts logical-verbal skills and critical thinking at the centre of every training programme on AI. Language is the instrument with which we govern the algorithm, and a trained mind remains the only requirement for using any technology well. It is not about using AI less, but about using it as aware humans — staying the owners of our thinking, not its spectators.

To go deeper, read our manifesto or get in touch.

Methodological note

We selected studies published in scientific journals and conferences or produced by recognised research centres. Several of these works have stated limits — small samples, correlations that do not prove cause and effect, specific tasks — and we flag them deliberately. The strength of the whole lies not in any single data point but in the convergence of independent evidence toward the same direction. The PIAAC data, in particular, describe a trend that predates AI and should not be read as its effect.

Sources