Hallucination

Output produced by a large language model that is presented as factual but is not grounded in real evidence — fabricated citations, invented quotations, non-existent companies, made-up statistics, or claims about reality that the model cannot support. The failure mode is structurally inherent to next-token prediction trained on noisy data, not a bug to be patched.

Hallucination is the failure mode the 'Mata v. Avianca' attorneys discovered in 2023 when their AI-research tool fabricated six legal citations. It remains the dominant 2026 enterprise risk for agent-mode deployments where the agent's output reaches a regulator, a counterparty, or a customer-facing surface. The defensive primitive is grounding: retrieval-augmented generation against a verified corpus plus citation-with-source-link enforcement at the policy layer. Models do not stop hallucinating; deployers stop letting hallucinations exit the system unverified.

Related frameworks

• MTTD-for-Agents →
• GAUGE →

Articles that analyse this term

• Agent incident response: the six-step playbook for when an autonomous-AI deployment breaks production
• Agent observability stack: the four layers production agentic-AI actually needs (and what each one misses)
• The unverified citation chain: where enterprise AI decisions actually come from

Primary sources

• Stanford HAI. AI Index Report 2026 — model evaluation
• Court Listener. Mata v. Avianca — sanctions order (June 2023)