Some time in the last two years, you probably encountered, for the first time, a generative AI system that felt genuinely useful. You asked it something. It produced an answer. The answer was, at least, good enough that you used it. Maybe you used it for an email, a first draft of an essay, a summary of something you didn’t want to read in full, a piece of code you couldn’t quite figure out, a piece of advice on a problem you were stuck on.
Whatever the specific use, you joined several hundred million people who have, in the last few years, started quietly integrating AI into the middle of their thinking lives. Not as a research tool or a search engine, but as something closer to a thinking partner — or, depending on how you use it, something closer to a thinking replacement. The difference between those two uses is one of the more important questions now facing anyone doing serious intellectual work, and it’s worth thinking about carefully before the habits set.
The early productivity research
The most-cited piece of early empirical research on generative AI in professional work came from two economists, Shakked Noy and Whitney Zhang, at MIT. In 2023, they ran a randomised controlled trial in which professional writers (marketers, grant writers, consultants, HR professionals) were given two writing tasks — with some participants permitted to use ChatGPT on the second and others not.
The results were striking. Participants with AI access completed the task 40 per cent faster on average and produced output that was rated 18 per cent higher in quality by blind reviewers. Perhaps more interestingly, the benefit wasn’t evenly distributed. Workers who had initially scored lower on the unaided task benefited more from AI access than workers who had initially scored higher. The distribution of quality compressed — less-skilled writers caught up with more-skilled ones, at least on tasks of this kind.
This finding has been extended in other professional domains. Studies of programmers, consultants, customer-service agents and others have generally found similar patterns: meaningful productivity gains, with the gains concentrated among the less-skilled practitioners.
The business writer Ethan Mollick, at Wharton, has been one of the most thoughtful popularisers of this research. His framework — which he calls the jagged frontier — describes how AI tools are not uniformly better or worse than humans across tasks. Some tasks they handle remarkably well; others they handle badly, often in ways that look superficially similar to the tasks they handle well. A professional who doesn’t know where the frontier is will sometimes get excellent help from AI and sometimes get misleading answers confidently delivered, without a clear signal about which kind of task this one is.
Mollick’s practical message, roughly: learn by using. The frontier isn’t fixed — it shifts with every new model release — and the only way to develop a sense of where it is, for your specific work, is to actually use the tools across many different kinds of tasks and notice where they help and where they mislead.
The counter-evidence on cognition
At the same time, a different body of research has raised questions about what sustained AI use does to human cognition itself — particularly in educational settings.
A 2024 study by Nataliya Kosmyna and colleagues at MIT scanned the brains of university students while they wrote essays in three conditions: unaided, using search engines, and using ChatGPT. What they found was that students using ChatGPT showed substantially weaker neural engagement across several measures — reduced activity in regions associated with memory formation, reduced connectivity in regions associated with integrative thinking, and, months later, reduced recall of what they had supposedly written.
The students weren’t, in other words, doing the thinking the AI was doing on their behalf. They were outsourcing the cognitive work to the tool, and the cognitive work that their own brains were doing — the work that produces the neural patterns that eventually become knowledge — was being quietly skipped.
This matches what many educators have been observing informally. Students who use AI to draft essays produce essays. The essays may be reasonably good. But the students don’t appear to know, afterward, what was in the essays they submitted. They haven’t learned the material. They’ve delegated the learning to a system that can’t remember it for them.
The Kosmyna study is early, and its findings should be held with appropriate care — brain-imaging research is notoriously sensitive to methodological choices, and individual studies often don’t replicate cleanly. But the broader concern it raises is consistent with what cognitive scientists have long known about how learning actually works. Learning, as we’ve covered in other articles in this series, is an active reconstructive process. The brain stores what it has worked to produce, not what has been handed to it. An educational system in which students delegate the producing to an AI is one in which the students are being credentialed for outputs their brains never produced and therefore never learned from.
What this suggests about use
The careful reading of the current evidence suggests a distinction that’s worth holding. Thinking with AI and thinking because of AI are different activities, and they produce different outcomes.
Thinking with AI looks like using the tool to accelerate specific sub-tasks while you do the core thinking yourself. You’ve formed a view. You ask the AI to check it against counter-arguments you might have missed. You’ve drafted something. You ask the AI to identify weaknesses in the structure. You’re stuck on a specific problem. You ask the AI to suggest five different angles of attack, then choose one and work through it yourself. In each case, you’re remaining the intellectual agent. The AI is accelerating specific parts of the process, but the understanding is being built in your own head.
Thinking because of AI looks like using the tool to produce the thinking itself. You haven’t formed a view. You ask the AI what to think. You haven’t drafted anything. You ask the AI to write it. You’re not stuck on a specific problem; you haven’t really tried to solve it. You just ask the AI for the answer. In each case, the output exists, but your own understanding doesn’t develop. You’ve produced artefacts without producing knowledge.
The first use is probably net beneficial, especially for routine tasks. The second use, if it becomes the dominant pattern, is almost certainly costly — not because the outputs are worse (often they’re fine) but because you’re not developing the capacities that would produce those outputs without the tool. Over time, you become less capable, not more, because the tool is doing the work that was supposed to be building your own intellectual muscle.
The skill-floor concern
A related and underdiscussed concern is what the current generation of AI tools does to the floor of competence in any given field.
Traditionally, the threshold to produce competent work in a professional field required real skill. A lawyer had to understand the law well enough to draft a competent brief. A programmer had to understand the language well enough to write working code. A writer had to understand the subject well enough to produce coherent prose on it. The floor of competence was gate-kept by the cognitive work required to clear it.
With AI, much of that floor is now accessible to people who haven’t developed the underlying skill. The competent brief, the working code, the coherent essay — all of these can now be produced by someone whose own capacities, unaided, would not have been able to produce them. This is partly what Noy and Zhang were documenting: the compression of the quality distribution means that the previously-incompetent can now produce outputs that look competent.
The concern is what happens next. If the floor of competence can be reached with AI assistance, there’s less pressure on learners to develop the underlying skills themselves. A student who gets good grades on AI-assisted essays has no particular reason to develop the capacity to write those essays without help. A junior professional who produces passable work through AI delegation never builds the underlying judgement that would, over a career, allow them to do the kind of work AI can’t yet do. The skill distribution gets compressed at the low end — but the high end may become harder to reach, because the developmental path to it (years of struggle producing merely-competent work) has been shortcut.
This is speculative. The full generational effects won’t be knowable for a decade or more. But it’s worth taking seriously, because the choices students and early-career professionals make now about how they use these tools are likely to shape the intellectual capacities they have at forty.
What to actually practise
For a Year 12 student entering university in the next year or two — into an environment saturated with generative AI — a few working principles seem defensible.
Use AI aggressively for tasks you already know how to do. The time you save is real benefit. The skill is already in you, so nothing is lost.
Use AI cautiously for tasks you’re trying to learn. The temptation is enormous to delegate rather than struggle, and the struggle is exactly where learning happens. For tasks that are supposed to be teaching you something — essays, problem sets, anything you’re supposed to understand after having done it — err toward doing the actual cognitive work yourself, and using AI only for specific narrow sub-tasks after you’ve worked through the main problem.
Learn where the frontier is for your specific field. Different domains interact with current AI tools differently. Some work very well with AI assistance; some don’t. Where the assistance is strongest, lean into it. Where it’s weakest, notice, and don’t trust confident-sounding outputs in those areas.
Keep practising without AI sometimes. This is probably the most important habit to build early. Just as athletes train without all their equipment, thinkers should occasionally work through problems using only their own minds. The capacity degrades if never used. The person who can, when they want to, close the laptop and think clearly for an hour is doing something fewer and fewer people can do.
The question that remains
The deepest thing about the AI question, right now, is that the honest answer is we don’t fully know yet. The tools are powerful, the effects are complex, and the long-term consequences for human cognition are genuinely uncertain. We’re all running an experiment on ourselves, and the results won’t be in for many years.
What we do know, from the cognitive-science literature we’ve drawn on throughout this series, is that learning is an active process that can’t be outsourced. Understanding requires doing the cognitive work. A life in which all the difficult cognitive work is delegated to AI is a life in which the person’s own capacities don’t develop — and, over time, may not even be noticed as missing, because the outputs are still there.
This suggests a specific discipline, worth developing while the habits are still forming: ask yourself, in each interaction with an AI tool, whether you’re using it to help your thinking or to replace your thinking. Over a year, the pattern will shape your actual capacity to think. Over a decade, it will shape who you’ve become.
The question worth carrying, especially if you’ve started leaning on these tools heavily:
In a month, or a year, or ten years, will the skill of thinking through a hard problem still live in your mind — or will it live only in the tool?
Key research referenced: Shakked Noy and Whitney Zhang’s 2023 paper on productivity effects of ChatGPT (Science); Ethan Mollick’s writing on the “jagged frontier” (Co-Intelligence, 2024); Nataliya Kosmyna and colleagues’ 2024 MIT study on AI use and neural engagement in essay writing; the broader cognitive-science literature on active learning covered elsewhere in this series.