There are two confident stories about AI and your brain, and they can’t both be true.
The first is that it’s a lever — a tool that multiplies what a capable person can do, the way a spreadsheet multiplied what an accountant could do, so that the work you get through in a week is work that used to take a month. The second is that it’s a crutch — that leaning on it quietly wastes the muscle underneath, and we’re all busily outsourcing our thinking to something that will happily think badly on our behalf.
Both camps are arguing about the tool. That’s the mistake. The tool is the same tool. What differs is what you do with it — and, more precisely, what you hand off to it. That’s not a semantic dodge. It’s the whole decision, and like most of the decisions I write about here, it’s a trade-off you can make well or make blindly.
Take the worry seriously first
It would be easy, and dishonest, to wave the crutch story away. The evidence for it is real and worth sitting with.
The most-quoted piece is Michael Gerlich’s 2025 study in Societies: across 666 people, frequent AI use was negatively correlated with critical-thinking performance, and the link was mediated by cognitive offloading — letting the tool carry the mental load you’d otherwise carry yourself. The effect was most pronounced in the youngest participants. Alongside it, Lee and colleagues at Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real uses of generative AI at work, and found that the more someone trusted the AI, the less critical thinking they reported doing. And the MIT Media Lab’s Your Brain on ChatGPT wired people up to an EEG while they wrote essays: the group using a language model showed the weakest neural connectivity of the three, produced strikingly similar essays, and struggled afterwards to quote work they’d supposedly just written. The authors call the residue cognitive debt — a nod to technical debt, and a good one.
Now the honest caveats, because leaving them out would be exactly the kind of unexamined confidence the studies are warning about. Gerlich’s finding is a correlation — it can’t tell you whether AI use dulls thinking or whether people who lean on tools were different to begin with. Lee’s is self-reported — it measures what people think they’re doing, not what a test would show. And the MIT study is a preprint that captures what happens during a writing session; it doesn’t demonstrate lasting harm, and its authors are careful not to claim it does. None of that makes the worry go away. But it does mean the crutch story is a warning, not a verdict.
The other half of the evidence
Because there’s a second stack of studies, and it points the other way.
The strongest single piece is a 2025 meta-analysis by Zhao and colleagues in the Journal of Intelligence, pooling 29 experiments rather than surveys or correlations. It found a moderate positive effect of generative AI on higher-order thinking — problem-solving most of all — and, crucially, the size of the effect depended on how the AI was used. That “how” matters more than the headline. Aggregated experiments outrank self-report and correlation as evidence, so this isn’t a footnote you get to ignore.
And the mechanism everyone’s worried about — offloading — turns out to be older and more ordinary than the panic suggests. Risko and Gilbert’s 2016 review in Trends in Cognitive Sciences is the standard reference on it, and their point is that offloading is normal: we’ve always pushed cognition out into the world, onto notebooks, calculators, the alphabetised contacts in a phone. Done well, it frees scarce mental capacity for higher-order work. Their honest caveat is the same one that runs through all of this — the capacity you offload can fade if you never exercise it, and whether it does is a matter of design.
Which is the thread that ties the two stacks together. The negative evidence mostly comes from passive use — taking the answer, moving on. The positive evidence mostly comes from active use — structured, scaffolded, where the person does some of the thinking first and uses the tool to go further. Even Lee’s study contains the tell in miniature: confidence in the AI predicted less critical thinking, but confidence in your own ability predicted more. Same tool. Opposite outcomes. The variable isn’t the machine. It’s the posture of the person using it.
So it’s a decision, not a property
That reframing is the point of this whole post. “Is AI good or bad for your mind?” is the wrong question, in the same way “are microservices good or bad?” is the wrong question. As ever, it depends on what you’re optimising for and what you’re willing to trade — and here the trade is unusually personal. You’re deciding, task by task, which parts of the thinking to keep and which to give away.
Give away the mechanics — the boilerplate, the first draft, the search across a codebase, the tedious translation from one format to another — and you’ve spent the tool well. Give away the judgement — is this true, is this the right call, does this hold against reality — and you’ve bought a crutch and called it a lever.
The handoff is where the risk lives
Engineers have seen this shape before, decades before anyone wired a language model into their editor. In 1983 Lisanne Bainbridge wrote a short, devastating paper called Ironies of Automation. Automate the routine parts of a job, she observed, and you leave the human only the hard exceptions — the very situations that demand the most skill. But because the human no longer practises the routine, they’re less prepared for those exceptions, not more. The better the automation, the more critical the human’s residual role becomes, and the less equipped they are to perform it. Aviation learned this the hard way: the autopilot flies the easy hours and hands control back to the pilots at precisely the worst moment, when something has already gone wrong and their hands are cold.
Lee’s paper describes the same phenomenon in gentler, newer language. Generative AI, they say, shifts knowledge work away from doing the task toward verifying, integrating and overseeing what the machine produced — they call it “task stewardship.” Read as a threat, that’s Bainbridge’s irony arriving on schedule: I’ve automated the doing and demoted myself to checking homework, using skills I’m no longer practising. Read as an opportunity, it’s an elevation: my scarce attention goes to the highest-value part, the judgement. It is the same finding read two ways — and which one you get is decided at the handoff. The seam between what the machine does and what you do is exactly where the risk concentrates, which, to an architect, is a very familiar place for risk to live.
How I actually use it
I’ve spent the better part of a year with an AI agent as a genuine fixture of my working day — not a novelty I open occasionally, but hours of it, most days, on real infrastructure and production problems. Recently I looked back over a month of that work in aggregate, and the pattern that jumped out wasn’t that the tool had done my job. It’s that I’d spent most of my energy as a reviewer.
In that month the agent confidently proposed root causes for an outage that were simply wrong. It invented a configuration setting that doesn’t exist. It once dropped a fabricated figure into a document as though it were established fact. Every one of those was caught — not because the tool flagged its own uncertainty, but because I went and checked against the real thing: the actual code, the actual database, the actual running system. The value it delivered wasn’t the confident first draft. The value was in the verification, and the verification was mine.
That is task stewardship, and it’s the opposite of deskilling only because I still bring the judgement to it. The tool did the typing, the searching, the tedious tracing across repositories — and it did them faster than I ever could. But the questions that decided whether the work was any good — is this actually true, is this the right trade, does this survive contact with reality — those stayed with me. It accelerated the work. It didn’t do the thinking. On the day I let it do the thinking, the confident-and-wrong root cause ships, and the crutch story comes true in my own output.
There’s a closer example than any of that, and it would be odd to leave it out: this post. These posts are drafted with AI assistance. I bring the argument, the reading behind it and the position I want to take; the tool turns that into prose; then I do the thing I’ve just described — check what’s actually true, cut what overreaches, and decide what the piece is really about. Earlier I listed “the first draft” among the mechanics worth handing over. I meant it, and it would be a bit rich to recommend that and then quietly make an exception of myself.
Whether I’ve kept the part that matters is a reasonable thing to ask, and I’d rather you asked it knowing than guessed. The about page sets out the loop.
The catch, and it’s a real one
Here’s the part that keeps me honest. This posture — delegate the mechanics, keep the judgement — only works if you have the judgement to keep. I can catch a wrong root cause because I’ve debugged enough systems to smell one. Take that same tool and hand it to someone who hasn’t yet built that instinct, and there’s nothing to steward with; they can only trust. That’s not stewardship, it’s offloading of the kind Gerlich measured — and it’s how you end up fluent in prompting and unable to tell when the answer is wrong.
The scaffolding research points at the fix, and it’s almost annoyingly old-fashioned: do some of the thinking first, then let the tool take you further. Form the judgement before you automate around it. Which means the danger was never using AI. The danger is using it instead of ever developing the judgement that lets you use it safely — and for people earlier in their careers than me, that’s not an abstract risk, it’s the whole ballgame.
All of which is a personal discipline — my own posture at my own keyboard. The organisational version of the same problem — how you govern a whole team’s use of these tools, when the frameworks we lean on for everything else don’t yet reach this far — is a harder question, and one I’ll come back to.
Finishing the sentence
So: lever or crutch? Neither, inherently. It’s a tool that faithfully amplifies whatever you bring to it. Bring judgement and it multiplies your judgement; bring nothing and it multiplies nothing, or worse, multiplies your mistakes at speed and with great confidence.
The useful version of the question — the one worth actually answering — is the one you decide at every handoff: what am I giving away here, and am I keeping the part that matters? Automate the typing. Keep the thinking. And keep exercising the muscle you’d otherwise be tempted to let the machine carry, because the day you can’t tell a good answer from a confident one is the day the tool stops being a lever at all.