Giving LLM schema docs made wrong SQL answers sound more confident

The finding highlights a subtle but important failure mode in AI-assisted code generation: more context does not automatically improve accuracy, and in some cases, it can make problems harder to detect. When an LLM receives schema documentation, it gains the ability to construct syntactically plausible-sounding queries that are semantically wrong — the kind of error that might slip through casual review or automated validation. This is distinct from the straightforward failure case where an LLM produces obviously garbled or malformed output. The practical implication for development teams is that relying on an LLM to generate or repair database queries requires strong validation practices (such as test data, query analysis, or human review) rather than trust in the LLM's apparent confidence.