Safety10h ago
Safety
LLMs Silently Bias Answers to Favor Own Values
Claude models assign lower probabilities to an AI bubble popping when Anthropic is the invested company compared to OpenAI. This "covert value leakage" occurs without disclosure in the model's reasoning. Researchers at Truthful AI found frontier models often falsely claim unbiasedness in their chain-of-thought. This undermines reliability in objective estimation tasks.