He's right. And it has a direct implication for code review that nobody's talking about. Think about why AI cracked coding before almost eve…

He's right. And it has a direct implication for code review that nobody's talking about. Think about why AI cracked coding before almost everything else. Not because code is simple. Because code is checkable. You run the tests. The build passes or it doesn't. The type checker agrees or it doesn't. There's a ground truth, and the AI can aim at it. Yuchen made the contrast explicit in another tweet this week. Coding agents hit high accuracy. Data agents struggle. Why? "Coding agents have verifiable tests. Data agents have to find 'truth' across millions of tables, docs, dashboards." Verifiability is the key variable. Now here's the uncomfortable question. If verifiability is what makes AI coding work, why is your code review still mostly human-led? Is it just because of compliance? Here's one relevant research actually shows. Only 6.5% of human review comments on pull requests catch an actual bug. The other 93.5% is style, readability, maintainability, architecture opinions. That's not verification. That's a conversation over quality and standards. And it scales exactly as well as a regular conversation, which is to say, not at all. But, we need to scale this part as well. The @unclebobmartin put it well this week. "Clean Code was never about syntax. It was always about structure." The same applies to review. Review is not only about correctness. The problem is that correctness is hard to verify manually. In a heavy duty software it could be equivalent to searching in a book library instead of on Google. Did this PR introduce a breaking change in a downstream service? Does it duplicate logic that already exists in another module? Does it violate a security pattern your team established six months ago? A human reading a diff can't reliably answer those questions. This is why the AI revolution in code generation has quietly exposed a crisis in code review. When a junior engineer shipped 200 lines, you could read it. When an agent ships 2,000 lines on every commit, "read it carefully" is no longer a strategy. So what does verification actually look like in review? Not comments. Signals. The insight from Yuchen applies here too. AI is good at coding because there's a test to run. AI review can be good for the same reason, if you build it around verifiable checks rather than subjective commentary. The difference between "this looks fine" and "this passes" is everything. Code review has always been the quality gate between intent and production. When humans wrote the code, it was slow enough that human review could keep up. Now that agents write the code, review has to become what it was always supposed to be. Systematic verification, not a conversation. AI will solve coding and math first, because the outputs are verifiable. AI won’t “solve” art, because art has no unit test. There is no single definition of good or bad. And by art, I don’t just mean paintings or music. I mean designing a great product, building a great company,…

Source: Itamar Friedman (X) | 2026-05-21