Shipping AI code nobody understands, and who vouches for it
The most revealing thread was 'How to deal with shipping AI-written code you don't understand?' Answers split. Some said they trust it like a colleague's work. One called it another layer of 'conceptual compression,' like ActiveRecord. One said product doesn't care and 30% of engineers will keep their jobs. The sharpest point: when the AI writes both the code and the tests, they can agree with each other and miss the same bug. The practical fixes people named were dev plans up front, asking the model to reuse existing libraries, and extra runtime observability and alerting.
The mood around it is mixed. 'It's still ok to enjoy programming' hit the front page. An earlier thread had a 60-year-old saying Claude Code reignited his passion, and another dev found the same workflow unfulfilling. A separate Ask HN proposed a flag for AI-generated articles, and the guidelines already say HN is for conversation between humans. People noted detection is unreliable and that anti-LLM stigma is already driving people off Lobsters.
The pattern: output is cheap, so verification and provenance are where the cost moved.
So what?
The product opportunity is in checking, not generating: independent test generation, runtime monitoring, and audit trails for AI-written code. Internally, treat 'who understands this code' as a real risk line item before the first production incident forces the question.
Read these
Ask HN: How to deal with shipping AI-written code you don't understand?
It's still ok to enjoy programming
Ask HN: Add flag for AI-generated articles
Tell HN: I'm 60 years old. Claude Code has re-ignited a passion