Professional liability
Hallucination has become a professional liability claim.
Across the seventeen European cases we have recorded, professional liability is the best represented line. The mechanism is always the same.
A professional — lawyer, auditor, researcher, journalist — asks a generative system to produce or source a piece of content. The system invents references: fabricated citations, case law that does not exist, preparatory works that were never written. The professional does not check. The document goes out.
Our dataset records four files of this kind between 2025 and 2026, in legal services, audit, academic publishing and the press. The recorded consequences run from retraction of the article to professional negligence found by a court.
What stands out is the level of autonomy involved: 2 or 3 on our scale — cognitive assistance, content generation. No robot, no automated decision, no agent executing anything. Today's most frequent risk comes from the least autonomous uses.
The conclusion for an underwriter is counter-intuitive: loss frequency does not follow technical sophistication. It follows the absence of human control over the output. A firm that reviews everything does not carry the same risk as a firm that signs whatever the machine proposes, with the same tool.
This is why we do not score a model. We score a deployment: who checks, at what point, and what happens when nobody checks.
Source: European Agentic AI Loss Database, compiled by Garenzia from the public OECD AIM screening and the AI Incident Database.