AI-Generated SQL Runs Fine But Can Silently Return Wrong Data, Dev Warns

A software developer has highlighted a subtle but serious risk with AI-generated SQL queries: they can execute without errors while still returning logically incorrect results. In a real-world example, an AI produced a complex query using window functions and CTEs that passed code review but used MIN instead of MAX, returning the earliest matching row rather than the most recent one. The mistake went undetected for two weeks after deployment, affecting live data for specific edge cases. The developer notes that leading large language models achieve only around 60% execution accuracy on complex natural language-to-SQL tasks, making direct sign-off unreliable. As an alternative, the author advocates for a step-by-step, verifiable approach to SQL generation where intermediate results can be inspected before a final query is produced.