Google Researchers Find a Way to Make AI Admit It's Guessing — Instead of Making Things Up

Google researchers have introduced a technique called 'faithful uncertainty' that trains large language models to express when they are unsure, offering a best guess rather than confidently stating a falsehood. The approach is designed to reduce hallucinations — the well-known problem where AI models invent facts — while still giving useful responses. The research was published in a new paper.