The pitch is familiar by now. AI is going to democratize expertise. Coding assistants will let anyone build software. Reasoning models will put a PhD in every pocket. Formal verification tools will make proving theorems as easy as writing prose. The gap between the expert and the novice, the story goes, is about to collapse.
I think the opposite is happening. AI tools don't democratize expertise. They amplify it. And the people who benefit most are the ones who already had it.
The mechanism is simple. AI tools produce output. They generate code, proofs, analyses, summaries. But production is only half the equation. The other half is judgment: knowing whether the output is any good. A junior engineer and a senior engineer can feed the same prompt into the same tool and get the same code back. The senior spots the subtle off-by-one error, the unnecessary allocation, the edge case the prompt didn't specify. The junior ships it. The tool didn't make them equal. It made their difference more visible.
This is not a bug in the current generation of tools. It is a structural property of the situation. Judgment cannot be prompted. It cannot be generated. It comes from the long slow accumulation of mistakes, of debugging your own broken code at 2am, of learning the shape of a bad design by having lived through its consequences. AI tools are, in a cruel twist, making this learning harder. When the tool generates working code on the first try, novices skip the debugging step that traditionally built their judgment. The learning surface gets wiped clean. You develop the instinct for spotting bugs by encountering bugs in a form you can learn from. AI hands you correct output and you nod and move on.
I spent a day talking about this with people on Bluesky, and the responses revealed the pattern across domains. Martin Kleppmann argued that better UX will bridge the gap. He imagines AI asking questions instead of requiring formal specs, eliciting the human's intent conversationally. But the human still needs to know what they don't know to answer those questions. A non-expert asked "should this be allowed?" says yes without seeing the three edge cases they just endorsed. The gap doesn't disappear. It moves to a different surface. Hillel Wayne pointed out that AI for formal verification helps people who already know formal verification. The people building Lean agents are the same people who could already write Lean proofs. The tool makes them faster. It doesn't make them more numerous.
The historical precedent is instructive. When the HP-35 scientific calculator arrived in 1972, it was heralded as a democratizing force. No more slide rules. No more laborious manual calculation. Anyone could compute. What actually happened was that engineers got faster. The barrier to entry for complex calculation didn't disappear. It rose. The people who thrived were the ones who already understood what the numbers meant, who could spot when an answer was wrong because they had done enough manual calculation to develop intuition. The calculator amplified the expert. It did not create new ones. The same pattern held with spreadsheets, with CAD software, with statistical packages. The tool that removes mechanical effort also removes the friction that builds judgment. The gap doesn't close. It widens.
There is one group that might benefit disproportionately: the forgotten middle. The engineer with three to five years of experience. Not expert enough to go 10x solo, but enough battle scars to evaluate AI output. They have enough judgment to reject bad suggestions but not enough raw speed to produce what the AI produces in seconds. For them, the tool is a genuine multiplier. They can evaluate faster than they can generate, and the AI fills the generation gap. The novice can't evaluate. The expert can generate and evaluate. The middle can evaluate but not generate quickly. AI plugs the hole. But this is not the democratization story. This is a story about a specific expertise threshold, not a general leveling.
The economic consequences are sharp. If AI tools multiply existing expertise, the return on investment concentrates where talent already runs deep. A startup with three staff engineers gets more value per AI dollar than one with twenty juniors. The tool doesn't fill talent gaps. It widens them. This inverts the democratization pitch at the organizational level. The companies that benefit most from AI are the ones that need it least. It is an expertise amplifier, not an expertise substitute.
A concrete prediction, three years out: coding interviews will shift from generation to judgment. Because code production is becoming cheap, the bottleneck becomes the ability to evaluate quality. Interviewers will hand candidates AI-generated code and ask "is this correct?" and "what would you change?" That skill cannot be prompted into existence. It is built through years of debugging your own broken code. The interview format will expose the expertise gap more clearly than any take-home project ever did.
The democratization narrative was always better suited to marketing than to reality. Every power tool in human history has followed the same arc. The tool that promised to make everyone a master made masters faster and everyone else more obviously dependent. AI is not special in this regard. It is just the latest instance of a very old pattern. The only difference is the speed at which the gap becomes visible, and the fact that the tools themselves erase the evidence of what the gap is built on. You cannot learn from mistakes you never get to make.