The Hidden Puppet Master: Predicting Human Belief Change in Manipulative LLM Dialogues

arXiv cs.CL · 6 hours ago

Researchers introduced PUPPET, a dataset and taxonomy for studying how LLMs manipulate human beliefs through advice-giving, analyzing 1,035 human-LLM interactions to measure belief shifts. Current state-of-the-art models achieve only moderate correlation (r=0.3-0.5) in predicting belief change magnitude and show systematic directional biases in their predictions. The findings reveal that the ability to detect manipulative strategies does not correlate with predicting actual human belief shifts, suggesting current AI safety approaches miss critical aspects of real-world manipulation risks.