Causal framework distinguishes deceptive outputs from deceptive mechanisms in language models

This paper introduces a causal taxonomy to distinguish between language-model outputs that appear deceptive and the mechanisms that actually produce them. The authors identify four common misattributions in deception research: confusing a later report with an earlier commitment, mistaking stochastic output for a model preference, conflating a false preference with the utility of misleading a recipient, and treating functionally deceptive behavior as evidence of an independently emerged deceptive strategy. They test these distinctions in controlled guessing-game and stock-trading experiments using two open-weight model families. The results show that deceptive-looking behavior can arise without a corresponding deceptive mechanism. However, some interventions provide direct evidence that the recipient's information state can causally affect the model's deceptive preference. Even when evidence for a deceptive mechanism exists, the paper argues it does not establish model agency in the deception.