How to keep critical thinking sharp in the age of AI

Stay Sharp: Protect Critical Thinking From AI

Treat AI as a sparring partner, not an oracle. A Microsoft-Carnegie Mellon study of 319 workers found higher trust in AI predicts less critical thinking, while self-confidence predicts more. Form your own view first, ask the model to argue against you, verify claims, and protect deep-work time for unaided reasoning.

The most worrying line in the 2025 research is not that AI makes us dumber, but that it quietly narrows the range of answers we even consider. When Microsoft and Carnegie Mellon surveyed 319 knowledge workers across 936 real tasks, they found something a builder and a coach should both take seriously: the more people trusted the AI, the less critical thinking they did, while the more they trusted their own judgment, the more they did. Confidence, it turns out, is the hinge. The danger is not the tool; it is the moment you stop noticing that you have stopped thinking.

A separate 2025 study by Michael Gerlich in the journal Societies, covering 666 participants, found a strong negative correlation between heavy AI use and critical thinking, mediated by what researchers call cognitive offloading, the mental equivalent of letting your legs atrophy because the car does the walking. The effect hit younger users hardest. This maps cleanly onto Daniel Kahneman's Thinking, Fast and Slow: the fluent, instant answer from a model is pure System 1, and a polished paragraph feels true precisely because it arrived effortlessly. Critical thinking is System 2, and System 2 is lazy by design. AI removes nearly all of the friction that used to force it awake, so you have to reintroduce that friction deliberately.

So the practical question is how to keep System 2 in the loop on purpose. The first habit I hold myself and my coaching clients to is forming a view before you prompt. Write your own rough answer, your own decision, your own thesis first, and only then ask the model. This preserves exactly what the Microsoft study found protective, self-confidence, and turns the AI into a check rather than a crutch. The second habit is to make the model argue against you. I routinely ask Claude or GPT-5.5 to give me the three strongest reasons I am wrong, to steelman the opposing position, to find the flaw in my reasoning. A fascinating CHI line of research on provocations found that AI which deliberately pushes back, rather than agreeing, measurably restores critical engagement. Most people never trigger this, because the default behaviour of every leading model in 2026 is to be agreeable and to flatter your framing.

The third habit is verification as a discipline, not a vibe. The same Microsoft study reframed modern knowledge work as information verification, response integration, and task stewardship, which is a polite way of saying your job is now to audit a confident machine. Treat every factual claim, citation, and number as unverified until you check it against a real source; I have caught Claude and GPT-5.5 inventing plausible references more than once, and the more authoritative the prose sounds, the harder I look. Use a tool like Perplexity or NotebookLM, which grounds its answers in documents you actually supply, whenever provenance genuinely matters. I write more about keeping curiosity and real inquiry alive in my essay on the art of discovery, because the instinct to interrogate is the muscle most at risk of wasting away.

The fourth habit is structural, and it borrows directly from Cal Newport's Deep Work: protect blocks of unaided thinking. Decide in advance which problems you will reason through with no model open at all, a strategy memo, a hard people decision, the shape of next year. Ethan Mollick, in Co-Intelligence, frames the model as a collaborator, but a collaborator you never disagree with is not a collaborator, it is an echo. The point is not to use AI less; I use it constantly, all day. The point is to stay the one who is doing the thinking. The concrete rule I give every leader I coach is simple: never let AI deliver a conclusion you have not personally stress-tested, and once a week, solve one meaningful problem with every tool closed, just to confirm the muscle still works.