Researchers find AI is pretty bad at debugging—but they’re working on it

A recent study explored using AI models to debug code, achieving a best-case success rate of 48.4%. This approach shows promise but highlights limitations in the models' understanding of tool use and lack of specialized training data. Researchers plan to develop info-seeking models to improve performance. While AI coding tools can sometimes create basic applications, they often produce buggy, insecure code. The goal remains to create AI agents that assist human developers rather than replace them entirely.