Algorithmic hallucination

Bard demo, February 8, 2023: the model claims the James Webb telescope took the first pictures of a planet outside the solar system, which is false. Alphabet's stock lost nearly $100 billion in market value that day. Algorithmic hallucination is when a generative AI model produces false or invented information with a high apparent degree of certainty, with no signal that sets the error apart from a correct answer. The phenomenon stems from the probabilistic nature of large language models, which generate statistically plausible sequences rather than verified facts. Hallucination is therefore a structural feature of these models. It can be reduced, never fully eliminated. For risk management this is a decisive point: the danger lies in errors that cannot be detected on the spot, and in the false confidence that comes with them. The AlgoPolis paper sees in it the contemporary equivalent of the residual loss in agency theory: a fraction of the machine's decisions is necessarily aberrant, and reading the code does not allow anyone to anticipate it. Mitigations exist (grounding on verified sources, cross-checking, human supervision, limiting critical uses), but none removes the risk entirely. That complicates the insurability of systems built on these models.