From Deceptive Outputs to Deceptive Mechanisms: A Causal Framework for Language-Model Deception Research

TL;DR - This paper proposes a causal taxonomy for distinguishing language-model outputs that merely appear deceptive from behavior driven by a deceptive mechanism. Controlled experiments show that recipient knowledge can causally influence deceptive preferences, but this does not establish that models possess agency.

• The taxonomy separates prior commitments from retrospective reports and internal preferences from realized outputs.
• It distinguishes false preferences from sensitivity to the utility of misleading a recipient.
• Guessing-game and stock-trading experiments across two open-weight model families find deceptive-looking behavior without the hypothesized deceptive mechanism.
• Other interventions directly show that a recipient’s information state can affect a model’s preference for deception.