New AI Training Trick Makes Chatbots Lie Less About Photos

AI that can look at pictures, charts or video and talk about them is increasingly common — think tools that summarize a screenshot, read a receipt, or describe a medical scan. The problem is hallucination: the AI states something that simply isn't in the image, and it says it with total confidence. That's the failure people notice most, because it's hard to spot and expensive to trust.

A team of researchers says they found two specific reasons this happens during training. First, when a question is really difficult, the AI tends to get every practice attempt wrong, so its learning signal flatlines exactly when it needs help most. Second, when the AI is confidently wrong, the math of training makes those mistakes nearly invisible to correction — the predictions needing the most fixing get the weakest nudge. Their fix, DEEPO, does two things: it injects expert guidance on high-uncertainty questions, and it reweights the training so confident errors actually get corrected.

On VideoMMMU, a demanding test involving long videos, the method improved results by about 4 points, and the authors report the gain is statistically solid rather than random noise. Each of the two fixes helped on its own; combined, they worked better still. Importantly, the AI didn't get less accurate elsewhere or become unstable during training — a common trade-off when researchers try to reduce hallucination.

What to keep in mind: this is an academic preprint, meaning it hasn't yet been reviewed by other scientists, and the gains were measured on specific benchmarks rather than on the messy real world. There's no app or feature you can buy because of it. Still, the direction matters. As AI gets wired into customer service, legal research and health tools, a system that confidently invents details is far more dangerous than one that simply says it doesn't know.