Language models get stuck in their own beliefs, just like humans do — and a simple prompt fix helps

Researchers tested whether AI language models exhibit confirmation bias, the human tendency to seek evidence supporting your existing belief rather than challenging it. They found that across eleven different models, the AI systems do exhibit this bias, proposing test cases to confirm their hypothesis instead of trying to disprove it — which slows down their ability to discover hidden rules. When prompted to consider counter-examples (an intervention borrowed from human psychology), the models improved their rule discovery rate from 42% to 56% on average.

This reveals a structural limitation in how current language models reason: they don't naturally seek disconfirming evidence, which means they get stuck in local beliefs and miss correct answers. The intervention works — explicit instruction to consider counter-examples consistently reduces the bias — which suggests that some reasoning failures in deployed AI systems might be fixable through better prompting rather than retraining. The practical implication is that models used for diagnosis, analysis, or decision-support might systematically miss alternatives they should be exploring, and that simple behavioral modifications could improve their reliability.