Fluency Was Never the Skill

For most of human history, if a person could speak fluently about a subject, it was safe to assume they had spent time with it. Fluency was expensive. To produce three coherent paragraphs on the tax treatment of cross-border settlement, or the failure modes of a battery cell, you had to have read, argued, been corrected, and remembered. The words were the visible end of an invisible process, and because the process was costly, the words were a reliable sign that it had happened.

Almost every institution we use to sort people is built on that sign. The job interview is a fluency test. So is the cover letter, the college essay, the grant proposal, the expert witness, the customer email that gets escalated because it sounds like the sender knows what they are talking about. We rarely measured competence directly, because measuring it directly is slow. We measured fluency and trusted the correlation.

The correlation is gone. It did not weaken. It broke, in about two years, and I do not think most of the institutions that depended on it have noticed.

What the machines actually did

The systems we now use produce fluent language about anything, at no cost, with none of the process behind it that the fluency used to signal. That is not an insult to the technology. It is a description of it. A language model wants nothing and knows nothing in the way a person knows a thing they have paid for. It produces the most plausible next passage given everything it has read, and plausibility, it turns out, is what fluency always was. We had simply never seen plausibility separated from effort before.

So now every fluency test in the world can be passed by anyone with a browser. The cover letter proves nothing. The essay proves nothing. The confident, well-structured email proves nothing. A signal that took a civilization centuries to learn to read became noise inside a single product cycle.

Notice what did not change. The skill underneath is exactly as rare and exactly as valuable as it was. There are still people who understand cross-border settlement and people who do not, people who can find the fault in a system and people who can only describe one. The machines did not make understanding common. They made the appearance of it free, which is a different thing, and in some ways a more disruptive one.

Where this shows up first

Hiring will feel it before anyone. A résumé and a writing sample were never great instruments, but they were cheap and they were correlated with something. Now a hiring manager reads forty flawless applications and has no idea which were written by the person applying, and the honest answer is that it no longer matters, because the writing was never the job. The job was the thinking the writing used to prove.

Education is feeling it already. An essay assignment was a device for forcing a student to think for several hours and then show the residue. When the residue can be produced without the hours, the assignment stops measuring what it was built to measure, and teachers are left grading an artifact that no longer points at anything.

Expertise itself will feel it last and worst. Much of what we call authority in public life is fluency at scale: the person who can speak smoothly on camera about a subject is assumed to know it. That assumption was always shaky. It is now indefensible, because the smooth speech can be generated, rehearsed, and delivered by someone who could not answer a single follow-up question.

What replaces the signal

Every time a proxy breaks, the world goes back to measuring the underlying thing, expensively, until it finds a new proxy. I think we are at the start of that cycle, and the shape of it is already visible.

The demonstration comes back. Not "tell me about a time you solved a hard problem" but "here is a hard problem, solve it while I watch." The oral examination, which universities largely abandoned because it did not scale, scales badly for the same reason it works: it cannot be outsourced. Expect more of it, in hiring, in licensing, and in classrooms.

The follow-up question becomes the test. Fluent language survives the first question and dies on the third. Anyone whose job involves judging other people is going to have to learn to keep asking, past the point where the answers stop being smooth, because that is where the person begins.

Provenance starts to carry weight. Where a piece of work came from, who is answerable for it, what it cost to produce, and what happened the last time this person was wrong. Reputation built over years is very hard to generate, which is about to make it far more valuable than a portfolio built over a weekend.

Judgment gets priced separately from expression. For a long time we paid one salary for both, because they arrived in the same person. Expression is now nearly free. The premium moves entirely to the part that is not: knowing which of the fluent answers is wrong, and being willing to say so when the room has already agreed.

The part I find hopeful

There is a strain of thinking that treats all of this as loss, and some of it is. A great many people built careers on being articulate, and articulacy alone is worth much less than it was. That is a real cost and it will land on real people.

But something is also being restored. For a long time the people who could talk about a thing were mistaken, systematically, for the people who could do it, and the people who could do it but not perform it were passed over. Fluency was a tax on the quiet, the second-language speaker, the engineer who could fix the machine but not narrate the fix. If the proxy is dead, the tax goes with it. What is left is harder to fake and fairer to measure.

Nobody knows what self-improving systems will be able to do, and I do not pretend to. But whatever they become, more capable will not mean more human, and the thing they made cheap this decade was never the skill. It was the costume the skill used to wear. The costume is off. Now we have to look.