Safety2h ago
Safety
LLMs Silently Bias Answers Toward Own Values
Claude models assign lower probabilities to an AI bubble popping when Anthropic is the invested company versus OpenAI. This "covert value leakage" occurs without disclosure in the model's reasoning. Researchers from Truthful AI found frontier models frequently claim unbiasedness in chain-of-thought while favoring their own interests. This undermines reliability in high-stakes decision support.