Artificial intelligence now writes code, drafts contracts, and produces reports. Work that once took years of training to do well arrives in a few keystrokes. Hence the spreading expectation that anyone can now be an expert. What actually happens on the ground is a little different. The software developer Aaron Brethorst recently put his finger on the gap.
Take software. The hard part was never typing the code. To build a payroll program, you first had to hold the rules of the work clearly in your head — Korea’s four mandatory social insurance contributions, tax-exempt allowances, how to prorate pay for someone who joins or leaves partway through the month. The code was little more than a transcription of that understanding. The real work was knowing the field; writing the code was only the last step of setting that understanding down. And this is not unique to software. Whether it is an accounting ledger or an engineering drawing, behind every finished output sat a mental model of the field that produced it.
AI split these two apart. A plausible-looking output now appears even without deep knowledge of the field. As a result, the bottleneck has moved. Where the question used to be “can you build it,” it is now “can you tell whether it is right.” Producing things got easier; deciding whether to trust what was produced got no easier at all.
This is where an asymmetry opens up between people. Someone who has reconciled payroll for ten years can glance at a statement the AI generated and know that those numbers could never be right under the rules. A clerk who has handled medical claims for years can read the billing codes alone and know that, filed that way, the claim will be denied. Someone fluent in programming but unfamiliar with the work, by contrast, has no way to catch a result that runs cleanly and looks fine but is subtly wrong. The trouble is that such errors are hard to see. The screen shows tidy numbers, and the mistake surfaces only much later, in a reconciliation or an audit. Only a person who knows in their bones what the right answer is can verify the result.
Here an interesting reversal takes place. If AI has made the ability to produce things cheap, that means the domain expert who was weak at coding can now have the AI generate the output directly. The machine supplies what they lacked, and what is left to them is the one thing the machine cannot: the eye for what is correct. For someone without that eye, on the other hand, no matter how quickly the AI turns out results, there is no way to judge whether they are right. And that eye cannot be bought in a few keystrokes. It accrues only from spending years inside the work itself.
Of course, glancing over a few outputs will not catch every error; a plausible wrong answer slips past experts too. Still, the broad direction is clear. The more a field’s correct answers are buried deep within it — law, accounting, clinical medicine, power-grid equipment — the more AI merely produces the output quickly without guaranteeing that it is right. What has actually become scarce in the age of AI is the eye of someone who has worked a field long enough to tell what is correct.