Dernière mise à jour :
When generative AI produces incorrect or fabricated information that appears plausible but has no basis in the training data.
A failure mode of large generative models (text, image, audio) where the system confidently invents details - facts, quotes, references - that sound coherent yet are false. Hallucinations arise from the model’s probabilistic sampling and lack of grounding. Governance approaches include grounding on trusted knowledge sources, retrieval-augmented generation, calibrated confidence scores, and post-generation fact-checking layers to catch fabrications before release.
Exemple concret :
A legal-tech chatbot invents a case citation “Smith v. United Republic, 2021” when summarizing contract law. The firm integrates a citation-check service: after generation, the chatbot cross-references each case against an authoritative database and flags any unverified citations for human review, preventing reliance on bogus precedents.
