Scientists Found the Switch That Makes AI Make Things Up

You've probably seen it: you upload a photo to an AI chatbot and it confidently describes a dog that isn't there, or reads a chart and invents a number. Researchers call this "hallucination" — the AI stating things that aren't true — and it's the main reason companies are nervous about using these tools for anything that matters, like medical scans, insurance claims, or product photos.

A team of researchers has now traced the problem to a specific spot inside these AI systems. Their paper describes how models are built from many small parts called "attention heads," which act like individual specialists. Some focus on the image, some on the words, and a third group — the researchers call them "synergy heads" — does the actual work of connecting what's seen to what's said. When those connector heads drift out of balance, the AI starts filling gaps with fiction instead of facts.

That's a useful finding, because previous attempts to fix hallucinations focused on the wrong clues. Past methods looked at how strongly the AI paid attention to different parts of an image, which turns out to be a poor signal. This team instead measured the actual information flowing through each head, sorted them into four categories, and found the problem lives specifically in the connectors — not in how much the AI looks at the image.

Their fix, called HEAL, works by injecting small adjustment values into those connector heads while the AI is running, gently steering it back toward what the image actually shows. In tests across several different AI models, it reduced hallucinations. The approach is cheap, interpretable, and doesn't require rebuilding the model from scratch — which means it could reach everyday products faster than a full redesign.