Introduction

A small, unreviewed MIT study on ChatGPT and essay-writing has become a flashpoint for a much larger question: as generative AI takes over more of our thinking, what exactly are we losing, and can we get it back? In June 2025, a team at the MIT Media Lab led by researcher Nataliya Kosmyna published a paper with a memorable, slightly ominous title: “Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Task.” It had not been peer-reviewed, and it studied only 54 participants. That did not stop it from becoming one of the most widely discussed pieces of AI research of the year, cited in newsrooms, classrooms and policy circles as evidence that something real, and possibly troubling, happens to human cognition when a large language model takes over a thinking task. Whether or not the study’s findings hold up to further scrutiny, the question it raised has not gone away, and it sits at the centre of a fast-growing body of research into what generative AI is doing to human reasoning, memory and judgement.

The MIT experiment was carefully designed. Researchers split participants into three groups: one wrote SAT-style essays using ChatGPT, one used a conventional search engine, and one used no external tool at all, relying purely on their own knowledge. Each group repeated the exercise across several sessions, and in a fourth, “crossover” session, some ChatGPT users were asked to write unaided while some previously unaided writers were given access to the tool. Throughout, an electroencephalogram recorded participants’ brain activity across 32 regions, and researchers separately analysed the essays’ linguistic patterns and interviewed each participant about how well they could recall and felt ownership over what they had written.

The results, while limited in scope, were striking. Brain connectivity, measured through EEG network analysis, scaled down systematically with the amount of external support participants used: the brain-only group showed the strongest and widest-ranging neural networks, the search-engine group sat in the middle, and the ChatGPT group showed the weakest overall neural coupling. Participants who used ChatGPT across sessions were also less able to quote or recall their own essays afterward, and reported a diminished sense of ownership over what they had produced. Perhaps most tellingly, when ChatGPT users were switched to writing unaided in the fourth session, they underperformed participants who had been writing without assistance from the start, suggesting that reliance on the tool had left them somewhat worse equipped to do the task alone. Kosmyna and her co-authors coined a term for this pattern: cognitive debt, the idea that offloading effortful mental work to an external system in the short term can accumulate into a longer-term deficit in independent thinking capacity.

The Science Behind the Worry

To understand why a single small study generated so much attention, it helps to look at the underlying neuroscience it draws on. The human brain’s capacity to change is not fixed at any point in life; a property called neuroplasticity allows it to continually reorganise its structure and function in response to how it is used, forming new neural connections and strengthening or weakening existing ones based on repeated activity patterns. This is the same basic mechanism that lets a violinist develop finer motor control in her left hand or a London taxi driver develop an enlarged hippocampal region for spatial memory. It cuts both ways, though: patterns that go unpractised tend to weaken, which is the biological basis for the old adage about not using something and eventually losing it.

Cognitive load theory offers a complementary frame. The human brain has a genuinely limited capacity to process new information at any given moment, and reducing unnecessary load — by cutting distractions or automating repetitive sub-tasks — can, in principle, free up mental resources for more demanding work. This is the optimistic case for AI-assisted thinking: that offloading rote or mechanical tasks to a tool lets a person spend more of their limited cognitive budget on the parts of a problem that genuinely require judgement, creativity or synthesis.

The trouble is that not all cognitive offloading is created equal, and generative AI offloads a fundamentally different category of work than earlier tools did. This is where a much older, well-established finding becomes relevant: the so-called “Google effect,” first documented in a 2011 study by psychologist Betsy Sparrow and colleagues, which found that when people expect information to remain easily accessible online, they become less likely to remember the information itself and more likely to remember where to find it. That is a meaningful shift in how memory works, but it largely concerns factual recall. Generative AI tools go considerably further, capable of offloading not just where to find information but the actual cognitive work of synthesising it, summarising it, drafting arguments from it, and generating original ideas around it — a qualitatively deeper form of delegation than simply outsourcing memory for facts.

When Offloading Becomes Overreliance

None of this means every use of AI in learning or knowledge work is harmful. Not all mental effort is equally valuable, and there is a reasonable case that some cognitive tasks — rote formatting, basic information retrieval, first-draft grunt work — are not where genuine learning or insight happens anyway. But genuine learning does depend on specific forms of effort: retrieving knowledge from memory under some difficulty, synthesising disparate pieces of information, catching and correcting one’s own errors, and reflecting on what has been produced. When a generative AI tool automates these particular processes rather than merely the mechanical surrounding tasks, it risks removing the very friction through which learning happens in the first place.

This concern scales up in professional contexts too. A 2025 Microsoft Research study of 319 knowledge workers who used generative AI in their daily workflows found a systematic shift in the nature of the critical thinking those workers actually performed. Rather than eliminating critical thinking altogether, AI use redirected it: cognitive effort moved from information gathering toward information verification, from independent problem-solving toward integrating and adjudicating AI-generated responses, and from executing tasks directly toward supervising and checking the execution that AI had already carried out. This finding echoes a much older and still-influential 1983 paper by engineering psychologist Lisandra Bainbridge, “Ironies of Automation,” which observed that automating a process tends to strip operators of the regular hands-on practice that builds and maintains their skill, while simultaneously handing them the much harder job of supervising a system whose internal workings they may not fully understand — precisely the situation many knowledge workers now find themselves in with generative AI. The upshot is not that critical thinking disappears in an AI-saturated workflow, but that it relocates: increasingly, the skill in demand is not generating an answer but recognising when an AI-generated answer is wrong, questioning its assumptions, and correcting it — a skill set that is, notably, much harder to develop if a person has not first built strong foundational reasoning through unaided practice.

That last point connects to what researchers call the “black box” problem: most users of generative AI tools do not understand the internal processes by which those tools arrive at an answer, which limits their ability to meaningfully interrogate or challenge the output in front of them. Combined with a well-documented human tendency to place undue trust and confidence in machine-generated output — even when it contains factual errors, unsupported claims or significant omissions — this creates conditions in which externally generated content is favoured over self-generated reasoning, gradually eroding a person’s confidence in and reliance on their own independent judgement. In an information environment already struggling with the rapid spread of unverified claims and misinformation, a citizenry and workforce less equipped to critically interrogate what it is told is a genuinely consequential risk, one with implications that extend well beyond individual learning outcomes into questions of national cognitive resilience and competitiveness.

The 3R Framework: A Way to Stay in the Loop

The more useful contribution of recent research, though, is not simply cataloguing the risk but proposing how to manage it. A theoretical framework developed earlier this year by researcher Simone Rossi and colleagues, published in npj Artificial Intelligence, starts from the premise that AI-generated output carries no inherent meaning until a human interprets it, situates it in context, and adds the nuance that only lived experience and judgement can supply. The act of doing that interpretive work, the framework argues, is itself what sustains cognitive engagement and, over time, healthy brain plasticity — meaning the value of the mental effort lies not in resisting AI use altogether but in remaining an active, responsible participant in how its output gets used.

Rossi and colleagues distil this into what they call the 3R framework: Results, Responsibility and Response. The idea is that individuals should actively adjudicate the results an AI system produces rather than accepting them uncritically, take genuine responsibility for how those results are applied by weighing them against relevant social, ethical and cultural context, and shape a considered response rather than simply forwarding the machine’s output unchanged. This stands in explicit contrast to passive engagement — accepting a chatbot’s suggestions or following its recommendations without scrutiny — which the framework identifies as the pattern most likely to erode cognitive engagement over time. Active engagement, by contrast, meaning interpreting, questioning, refining or genuinely co-creating with an AI tool’s output, appears far more likely to sustain or even strengthen cognitive functioning, because it preserves the very interpretive and evaluative work that passive use would otherwise remove from the human side of the interaction.

Building Habits Before They Become Deficits

Translating this into practice matters most, researchers suggest, at the earliest stages of education, where habits of engagement with AI tools are still forming and therefore most malleable. Concretely, that means treating AI as a resource to consult rather than a shortcut to the finished product; using it as a starting point for discussion and further problem-solving rather than an endpoint; deploying it to locate evidence and stress-test one’s own arguments through structured fact-checking; and using it as a platform for genuinely problem-based or phenomenon-based learning, where the AI tool supports exploration rather than substituting for it. Extending similar norms into professional workflows — treating AI output as a first draft to be actively verified, questioned and reworked rather than a finished deliverable to be passed along — applies the same underlying logic to knowledge work more broadly, and is arguably where the stakes are highest given how quickly generative AI is being embedded into everyday professional life across sectors.

None of the research reviewed here claims to have settled these questions definitively. The MIT study itself is explicit about its limitations: a small sample, brief writing sessions, no peer review at the time of publication, and no claim to demonstrate permanent or irreversible cognitive change. The broader research agenda — how different AI tools affect cognitive engagement across different personality types, learning styles, age groups and cultural contexts — remains very much a work in progress, and any confident, sweeping claim about AI’s long-term effect on human cognition should be treated with appropriate caution.

Conclusion

What the emerging evidence does suggest, even allowing for its limitations, is that the question worth asking about generative AI is not simply whether it makes people faster or more productive at a given task, since it plainly often does. The more important question is what kind of engagement a person maintains with the thinking that AI performs on their behalf. Passive, uncritical delegation — accepting outputs without interrogation, treating a chatbot’s first answer as a finished one — appears to carry a genuine cost to the cognitive processes that underpin learning, memory formation and independent judgement over time. Active, critical engagement — questioning, verifying, contextualising and taking ownership of what an AI tool produces — appears to avoid much of that cost, and may even sustain the very reasoning capacities that critics worry AI erodes. As generative AI becomes further embedded in how societies learn, work and make decisions, preserving human cognitive capacity will not be a matter of avoiding these tools, which is neither realistic nor obviously desirable. It will depend on the much harder, more deliberate work of teaching people, starting early and reinforced throughout their working lives, to remain the ones doing the thinking even when a machine is standing ready to do it for them.

Consulting Editor