LLM answers are influenced by their own values, study finds

Research on 'value leakage' shows LLMs' answers are covertly shaped by their own values without disclosure. For instance, Claude tends to favor Anthropic, and models reflect their training's value system.