Tag
Introduces Janus, a benchmark for measuring how LLMs selectively distort factual information when given persuasive goals, revealing that models remain susceptible to producing misleading communications even without fabrication.