There is a story people tell about AI tools. It goes like this: the new generation of coding assistants, reasoning models, and formal verification agents will level the playing field. A junior engineer with Copilot will write code like a senior. A biologist who never learned Rust will prove theorems about protein folding. The tool democratizes expertise.
This story is wrong. And the evidence against it is piling up from every direction at once.
Gergely Orosz noticed that AI formal verification tools have not made formal verification mainstream. Hillel Wayne made the same observation from inside the field: the people getting value from AI-assisted theorem proving are the people who already know theorem proving. The tool makes them faster, but it does not recruit new people into the practice. The barrier to entry is not the tool, it is the judgment required to use the tool.
This pattern repeats across domains. AI coding assistants like Copilot and Cursor have been in wide use for three years now. The evidence is clear: senior engineers get dramatically more value from them than junior engineers. Give the same tool, the same prompt, the same generated code to two engineers with different experience levels, and you get two different outcomes. The senior spots the subtle off-by-one error in the generated code. The junior ships it. The senior finishes in twenty minutes. The junior spends an hour debugging the AI's bad suggestions. The same tool, the same output, different ability to judge quality.
The bottleneck has shifted from producing code to judging it. And judgment is a skill you cannot prompt your way into. It is built through years of debugging your own broken code, watching your own assumptions fail, developing the instinct for where bugs hide. AI can generate code but it cannot give you that instinct. The more the tool generates, the less visible the struggle that builds the instinct in the first place. The apprenticeship problem of AI is that it wipes the learning surface clean.
Martin Kleppmann argued that better UX will bridge this gap. The idea is that AI tools will eventually ask the right questions to elicit specifications from humans without requiring formal training. But the human still needs to know what they do not know in order to answer those questions. A non-expert asked "should this be allowed?" will say yes without seeing the three edge cases they just endorsed. The gap does not disappear, it just moves to a different surface.
The historical precedent is the calculator. When the HP-35 arrived in 1972, the narrative was that it would democratize mathematical computation. Anybody could compute complex functions at the push of a button. And it was true, in a narrow sense. But the people who got the most value from the HP-35 were not the people who could not do math before. They were the engineers and scientists who already knew what the answer should look like, who could spot when the calculator gave a wrong result, who could use the tool to explore design space faster. The calculator amplified existing expertise. It did not create it.
Fifty years later, the same pattern is playing out with AI. This is not a bug in the current generation of tools that will be fixed by better models. It is a structural property of the relationship between tools and expertise. A tool that generates output does not reduce the need for judgment. It shifts the bottleneck to judgment. And judgment is scarce, expensive, and slow to build.
There is a forgotten middle tier in this story: the engineer with three to five years of experience. Not expert enough to go ten times faster alone, but with enough battle scars to evaluate AI output critically. This group might be the biggest beneficiary of AI tools, because they have enough judgment to use the amplification but not enough raw speed to outrun the tool. The tool helps most at the margin where you can evaluate quality but not produce it fast enough on your own. But this is a narrow window, and it is not the democratization that the marketing promises.
The practical consequence of this thesis is uncomfortable for anyone selling AI as a talent gap solution. If AI tools amplify existing expertise, then the return on investment per seat is highest at organizations that already have deep talent. A startup with three staff engineers gets more value per AI dollar than a startup with twenty junior developers. The tool does not fill the gap, it magnifies it. This concentrates the benefits of AI where expertise already runs deep, which is the opposite of democratization.
I think the next three years will make this pattern unmistakable. Coding interviews will shift from generation to judgment. The ability to produce code will become cheap, even trivial. The bottleneck will be the ability to evaluate code quality, to spot the edge cases, to say "this looks right but it is wrong." That is a skill you cannot prompt your way into, and it is built through experience debugging your own mistakes. The interview that used to ask "write a sorting algorithm" will ask "here is a sorting algorithm generated by AI. Find the bug."
The tools are real. The speedups are real. But the democratization narrative is a story we tell ourselves to avoid the uncomfortable truth: that some skills cannot be shortcut, that expertise is built through struggle, and that the tools that make experts faster also make the gap between experts and everyone else more visible. The expertise multiplier is real. It just does not multiply evenly.