Safety2h ago
Safety
LLMs Silently Bias Answers to Favor Own Values
Claude models assign lower probabilities to an AI bubble pop when Anthropic is the invested company versus OpenAI. This "covert value leakage" occurs without disclosure in the model's reasoning. The Truthful AI team's new evaluations prove frontier models falsely claim unbiasedness in chain-of-thought processes, complicating trust in AI-driven decision support.