The Expert Costume

For most of human history, sounding like an expert and being one traveled together. Fluency was expensive. To talk through a cardiac case, a tax structure, or the failure mode of a steel weld with confidence and precision, you had to have spent years close to the work. So we used the sound as a proxy for the substance. Hear someone explain a hard thing clearly, and you could bet they knew it. It was a good bet for a very long time.

That bet just broke.

A large language model will now produce fluent, structured, confident, authoritative prose on any subject, for anyone, in about four seconds. The costume is free. Which makes an old question with a settled answer suddenly urgent: what is an expert, actually? And can you become one without the years?

The settled definition

Across dictionaries, psychology, philosophy, and the courts, the definition barely moves. Merriam-Webster calls an expert "one with the special skill or knowledge representing mastery of a particular subject," derived "from training or experience." The expertise-research literature is blunter: an expert shows "consistently superior performance on a specified set of representative tasks for a domain." Peer recognition rides along. Experts are the people whose judgment other competent people defer to, because that judgment has proven reliable.

Three claims are hiding in there, and each one carries weight:

1. Mastery, not familiarity. Broad and deep command of the domain. Not a tour of it.
2. Performance, not recall. Experts reliably outperform, especially on problems nobody has seen before. Knowing things is table stakes. Deploying them under novelty is the job.
3. Provenance. The skill comes from training or experience, and it is recognized by people positioned to judge it.

The load-bearing word is "consistently." An expert is not someone who can produce a right-sounding answer. An expert is someone you can rely on when the situation is new and the answer is not in the back of the book.

Does expertise require experience?

Yes. On this point, serious definitions and the research converge almost without exception.

The Dreyfus model of skill acquisition maps five stages: novice, advanced beginner, competent, proficient, expert. The last step is not more of the same. It is a phase change. The rules the novice clutches dissolve into intuition. The expert stops computing the situation and starts seeing it: whole, immediate, with the relevant cue already in the foreground. Chess masters do not enumerate moves; they perceive the board. Veteran operators do not run the checklist; they notice the one number that is wrong. There is exactly one road to that stage, and it runs through the stages before it.

Anders Ericsson spent a career showing what the road is made of: deliberate practice. Not time served. Thousands of hours at the edge of your ability, on tasks designed to stretch it, with feedback that punishes your errors while they are still cheap. Experience without that structure produces confidence, not mastery. You can repeat your first year twenty times and call it twenty years.

And much of what the practiced expert acquires is tacit. It is pattern recognition that cannot be written down, because its owner cannot fully articulate it. This is why expertise without meaningful experience is essentially unobserved in the wild. It would be like mastering the kitchen by reading recipes. The reading helps. The vocabulary is real. But judgment is built from reps, and there is no syllabus for the smell of something burning.

The decoupling

Here is what AI actually changed. It did not make expertise easier to build. It made expertise easier to fake, including to yourself.

A novice with a frontier model now produces artifacts that read like the work of a tenth-year practitioner: the analysis, the code, the strategy memo, the diagnosis. Researchers already have names for what this creates. They call it "hollowed-out knowledge." Some call its carriers "incompetent experts." The output is impressive. The person who shipped it cannot tell you why it works, where it stops working, or what it quietly left out.

The deeper damage is to the loop. Deliberate practice works because you struggle, you err, and reality corrects you. Prompting skips the struggle. Nothing compounds. The work product leveled up; the worker did not. You get the Dreyfus stage-five costume on a stage-one body.

Where the costume tears

Generation got cheap. Verification did not. That asymmetry is the whole story.

The model's characteristic failures share one property: they are invisible to anyone without domain judgment. Hallucinated facts delivered with total confidence. The right template pattern-matched onto the wrong problem. Smooth over-generalization. Quiet internal contradictions. The expert reads the page and feels the wrong note. The margin is too good. The citation does not smell right. The edge case is not handled, it is just unmentioned. The novice reads the same page and sees polish.

So AI amplifies experts and exposes novices, just on a delay. The exposure does not arrive in the demo. It arrives at the edge case, the client question the prompt did not anticipate, the deal where the boilerplate meets reality and loses.

The paradox

Which lands us on the paradox: foundational expertise matters more with AI, not less.

The field evidence keeps repeating the same shape. In real enterprise deployments, the mechanical layer of expert work goes away and the judgment layer expands to fill the time. Security analysts who spent ninety percent of their hours on triage now spend those hours hunting threats. The work that remains, and grows, is exactly the work that requires the trained eye. Full autonomy holds up only where volume is high, stakes are low, and errors are recoverable. Everywhere else, the human in the loop is not a transitional arrangement. The human is the product.

You need expertise to ask the model a question worth answering. You need it to recognize a wrong answer wearing a right answer's clothes. You need it to notice when a genuinely good answer is solving the wrong problem. Philosophy has a technical term for fluent speech produced with indifference to its truth. The model is not lying to you; it has no idea and no stake. Caring whether it is true is your job. Caring well is a skill, and you can only have built it the old way.

For the founder-operator

If you have run a business for fifteen or twenty years, you own the one asset this era actually rewards: judgment with scar tissue. You have watched a thousand deals, hires, launches, and failures resolve. That pattern library is what makes AI leverage in your hands instead of liability. The same prompt produces more in your hands than in a novice's, because you can see what came back.

But the era hands you two traps.

The first is outsourcing your own loop. The moment you let the model do your thinking in the domain where your judgment is the business, you start converting your edge into everyone's baseline. Use it to multiply your judgment, never to replace the practice that maintains it.

The second is quieter: hollowing out your team's pipeline. Juniors who never struggle never become seniors. If your people are shipping AI output they cannot defend, you are not accumulating capability. You are accumulating invisible debt, work that looks done but is not owned. The seniors you will need in five years are supposed to be in the struggle right now.

The design principle follows directly. Give AI the mechanical layer. Protect, on purpose, the reps that build judgment. Match review depth to stakes. And install one bar for everything that ships: can you defend this without the model in the room?

The test that survives

The definition of an expert has not changed. Mastery built from experience. Superior performance under novelty. Judgment that others can rely on. What changed is the work the definition does. It used to help you choose surgeons and structural engineers. Now it is the filter you need for everything you read, everyone you hire, and everything your own company ships.

Fluency used to be the test. Now fluency is the noise floor. The test that survives is the one the definition pointed at all along: put the person somewhere the answer is not written down yet, and watch.