There's a particular kind of professional anxiety that AI has created, and it doesn't get enough honest examination. It's not the acute fear of job replacement - that conversation has been running for years and at least has the virtue of being concrete. It's something more diffuse: the fear that the long, difficult process of becoming genuinely good at something is becoming economically optional. That expertise, in certain domains, is being demonetized.

The fear is reasonable. A junior lawyer who used to spend years drafting documents can now have a capable AI produce a first pass in minutes. A copywriter who built their career on being able to hold a brand voice across a campaign is competing with a model that can be prompted to approximate that voice from a few examples. A financial analyst who spent a decade learning to synthesize complex information quickly can see that synthesis replicated, imperfectly but adequately, by tools anyone can access.

What does this actually mean? And should the people affected - or the people who would have been those people - be worried?

What expertise actually is

Before we can answer whether AI erodes expertise, we need to be clearer about what expertise is. The folk model - expertise as a large inventory of facts and procedures, accumulated slowly - is incomplete. What distinguishes an expert from someone who has merely read a lot is pattern recognition operating below the threshold of conscious deliberation. The chess grandmaster who "sees" the right move before they've calculated it. The experienced surgeon whose hands know something their verbal mind hasn't yet formulated. The senior editor who reads a paragraph and immediately knows what's wrong with it, without being able to fully articulate why.

This kind of expertise is tacit. It lives in the body and the unconscious cognitive systems as much as in explicit memory. And it has a particular property: it was forged by years of feedback-rich practice. The failures were real, the corrections were immediate, and the patterns were learned through consequence.

AI doesn't have this. What it has is a different kind of pattern recognition, operating over a much larger corpus than any human could read, but without the grounded feedback loop of real-world consequence. The outputs can look similar to expert outputs without being generated by the same underlying process.

The output question versus the process question

Here's where the analysis needs to fork. There are two questions that often get conflated:

First: can AI produce outputs that are as good as those of a human expert? In a growing range of domains, the honest answer is: often, for routine tasks, yes. Not always. Not for complex judgment calls requiring deep contextual knowledge. Not reliably at the frontier of a discipline. But for a lot of the bread-and-butter work that constitutes most of what experts actually do day-to-day - yes, increasingly.

Second: does the process of developing expertise matter, independently of the outputs? This is the more interesting question, and the answer is clearly yes - but the reasons are worth spelling out carefully, because some of them are more robust than others.

"The question isn't whether AI can produce a good document. It's whether people who've never had to struggle through producing bad ones can tell the difference."

What we lose

The first thing we lose is error detection. One of the things that genuine expertise gives you is the ability to evaluate the output you're reviewing, not just produce output yourself. An experienced lawyer can read an AI-generated contract clause and spot the subtle error that looks fine to a non-expert. If we skip the expert-formation process - if the next generation never has to write the bad drafts and get them corrected - we produce people who can operate AI tools but can't audit them.

This is a real problem. It's not hypothetical. It's the same problem that emerged when calculators became ubiquitous and a generation of students stopped developing numerical intuition. They could get the right answer from a calculator but couldn't sense when an answer was wrong.

The second thing we lose is frontier capacity. Expertise isn't just about replicating what's already known. The people at the cutting edge of any discipline got there through the same long apprenticeship that AI now seems to make unnecessary. You can't shortcut your way to knowing enough to push a field forward. The tacit knowledge built through years of practice is precisely what allows you to ask the right questions, recognize the anomaly, make the non-obvious connection.

If the economics of expertise-formation collapse - if there's no viable early career for the junior professionals who are supposed to become the next generation's seniors - we may be depleting a talent pipeline whose consequences won't be visible for a decade.

What we don't lose

That said, there's a version of the expertise-erosion argument that goes too far. The argument that because AI can produce a competent legal brief, the practice of law is therefore hollowed out, or that the accumulated wisdom of human expertise is simply being replaced - this overstates the case.

Most high-stakes expertise involves judgment in genuinely uncertain situations where the right answer isn't obvious from the training data, and where consequences are severe enough that "usually correct" isn't good enough. Medicine, law, engineering, strategic decision-making - the places where expertise genuinely matters most are precisely the places where AI assistance is most valuable as a tool and most dangerous as a replacement.

The surgeon who uses AI to assist in planning a procedure is not diminished as a surgeon. The analyst who uses AI to rapidly surface relevant data is not diminished as an analyst. The expertise is still doing real work - it's just doing it in concert with a very capable assistant rather than alone.

The calibration problem

The challenge, then, is calibration. We need to be honest about which parts of which expertise domains are genuinely commoditized by AI, and respond to that honestly - rather than either denying it (protecting economic interests at the cost of truth) or overgeneralizing it (pretending the entire edifice of professional expertise is now irrelevant).

For some domain-specific, high-volume tasks, AI has effectively replaced significant parts of junior professional work. That's real. The economic and educational structures built around that work need to adapt, and pretending otherwise is wishful thinking.

For deep expertise - the kind built over years, exercised in genuinely novel situations, operating at the frontier of a discipline - the story is more complicated and more hopeful. AI may accelerate how we get there. It may change what we spend our time doing once there. But the destination remains genuinely valuable in ways that are hard to automate.

The people who will navigate this best aren't those who resist AI, nor those who blindly outsource their thinking to it. They're the ones who develop genuine expertise and use AI to multiply it - who know enough to tell when the model is wrong, and have the depth to go where the model can't follow.

If you're a practitioner at this juncture, the AI for Experts path is built for people who want to multiply existing depth rather than start from scratch. And for a closer look at why AI's relationship to knowledge differs so fundamentally from ours, The intelligence we built doesn't think the way we do is the essay that sits alongside this one.