Is AI Making You a Worse Thinker? How to Stay Sharp and Sustain Critical Thinking

AI does not make you a worse thinker on its own. The harm comes from habit, from handing it the hard part, the reasoning, the first draft, the working out, until the skill you meant to build begins to fade. What follows is the honest version of the research, and a plain way to use these tools so they sharpen your thinking rather than stand in for it.

You feel it most in the half-second before you type.

There is a problem in front of you, an essay that won’t begin, a tricky message to a colleague, a concept that hasn’t clicked, and somewhere in that half second, your hand has already drifted towards the chat box. Not because you’ve tried and failed.

Because trying has started to feel optional. I caught myself last week asking a chatbot to rephrase a sentence I could have written in my sleep, for no reason except that asking was easier than thinking. The sentence wasn’t hard. I was just out of the habit of bothering.

That reflex is the whole story. The interesting question was never whether AI is clever. It obviously is. The question is what happens to you when you stop needing to be.

How you can use AI and critical thinking

Here’s the honest version, up front, because you’re busy and you came for an answer rather than a sermon. AI does not automatically make you a worse thinker. Offloading the mental effort that builds your thinking does, and AI happens to make that offloading effortless. Reach for it to skip the hard part, and the benefits of the hard part leave with it. Reach for it to pressure-test work you have already done yourself, and it can make you sharper.

The tool is neutral.

The habit is not. Almost everything that matters here comes down to telling those two apart, which is roughly the difference between using a calculator and leaning on a crutch you have forgotten you’re holding. One leaves your legs working. The other does the walking for you until, one day, you try to stand and find out what’s happened.

So no, you are not doomed. But the worry that sent you searching this is not silly either, and it’s worth understanding properly rather than through a frightening headline.

What cognitive offloading actually means

Cognitive offloading is the act of handing a mental task to something outside your head so your own brain does less of the work. You have done it your whole life and most of it is completely sensible. Jotting a number down instead of memorising it is offloading.

So is a shopping list, a calendar alert, the sat-nav telling you to turn left in two hundred metres. Externalising memory is one of the oldest and best tricks our species has.

What has changed is the scope of what we can now hand over. The older tools offloaded storage. Where things are, when things happen, what the figure was. Generative AI offloads the act itself, the reasoning, the drafting, the working-out, even the judgement about whether the result is any good. That is a different category of help, and it arrives with a bill attached.

There’s a name for that bill, and it’s the single most useful idea in this whole debate. Cognitive debt is what builds up when you repeatedly skip the effortful thinking that would have grown a skill, so the skill quietly fails to develop, or slowly fades. Like the financial kind, it stays invisible right up until it’s called in, which tends to be the precise moment you need to think without the machine and discover the muscle isn’t there. You can run up cognitive debt for months and feel marvellously productive the entire time. That is exactly what makes it sneaky.

The one study worth actually knowing about

If you read about this anywhere, you read about a single piece of research, so it’s worth understanding what it really found rather than what the internet decided it found.

In June 2025, a team at the MIT Media Lab led by Nataliya Kosmyna published a study with a title built for going viral. They called it “Your Brain on ChatGPT”, and the subtitle described the accumulation of cognitive debt when using an AI assistant to write essays. It did go viral, and within days it had been flattened into the claim that ChatGPT rots your brain. That is not quite what it found, and the gap between the two is the most interesting part.

The set-up was simple. Participants wrote essays in one of three conditions. One group used ChatGPT, one used a standard search engine, one used nothing but their own head. The researchers recorded brain activity with EEG across several sessions. The people leaning on the language model showed less of the connected, effortful brain activity you’d associate with deep engagement. They also remembered their own essays less well afterwards and felt less ownership of them, which makes a grim kind of sense. It’s hard to feel proud of, or even to recall, something you didn’t really make. Then came the finding that gave the paper its name. When the AI users were later asked to write unaided, they performed worse than the group that had never touched the tool. They had borrowed fluency, and ended up poorer for it.

Now the part the headlines skipped, and the reason you can use this without overclaiming. This was a preprint, not yet through formal peer review when it broke. The sample was small, in the region of fifty-odd people. It tested one narrow task, essay writing, under artificial lab conditions, which is a long way from proving anything permanent about how your brain works in general. The authors themselves were careful about all of this. So treat the study as a sharp, suggestive early signal, not a final verdict. That honesty isn’t a weakness in the argument. It’s the difference between a useful idea and a scare story, and your readers can tell which one they’re being handed.

What the study does give us, even at this early stage, is a clean illustration of the thing worth worrying about. When you let the machine do the cognitively demanding part, the part that would have stretched you, you tend to get worse at that part. The effort wasn’t a tax you were paying on the way to the answer. The effort was the thing building you. And that idea doesn’t rest on one study at all. It rests on something psychologists have understood for decades.

Why the effort was never the price, it was the point

This is the bit most coverage misses, and it’s where a little cognitive science earns its keep, without a single contested statistic in sight.

Psychologists talk about germane cognitive load, which is the good kind of mental effort, the effort that actually constructs understanding and lays down durable skill. The struggle to wrestle a messy idea into a clear paragraph is not wasted time on the way to a finished paragraph. The struggle is how your brain builds the capacity to think clearly in the first place. Hand that struggle to a machine and you get the paragraph, but you skip the construction. You’re left holding the product and missing the process that would have changed you. The output looks identical. You are not.

There’s a related and slightly counterintuitive idea from the memory researcher Robert Bjork called desirable difficulties. It turns out that learning which feels hard, slow and a bit frustrating often sticks far better than learning that feels smooth and easy.

The fluency is a trap. When information goes in effortlessly, your brain treats it as cheap and lets it go. When you have to fight for it, retrieve it, reconstruct it, you keep it. This is the reason rereading your notes feels productive and does almost nothing, while testing yourself feels horrible and works brilliantly.

Generative AI is, in effect, a desirable-difficulty removal machine. It is engineered to make everything feel smooth. Ask, receive, move on. That smoothness is exactly what makes it such a pleasure to use and such a quiet thief of the friction your brain needs to grow. The feeling of understanding it gives you is real, but it’s a feeling, not the same thing as understanding. You can read an AI’s flawless explanation of a concept, nod along, feel you’ve completely got it, and discover the next day that nothing landed, because you never did the work that makes knowledge yours. The confidence arrives instantly. The competence never shows up.

None of this means hard for the sake of hard. Suffering through a tedious formatting job teaches you nothing, and offloading that to a machine is a gift you should accept gladly. The whole art is learning to tell the difference between effort that builds you and effort that just drains you, then guarding the first while cheerfully giving away the second. Most people never draw that line. Drawing it is most of the battle.

Why does this land hardest if you are young or early in your career

If you are a student or in the first few years of work, you’re standing exactly where this hits hardest, for two reasons.

The first is that you’re still building the muscles. An experienced lawyer who offloads first drafts to AI is leaning on judgement they spent fifteen years developing the slow way. They have a deep model to fall back on, and to check the machine against when it goes confidently wrong. If you offload your first drafts before you’ve built that model, there’s nothing underneath to catch the errors or to grow from. You’re not freeing up expertise to do higher work. You’re skipping the apprenticeship that would have produced the expertise. The gym comparison is almost too tidy. Paying someone to lift the weights for you produces a finished workout and precisely zero muscle.

The second reason is structural, and it’s quietly becoming one of the defining career questions of the decade. The entry-level tasks that AI swallows fastest, the summarising, the basic drafting, the first-pass research, are the very tasks that used to be how juniors learned a trade.

Those boring jobs were never just busywork to keep you occupied. They were the lower rungs of the ladder. If those rungs are being sawn off, the people who’ll thrive are the ones who keep practising the underlying skills anyway, even when a machine would do them faster, not out of nostalgia but because that practice is what turns a beginner into someone worth promoting. The colleague who quietly keeps their own thinking sharp while everyone else outsources theirs is going to look remarkably valuable in about three years, and they’ll have no idea why it came so easily to them.

I’m not telling you to refuse the tools. Refusing them is its own kind of self-sabotage, and you’d be competing with one hand tied behind your back. I’m telling you to use them like someone who fully intends to still be good at their job once the novelty wears off and everyone else has gone soft.

How to use AI without going soft

So here’s the practical part: how to keep AI as a tool that extends your thinking rather than one that slowly stands in for it. None of this is about willpower or purity. It’s about a few small habits that put the friction back where it belongs.

1. Try it before you ask it. 

The single most important rule, and the one that fixes most of the problem on its own. Make your own attempt first, however rough, then bring the AI in to improve it. Draft the email badly, then ask for a sharper version. Sketch your own answer to the essay question, then have it poke holes in your reasoning. This keeps you in the thinking seat and turns the AI into an editor rather than an author. The order is everything. Effort first, assistance second. I now make myself write a scrappy opening paragraph before I’m allowed to open any AI tab, and the difference in how much I actually remember afterwards is not subtle.

2. Offload the chores, guard the thinking. 

Draw a line, explicitly, between tasks that build you and tasks that merely drain you. Reformatting a reference list, tidying a messy table, summarising a document you’ve already read closely, and generating boilerplate. Hand these over without a flicker of guilt. Reasoning through a hard problem, forming an argument, learning something you genuinely want to keep, making the judgment call, these you do yourself. The aim was never to use AI less. It’s to use it for the right things, so the hours it frees up get spent on harder thinking, not on no thinking at all.

3. Make it explain, not just answer. 

When you do use AI to learn something, don’t accept the finished result and move on. Ask it to show its reasoning, to walk you through the why, to define its terms. Then put the explanation aside and try to reproduce it in your own words without looking. If you can’t, you didn’t learn it, you just watched it happen. This one move converts AI from an answer vending machine into something much closer to a patient tutor, and tutoring is one of the few uses that genuinely helps rather than hollows you out.

4. Turn it into a sparring partner. 

Here’s the inversion that flips AI from a threat into an asset. Instead of asking it to do your thinking, ask it to make your thinking harder. Have it argue the opposite side of your essay so you’re forced to defend yours properly. Ask it to quiz you on a topic rather than explain it, which is simply active recall with a chatbot attached, and active recall is one of the most reliable study methods there is. Ask it to find the single weakest point in your plan, then sit with the answer. Used this way, AI adds desirable difficulty instead of removing it, and that is the entire game in one habit.

5. Keep one thing you always do unaided. 

Choose a regular task you will never delegate, and protect it like a stubborn old ritual. For me it’s the first draft of anything I care about; the machine doesn’t get near it until I’ve made my own mess on the page. For you it might be working through a problem set, planning your week longhand, or writing in a journal. The point is to keep at least one muscle in constant use, so the ability to think from a blank page never fully leaves you. It’s a tiny, cheap insurance policy against your own cognitive debt, and you’ll be quietly glad of it.

6. Treat every output as a claim, not a fact. 

AI is fluent, confident and regularly wrong, and it will state an invented source with exactly the same poise it uses for a true one. Get into the habit of verifying anything that matters before you trust it. This isn’t only about avoiding errors, though it will save you from some genuinely embarrassing ones. The act of checking keeps you in the sceptical, evaluating frame of mind, which is the very capacity most at risk of softening. Doubt is a skill like any other. Practise it on your tools.

You’ll notice that not one of these asks you to give up AI. They ask you to stay awake while you use it. That’s the whole prescription, really. Not abstinence, attention.

Frequently asked questions

Does using ChatGPT make you lazy?

It can make laziness frictionless, which isn’t quite the same thing. The tool doesn’t create the urge to skip effort, but it removes every obstacle to acting on it. The fix isn’t to avoid ChatGPT, it’s to decide in advance which tasks you’ll do yourself and to attempt things before you ask for help.

Is AI actually bad for your brain?

The early evidence suggests that habitually offloading demanding mental work to AI is linked to weaker performance on that same work when you later have to do it alone. That’s a finding about habits, not about damage. Used to add challenge rather than remove it, AI may well do the opposite and sharpen you.

Can AI improve critical thinking?

Yes, if you make it disagree with you. Asking it to argue the other side, test your reasoning or quiz you turns it into a sparring partner, and that kind of friction is exactly what builds thinking. The harm comes from using it to reach conclusions without doing the reasoning, not from using it at all.

What is cognitive debt?

Cognitive debt is the gradual cost of repeatedly skipping the effortful thinking that would have built or maintained a skill. It stays hidden while the machine does the work, and it shows up as a gap in your own ability the moment you have to think without it.

How can students use AI without cheating themselves?

Use it after your own attempt, not instead of it. Have it explain rather than answer, test you rather than tell you, and check its work rather than trust it. The students who’ll benefit are the ones who treat it as a demanding study partner, not a shortcut to a finished assignment.

The reflex is yours to keep or break

The tool isn’t the problem. The half-second reflex is, and that reflex belongs to you, which means it’s yours to retrain. Every time you reach for the chat box, you’re casting a small vote about the kind of mind you want to have in five years, and those votes add up far more quietly than you’d ever guess from any single one of them.

The good news is that none of this asks you to be a purist or to fall behind the people racing ahead with these tools. It just asks you to use them like someone who still intends to be able to think when the power goes out. Do your own bit first. Keep one muscle in use. Make the machine argue with you now and then. Stay a little sceptical of how smoothly it all goes down. Do that, and AI stops being a replacement for your thinking and turns into the thing it should have been all along, a reason to do more of it.