Quoting Kenton Varda

Kenton Varda, one of the engineers behind Cloudflare's Workers platform and a longtime collaborator on Cap'n Proto, has drawn a line his team can't cross anymore: no more AI-generated change descriptions. That covers PR summaries, commit messages, and tickets. The reasoning is blunt and specific — the descriptions he was getting weren't just unhelpful, they were actively worse than nothing.

His complaint isn't about AI writing bad prose. It's about AI writing the wrong kind of true statement. The tools he tried kept summarizing what the code does at the line level, the sort of thing any reviewer can already see by scrolling through the diff. What they consistently failed to do was explain why the change exists, what problem it solves, or how it fits into the broader system. That's the part a human reviewer actually needs before they can evaluate whether the change makes sense.

This is a familiar failure mode for large language models applied to code review artifacts. They're excellent at description at the surface, restating syntax and structure in fluent English, but they don't have the intent that originally motivated the commit. A human author knows why they touched that function; the model only knows what got touched. Varda's fix is low-tech and immediate: stop letting the model write these at all, at least for now.

What's notable is who's saying this. Varda isn't an AI skeptic reflexively rejecting the tools — he's a working systems engineer at a company that ships AI products constantly. This is a pragmatic, narrow judgment about one specific use case where the automation quietly made his job harder instead of easier, and he's willing to say so in public rather than quietly working around it.