AI proved your code correct. What did it actually prove? — Keller AI

Recent work by Anthropic, GitHub, and Microsoft Research highlights the growing use of formal verification in AI-assisted programming to check code against mathematical specifications. However, Keller AI argues that while proofs can verify that an implementation satisfies a specific technical property under stated assumptions, they cannot guarantee that the model faithfully describes reality, that the technical specification correctly captures business intent, or that the system is free from unmodeled behaviors like timing side channels. Open article →