For a long time, fluency was a reasonable shortcut.
If someone could write a clear brief, explain the numbers, structure an argument, or speak confidently about a technical subject, it was fair to assume they had done at least some of the thinking.
Not always. There have always been people who could turn a weak idea into a handsome slide deck. PowerPoint did not invent bluffing. It merely gave bluffing better transitions.
But producing fluent work took effort. The effort itself was a weak signal of understanding.
AI has made that signal much weaker.
A polished answer can arrive before a view does
Someone can now have an articulate memo, a plausible strategy, a legal-sounding explanation, or a confident analysis before they have decided what they think.
This is not automatically bad. A draft can help people think. It can give a shy person a way into a conversation. It can translate jargon. It can make an early idea legible enough to improve.
The danger comes when the polish ends the thinking instead of beginning it.
A clean answer creates social pressure. It looks finished. It has headings. It may cite sources. It sounds more certain than the person who is still honestly working through the problem.
Then the person with a messy but valuable question can look unprepared next to the person with a smooth but unexamined answer.
That is a bad trade for any organisation that claims to value judgment.
Expertise is changing shape, not disappearing
The response should not be to fetishise rough work. Nobody needs a return to the age when important ideas had to be hidden inside a badly formatted document to prove they were authentic.
The response is to become clearer about what we are actually trying to recognise.
Expertise is not the ability to produce a first answer. It is the ability to notice what the first answer leaves out.
It is knowing which source should change your mind. It is being able to say where the numbers came from. It is seeing the hidden assumption inside a tidy recommendation. It is recognising the customer, context, or consequence that makes the obvious answer unsafe.
In other words: expertise increasingly lives in the questions a person can ask of a fluent answer.
Research on how people judge AI-assisted work suggests this tension is social as well as technical. In one experiment, participants judged delegation to an LLM as less acceptable than delegation to a human and reported lower trust in the person overseeing the work. The study is narrow—it concerns research scenarios, not every workplace—but it captures an instinct people already have: when a machine did part of the work, they want to know whether the human still understands it. Niszczota and Conway
That instinct is not anti-AI. It is a demand for accountability.
Stop rewarding the performance of certainty
Leaders can make this worse by treating speed and polish as proof.
They reward the person who produces the neat answer before the meeting. They ask for a one-page summary before anyone has been allowed to investigate the question. They call uncertainty a lack of executive presence, then wonder why the company is full of confident documents and surprised people.
AI will amplify this tendency because it makes certainty cheap.
The countermeasure is not bureaucratic review for its own sake. It is better questions.
- What would change your mind?
- What did the model assume?
- Which part of this answer did you verify yourself?
- What would the person closest to the work say is missing?
- If this recommendation fails, where will it fail first?
These questions are not a test of whether someone used AI. They are a test of whether someone owns the conclusion.
The future will not belong to the person who can make the most convincing first draft.
It will belong to the person who knows when a convincing first draft should make everyone slow down.