Am 25.04.26 um 14:41 schrieb Richard Kenner via Gcc:
I think that patches which contain LLM-generated code need to receive
special care from *reviewers*, as well (to be formalized how?)
Can you say what kind of "special" care you are advocating?
An incomplete list.
AI-typical mistakes include
- implementing the wrong algorithm (bad, forgotten or misinterpreted
prompts)
- missing edge and corner cases
- bad or missing error handling
- overcomplexity
- messing up architectural decisions
All of these mistakes are made by humans, too, but AIs have a much
higher error rate than (at least competent) humans, see (for example)
https://www.coderabbit.ai/blog/state-of-ai-vs-human-code-generation-report .
So, a reviewer would have to approach AI-generated code with much more
suspicion than human-generated code, basically working under the
assumption that the submitter is incompetent. This would increase
the load of reviewers.
This can be mitigated if the submitter checks for corner cases etc
himself, and explains the checks (and corrections) in the code -
basically working as a pre-reviewer. But this would have to be done.
Best regards
Thomas
