Researchers at the Korea Advanced Institute of Science and Technology have developed a way to fix errors that occur when AI converts natural-language questions into SQL database queries without rebuilding the whole command. The technology, called SafeQL, identifies and corrects only the wrong parts using database error information and stored data. It was implemented inside PostgreSQL and tested on Bird and Spider benchmarks, reducing token use and repair time while improving execution accuracy.

Database and analytics companies are rolling out text-to-SQL tools that let users query databases in natural language, but experts warn non-specialists should be cautious. Toronto University computer science professor Nick Koudas said systems can generate queries that are grammatically correct but not what business users intended. He put current accuracy at about 80 percent, below human experts at around 93 percent, and said company-specific data and jargon are hard for LLMs without extra training.