Anti-LLM Sentiment Considered Harmful

AI code review is more a light floor raiser. It’ll catch a lot of oversights and will sometimes call out lack of following convention. It won’t raise the ceiling of your code. I’ve never seen it meaningfully suggest improved abstraction usage - new or standard (typeclassopedia). Let alone architecture etc. That’s in-line with the code it generates. AI Haskell feels very “unrolled” to me. Both code I’ve generated and AI PRs I’ve received. Maybe with enough context and prompting and skills and tools it can improve. But I just haven’t seen it think to use stuff like Align or lazy State or even more advanced Semigroup/Monoid techniques. And it does take constant effort to grok code and, as. reviewer, call out the potential for these improvements. It’s extra important for juniors / Haskell newbies. It’s even extra effort to read AI code generated by those sorts. But it’s necessary effort in 2026. Because accepting AI PRs because they work, speed things up, otherwise make number go up is easy. But the downsides compound over time. In the context of my comments above, your codebase becomes closer to bytecode meant for LLM consumption. That’s bad.