AI Writes Fluent But Factually Wrong Docs — Here Is How to Fix That

Large language models tasked with generating technical documentation tend to produce well-structured output that contains fabricated details, such as non-existent parameters or incorrect retry logic, because they pattern-match code to similar libraries rather than reading it precisely. The author argues that wrong documentation is more harmful than no documentation, as it discourages developers from consulting the source code. The proposed fix is to stop asking AI to invent documentation and instead ask it to transform artifacts that already encode ground truth, such as type signatures, passing tests, OpenAPI specs, database migrations, and git history. Constraining the model to these inputs — and instructing it to write 'UNKNOWN' for anything not present — significantly reduces hallucination. The author also recommends scoping documentation updates to individual diffs rather than whole-repository passes, making AI-generated suggestions small enough to review meaningfully.