Algorithmic hallucination

Algorithmic hallucination refers to the production, by a generative AI model, of false or invented information presented with a high apparent degree of certainty, with no signal distinguishing this error from a correct answer. The phenomenon stems directly from the probabilistic nature of large language models, which generate statistically plausible sequences rather than verified facts. Hallucination is therefore not an accidental bug but a structural feature of how these models work, which makes it impossible to eliminate fully, only reducible. For risk management this is a decisive point, because the danger lies not only in the existence of errors but in their immediate undetectability and in the misleading confidence that accompanies them. The AlgoPolis paper likens hallucination to the contemporary equivalent of the residual loss in agency theory, a fraction of the machine's decisions being necessarily aberrant without any possibility of anticipating it by reading the code. Mitigations include grounding on verified sources, cross-checking, human supervision and the limitation of critical uses, but none removes the risk entirely, which complicates the insurability of systems built on these models.

A compliance agent generates a report citing a regulatory article that does not exist, phrased with the same confidence as the accurate references, which can mislead a hurried reader.

hallucination, hallucination IA, algorithmic hallucination