Natural-language SQL breaks quietly when schema context goes stale

Natural-language SQL can still produce runnable queries after schema changes, but the results may be wrong if the model uses stale schema context.

The article warns that AI SQL tools need versioned schema metadata, drift detection, and clear provenance to avoid silent accuracy failures.

It recommends refusing or re-fetching context when schema versions, migration hashes, or catalog hashes no longer match.

This is a reliability issue for production database assistants, where outputs can look valid while relying on obsolete schema information.