The evidence increasingly says yes — not because AI reasons badly, but because outsourcing the effortful parts of thinking lets the underlying skills weaken. Heavy reliance on AI tools correlates with lower critical-thinking scores, and controlled studies find measurable drops in memory, neural engagement, and originality when people let AI do their thinking for them. The cost is deferred and easy to miss — researchers call it “cognitive debt” — and how you use AI determines whether you accrue it.
The concern is not that AI is unintelligent. It is that intelligence, like muscle, is maintained by use. When a tool removes the strain, it can quietly remove the training along with it — and the bill arrives later, in capability you no longer have.
Does using AI actually reduce critical thinking?
The strongest available evidence points to a real negative association, though causation is still being established. A 2025 study by Michael Gerlich, surveying and interviewing 666 people, found that frequent AI-tool use correlated significantly with weaker critical-thinking scores, with cognitive offloading — handing mental tasks to an external system — as the mediating mechanism. Younger, heavier users scored lowest; higher education partly buffered the effect.
Knowledge-worker research reinforces the pattern. In a 2025 study of professionals by Lee and colleagues, the more people trusted an AI’s output, the less critical scrutiny they applied — and they reported expending less cognitive effort on the task. The mechanism is intuitive: confidence in the tool substitutes for engagement with the problem. Correlational studies cannot prove that AI causes the decline, but the mechanism is coherent and converges across independent samples.
What is “cognitive debt”?
Cognitive debt is the deferred cost of repeatedly letting a machine do cognitive work you would otherwise do yourself: convenience now, diminished capability later. The term comes from a 2025 MIT Media Lab study (Kosmyna and colleagues) that tracked 54 people writing essays over four months under EEG, split into an LLM group, a search-engine group, and a no-tools “brain-only” group.
The LLM group showed the weakest, least-distributed brain connectivity, the lowest sense of ownership over their own writing, and — strikingly — difficulty quoting essays they had just produced. The deficit also persisted: when LLM users were later asked to write unaided, they still underperformed. They had borrowed against their own engagement, and the debt did not clear when the tool was removed.
Why does handing your thinking to AI weaken it?
Because the effort you skip is the exact process that builds and maintains the skill. Forming an argument, holding competing ideas in tension, retrieving and connecting what you know — these are the “reps.” Remove them and the capability atrophies, the same way navigation skill fades under constant GPS.
There is a second, subtler cost: homogenization. Work by Sourati and colleagues (2026) finds that large language models pull human expression and reasoning toward a generic, mainstream mean. Lean on them and you don’t only think less — you think more like everyone else, surrendering the idiosyncratic angle that distinguishes original judgment. Notably, the MIT data suggests sequence matters: people who did their own thinking first and brought AI in afterward stayed cognitively engaged; those who led with the tool did not.
Does this affect everyone equally?
No — and the asymmetry is the practically important part. Experts can use AI as a multiplier because they retain the judgment to evaluate, correct, and reject its output. Novices cannot, because they lack the baseline the tool is supposed to assist. Research by Shen and Tamkin (2026) on AI and skill formation underscores the sharper risk: experienced people lose skills to AI, but early-career workers and students may never build them, treating the model’s reasoning as their own before forming any of their own.
The regional stakes vary. In CEE transformation economies, where the human skill base is the competitive asset, eroding it undercuts the core advantage. In the Gulf’s young, rapidly scaling workforce, the formation question dominates — a generation is building its professional reasoning with AI in the loop from day one. In US-scale organizations, homogenized output multiplied across thousands of users compounds into systemic sameness.
How do you use AI without getting worse at thinking?
Treat AI as a sparring partner, not an oracle, and protect the cognitive work that builds you. Four operating rules:
- Effort-first, AI-second. Form your own position before you prompt. Bring AI in to pressure-test, not to originate.
- Make it argue back. Ask the model to challenge your reasoning, surface what you missed, and build the strongest case against you — not to confirm you.
- Keep judgment where it counts. Offload the routine; retain the calls that require your accountability. When you accept an AI’s conclusion without working through the reasoning, you end up laundering your own judgment — presenting a verdict as considered when the thinking never actually happened.
- Preserve the reps. Deliberately keep doing some hard cognitive work unaided. The point is not to refuse the tool; it is to refuse the debt.
Frequently asked questions
Does AI make you dumber?
Not in the sense of lowering raw intelligence. The better description is skill atrophy: capabilities you stop exercising — recall, synthesis, sustained reasoning — weaken, and the loss is gradual enough to go unnoticed until you need them.
Is cognitive offloading always bad?
No. Offloading is how every tool from the calculator to the map works, and it usefully frees attention for harder problems. It becomes a problem only when you offload the skills you still need to own — the ones your judgment, credibility, or profession depends on.
Can the effect be reversed?
Partly. Short-term studies suggest the deficit lingers after AI use stops, but deliberate practice and effort-first habits appear to help rebuild engagement. Long-term reversibility is not yet well established — a genuine gap in the current evidence.
How can leaders stop AI from eroding their teams’ thinking?
Build effort-first workflows, normalize using AI as a critic rather than a ghostwriter, and protect junior skill formation specifically. The failure mode to watch is “AI-drafted” silently becoming “unexamined” — treat AI output as a first draft to be interrogated, never as a finished verdict.
