Think about the last time you used AI for something. Maybe you asked it to summarise a dense report, draft a reply you didn’t want to write, or explain a concept you half-remembered. In each case you did something your brain is very good at and has done your whole life: you handed a mental task to something outside your head.
Psychologists call this cognitive offloading, and it’s worth understanding clearly, because the conversation around AI tends to collapse into two unhelpful camps — “AI is making us stupid” and “AI is just a tool, relax.” The truth is more precise and more useful than either. Offloading isn’t good or bad in itself. The brain is built to do it. What matters is what you offload, and whether you’re handing away a task you’ve already mastered or one you were supposed to be learning.
Your brain has always outsourced
Cognitive offloading is formally defined as using an external action or object to reduce the mental effort a task would otherwise demand (Risko & Gilbert, 2016, Trends in Cognitive Sciences). That definition is deliberately broad, because the behaviour is everywhere and most of it is ancient. Writing a shopping list instead of memorising it. Setting a reminder instead of holding an intention in mind. Tilting your head to read rotated text instead of mentally rotating it. Using a calendar, a calculator, a map.
None of this is cheating, and none of it is new. The human brain evolved to use the environment as scaffolding — to push cognitive load out into tools and notes and other people so that limited working memory is freed for whatever actually needs it. A surgeon with a checklist isn’t a worse surgeon. A pilot who offloads altitude and fuel to instruments is a safer one. Offloading, done well, is how skilled people stay sharp under load rather than drowning in detail.
So the instinct that offloading must be a loss is wrong from the start. The useful question is narrower: which offloading quietly strengthens you, and which quietly hollows you out.
The Google effect: you already trade knowing for finding
The clearest preview of the AI question came more than a decade before ChatGPT. In 2011, a team led by Betsy Sparrow published a study in Science that became known as the Google effect. Across several experiments, people who expected to be able to look information up later remembered the information itself less well — but remembered where to find it better. When a fact felt saved and retrievable, the brain quietly decided it didn’t need to store the fact, only the path to it.
This is an old idea wearing new clothes. Long before search engines, we used transactive memory — the practical knowledge of who knows what. You don’t memorise everything; you remember that your colleague handles the contracts and your sister remembers the family dates. The internet simply became the most capable member of that network. Offloading facts to a reliable external store isn’t dysfunction; it’s efficient, and it’s roughly what memory already does when it decides what’s worth keeping.
The Google effect is mostly benign because what gets offloaded is storage — the fact sits out there, intact, waiting. You can still reason with it when you pull it back. Which is exactly why generative AI raises a different and sharper question.
What’s actually different about offloading to AI
A search engine offloads where the answer lives. You still have to read, judge, synthesise, and decide. Generative AI offloads a step further up the chain: it can do the reading, the judging, and the synthesising, and hand you the finished conclusion. You’re no longer outsourcing storage — you’re outsourcing the thinking that turns information into an answer.
That’s a meaningful shift. The mental work you skip when you ask AI to “just give me the takeaways” isn’t the work of remembering a fact. It’s the work of wrestling with the material — the comparing, the doubting, the connecting — which is precisely the work that builds understanding in the first place. Offload a phone number and you lose nothing important. Offload the act of reasoning through a problem, every time, and you never build the circuit that does that reasoning.
This is why “AI is just a tool like a calculator” is only half right. A calculator offloads arithmetic you understand in principle. If AI offloads the understanding itself — before you’ve built it — the analogy breaks.
What the 2025 research actually found (and didn’t)
Two studies from 2025 are worth knowing about, precisely because they’re more measured than the headlines they generated.
Michael Gerlich surveyed 666 people and reported a significant negative correlation between frequent AI-tool use and critical-thinking scores, with cognitive offloading appearing to mediate the link — heavier users offloaded more, and more offloading was associated with weaker critical thinking (Gerlich, 2025, Societies). Separately, researchers at Microsoft and Carnegie Mellon surveyed 319 knowledge workers who use generative AI at least weekly. They found that the more people trusted the AI, the less critical thinking they reported applying — while the more confident people were in their own ability, the more they scrutinised what the AI produced (Microsoft/CMU, 2025).
Both findings are genuinely interesting, and both deserve honest caveats. They are correlational and self-reported: they show associations between behaviours people report, not proof that AI use causes a decline. The Gerlich study leans on surveys and interviews; the Microsoft one is a survey of self-described habits, and six of its seven authors work at the company that sells the AI assistant in question. None of this makes the results worthless — it makes them a signal to take seriously without overstating.
And the signal, read carefully, isn’t “AI rots your brain.” It’s subtler and more actionable: the risk rises with trust and falls with effort. When you assume the answer is right and move on, you skip the scrutiny. When you treat the output as a draft to interrogate, you don’t. The same tool produces opposite effects depending on how you hold it.
The Microsoft team put the shift in a useful way: AI doesn’t remove the effort, it moves it — from gathering information to verifying it, from solving the problem to integrating the AI’s response, from doing the task to overseeing it. The thinking doesn’t disappear. It changes shape. The danger is only when it disappears instead of changing shape.
Use it or lose it is real for skills, too
There’s a reason offloading a mastered skill is safe and offloading an unbuilt one is costly, and it comes down to how learning works. Understanding is built by effort — by the slightly uncomfortable work of retrieving, struggling, and doing the thing yourself. That difficulty isn’t a bug to be engineered away; it’s often where the learning actually happens. We’ve written before about how the struggle is frequently the point, not an obstacle to route around.
When you let AI do a task you’ve already mastered, you lose nothing — the circuit exists and stays warm enough. When you let AI do a task you’re still building, you skip the reps that would build it. The result isn’t dramatic or immediate. It’s the quiet kind of loss: a skill that never forms, or one that slowly fades because it’s no longer practised. “Use it or lose it” isn’t a slogan here; it’s a reasonable description of how under-used neural pathways get pruned.
This also explains the uncomfortable asymmetry: AI is most tempting exactly when a task is hard, and tasks are hard exactly when they’d teach you the most. The moments you most want to offload are often the moments you most shouldn’t.
The rule of thumb: offload the output, keep the reasoning
Here’s a single principle that resolves most of it in practice: offload the production, not the thinking.
Offloading the output is nearly always fine — formatting, drafting boilerplate, converting your own ideas into clean prose, generating options to react to, handling the mechanical parts of a task whose logic you already own. That’s the calculator case. It frees working memory for the part that matters, which is what good offloading is for.
Offloading the reasoning is where the cost lives — asking AI to form the judgement, reach the conclusion, or do the core intellectual work of something you’re meant to understand or be responsible for. The tell is simple: if you couldn’t roughly reconstruct or defend the answer yourself, you offloaded reasoning, not production. As a bonus, keeping the reasoning is also the only reliable defence against AI’s confident mistakes — you can’t catch an error in logic you never engaged with, which is the whole reason checking AI output has to be an active skill, not a glance.
How to offload on purpose
A few habits keep AI on the production side of that line:
Attempt first, then compare. For anything you’re trying to learn, take a real swing before you ask. Even a rough attempt primes your brain to actually absorb the AI’s version instead of nodding along to it. The comparison is where the learning is.
Use AI to pressure-test your thinking, not to supply it. “Here’s my reasoning — where is it weakest?” keeps you in the driver’s seat. “Just tell me the answer” hands over the wheel. The way you prompt largely determines which one you get, because the framing of the request is itself an act of thinking.
Keep deliberate no-AI reps for skills you’re still building. A musician doesn’t practise only with the backing track. If writing, analysis, or a particular kind of problem-solving is a skill you want to own, protect some reps where you do it unassisted — not out of purity, but to keep the circuit forming.
Treat every output as a draft to interrogate, never a verdict to accept. This is the single habit the research points at most directly. Trust is where scrutiny goes to die; mild, active skepticism is what keeps the thinking in your hands. If you want the fuller version of this, we’ve written a practical guide to using AI without letting it dull your edge.
And watch your cognitive budget. The point of offloading is to free capacity for higher-value thinking — so spend the freed capacity, don’t just pocket it. Offloading that buys you room to think harder is a win; offloading that buys you room to think less is the trap. (If “cognitive budget” is a new frame, here’s what cognitive load really means for learning with AI.)
The goal isn’t less offloading — it’s better offloading
Your brain was outsourcing cognition long before it had anything as capable as AI to outsource to, and that’s not a weakness to correct. The shopping list, the calendar, the colleague who remembers the thing you don’t — these make you more capable, not less. AI belongs in that lineage. It’s simply the most powerful external mind we’ve ever had, which means the stakes of using it carelessly are higher and the payoff of using it well is larger.
So the question was never whether to offload. It’s which parts to keep. Keep the reasoning, the judgement, and the reps that are still building you. Hand over the rest freely. Do that, and AI amplifies your thinking instead of replacing it — which, in the end, is the entire difference between a tool that makes you sharper and one that quietly makes you redundant.
Learning to work with AI in a way that strengthens your thinking rather than eroding it is exactly what we teach at UnAI Labs. If this is the relationship with AI you want, explore our courses — practical, science-backed, and built for how your brain actually learns.