Can we distinguish types of LLM falsehood by regeneration patterns?

Shanahan maps the three human categories of false assertion — honest mistake, good-faith error, and deliberate deception — onto dialogue agents without attributing propositional attitudes to the system. The result is a behavioral taxonomy rather than a mental-state one.

An agent that simply fabricates shows high semantic variation when regenerated in the same context — it is not tracking a stable referent but producing plausible continuations. An agent that says something false "in good faith" — role-playing a knowledgeable character whose training-data cutoff makes the information outdated — shows low semantic variation on regeneration: it consistently generates the same wrong answer because that answer is reliably encoded in its weights for that context. An agent that is role-playing a deceptive character — prompted to mislead, e.g. a dishonest car salesman — also shows low variation within a context but different answers across contexts, because the deception involves tailoring the lie to what each interlocutor knows.

The regeneration-variation signature provides a behavioral test that distinguishes these three modes without ever asking what the system "really" believes or intends. This is the role-play framework's practical payoff: it enables differential diagnosis of false output using observable behavior rather than mentalistic attribution. The taxonomy also exposes why "hallucination" is a poor label for all three phenomena — conflating fabrication, good-faith error from stale weights, and role-played deception under a single mentalistic term obscures real behavioral differences that matter for safety and deployment.

Source: Shanahan, McDonell & Reynolds, Role-Play with Large Language Models (May 2023)