Insights
Why AI Polish Undermines Critical Thinking
I noticed something uncomfortable about my own work recently. I was reviewing a document an AI had drafted for me. It was good. Clear structure, reasonable logic, no obvious errors. I read it twice, made three small edits, and approved it. Then I asked myself a question I did not enjoy answering: would I have caught a serious flaw if there had been one? I am not sure I would have. Not because I lack the ability, but because the draft arrived already looking finished.
Polish invites agreement
And I had unconsciously shifted from thinking about the problem to checking the answer. There is research on this now. Microsoft Research surveyed 319 knowledge workers who shared 936 real examples of using AI in their work. The pattern they found was not that AI makes people less capable of thinking. It was subtler and more troubling: the more confidence people had in the AI, the less critical thinking effort they applied. The more confidence they had in themselves, the more they applied. Confidence, it turns out, is a finite resource that moves. When we place more of it in the tool, we withdraw some from ourselves. A separate 2026 study of over 1,200 people found something related — participants were poorly calibrated at estimating how much time AI actually saved them. The felt speed and the real speed diverged. The authors called it a speedup illusion. I keep returning to what these findings imply for organizations, because I do not think it is a training problem. Every organization I know is currently measuring AI adoption. Seats deployed. Tasks automated. Hours saved. Not one is measuring whether the quality of human thinking inside the organization went up or down.
We have instrumentation for the machine and none for
And there is a strange asymmetry here. Microsoft’s 2026 Work Trend Index found that when workers were asked which skills matter most as AI takes on more work, the top two answers were quality control of AI output and critical thinking. Eighty-six percent said they treat AI output as a starting point rather than a final answer. So people know. They can name the skill. What they lack is an environment that protects the conditions for exercising it — time, doubt, dissent, the permission to say “this looks right and I still don’t trust it.” That permission is not a personality trait.
It is an organizational design choice
If a team’s meetings are structured so the fast answer wins, the AI draft will win. If disagreement costs social capital, no one spends it questioning something that already reads well. If the calendar has no space between receiving work and deciding on it, judgment gets compressed into approval. Contribution Leadership starts from the premise that AI should optimize work while organizations optimize humans. What I am learning is that this is not only about unlocking what people can add. It is also about defending what they might quietly lose. The most valuable thing a person brings to an AI-saturated organization may be their willingness to remain slightly unconvinced. Which raises a question I do not have a clean answer to yet: if confidence transfers from the person to the tool, what does a leader do to transfer some of it back? I suspect the answer looks less like a policy and more like a practice. Asking someone to explain the reasoning, not just the output. Reviewing decisions, not just deliverables. Making it normal — expected, even — to say “I used AI here, and here is where I still think it’s wrong.” Small rituals. But they may be the difference between an organization that thinks with AI and one that slowly stops thinking at all.