AI models fail quietly under pressure — new test reveals which ones catch their own mistakes

Researchers built a test that measures whether AI language models stay accurate when information gets degraded or when someone deliberately tries to trick them, rather than just testing them under ideal conditions. It turns out the smartest models often fail worst, and a model's ability to catch its own errors is the only thing that predicts whether it will stay reliable under stress.

Until now, AI benchmarks test models on clean questions with perfect information. They don't tell you what happens when a model encounters incomplete data, adversarial prompts, or real-world degradation. This test exposes that brittleness. The practical consequence is that deployment decisions for AI in hospitals, courts, or financial systems have been made blind to a critical failure mode. A smaller model that catches its own mistakes might be safer than a flagship model that hallucinates with confidence when pressed.