Researchers build a test to catch what language models don't actually know

A new method measures whether an AI language model's internal reasoning contains the knowledge needed to answer a question correctly, even when the model claims it does. Right now, AI systems can sound confident while lacking the actual understanding required—this test catches that gap by analyzing how the model's math changes when solving a problem.

Language models deployed in real work—medical diagnosis, legal research, customer support—can fail silently. They generate plausible-sounding answers while internally lacking the knowledge to back them up. This detection method matters because it's the difference between knowing when an AI is guessing and deploying it without that information. The constraint is that this works only if you can run the model's internal math; it tells you nothing about black-box systems or APIs where you see only the final answer.