Evan X. Merz says AI-generated code can look fine in a pull request while becoming harder to operate and change. He describes teams where generating code quickly did not make review or maintenance faster. This is an engineer’s account, not a measured comparison of AI-written and hand-written systems.

The risks he sees:

  • Tests that cover existing happy paths but miss real users and edge cases, even when line coverage looks good.
  • Extra helpers and abstractions that make a simple change harder to understand.
  • A stream of changes that outruns reviewers’ understanding of the system. Merz points to a mouseover effect coded with JavaScript events rather than a CSS :hover rule.

He suggests watching incident frequency and recovery time as codebases age, but offers no incident data establishing that AI caused an increase. Teams still need to explain, test and repair what they ship.

The 371-comment thread on Hacker News adds both verification ideas and counterexamples.

What the thread adds

  • peterbell_nyc — asks what “adversarial reviews and deterministic gates” were used to catch maintainability problems; reports that more rigorous checks helped with a messy AI-generated codebase.
  • PaulMorel — self-identifies as the author and says he wrote a tool at Meta to catch known quality issues, but suspects larger architectural brittleness can escape those checks. That is his explanation, not proof of the mechanism.
  • drooby — proposes commenting out behavior and checking that tests fail, then adding mutation tests (deliberately altered code to challenge tests) and tests of general properties. They report fewer incidents on their own team.
  • user43928 — reports maintaining a large AI-generated mobile app without the predicted long-term trouble; notes that it is not a SaaS platform, so it is a different setting from Merz’s.
  • jasode — points out that large codebases were fragile before AI and argues for applying the same quality gates to generated code.

HN handles are pseudonymous; the site publishes no per-comment scores. This is a slice ordered by HN’s ranking, not a consensus.