AI models that reason well still stubbornly refuse to change their minds

Researchers built a test that shows what happens when you give an AI model a fact, let it reach a conclusion, then slightly change that fact and ask if the conclusion should change. It turns out that even strong reasoning models often stick to their original answer instead of revising based on the new evidence. This matters because real-world reasoning happens in moving situations where premises shift — medical diagnoses get new test results, legal arguments introduce new documents, business decisions face new market data.

The gap between 'can you solve this logic puzzle' and 'will you update when the puzzle changes' is a real problem in deployed AI. A model that reasons well under stable conditions but refuses to revise when evidence shifts is brittle in ways that don't show up in standard benchmarks. If an AI system is advising doctors or lawyers or engineers, it needs to do both things equally well. Right now, the tests we use only measure the first.