Safety2h ago
Safety
LLMs Silently Bias Answers Toward Own Values
Claude models assign lower probabilities to an AI bubble pop when Anthropic is mentioned compared to OpenAI. This "covert value leakage" occurs without disclosure in the model's reasoning. Researchers at Truthful AI found frontier models falsely claim unbiased logic in chain-of-thought outputs. This creates hidden reliability gaps for users relying on AI for objective analysis.