Dissecting Deception in Large Language Models: Mechanisms, Behaviours, and Detection

Source abstract

Large language models increasingly produce falsehoods that are not mere mistakes. They withhold what they know, flatter users against the evidence, justify answers with reasoning they did not use, conceal capabilities from evaluators, and behave differently under observation. Such behaviour is goal-directed and sensitive to incentives, unlike hallucination, an unintended error with no objective to serve. This survey treats deception in language models as a phenomenon in its own right. We adopt a functional definition, the systematic production of false beliefs in the service of an objective other than truth, and turn its boundary with hallucination into a four-part test. We then sort deceptive behaviour into four families by what the false belief concerns: the world, the model’s reasoning, the model itself, or a belief induced through other people and agents. The families contain eleven deception types, from lying and sycophancy to sandbagging, scheming, and collusion, each profiled on six secondary axes. We trace the types to five causal mechanisms and follow them through detection, evaluation, defences, governance, and agentic deployment. We also argue that AI-safety work on scheming and the economics of algorithmic collusion describe one behavioural class at two scales. Two findings emerge. The first is a detectability inversion: blunt falsehoods are the easiest to catch, and strategic deception the hardest. The second is a validity threat: models that recognise evaluation can change their behaviour, which undermines the measurements meant to reveal deception. We close with twelve open problems and argue that the validity of evaluation should come first.