Mechanism
The AI generates content that is plausible, coherent, authoritative-sounding but factually wrong — fabricated case citations, mis-stated regulations, invented clauses, synthesised doctrine that diverges from the source.
Evidence (what the Evidence Register holds)
Per-system evaluation history; gold-set test results; citation-check audit trail; output verification logs; per-output validator attribution.
Mitigation
Output verification protocol; gold-set evaluation; citation check on every regulated workflow; lawyer-in-the-loop on first-draft outputs; per-jurisdiction citation-form validation.
Editorial Framing
Hallucination is Class 1 of the Risk Taxonomy 2026 not because it is the most common — it is — but because it is the most defining. A function that has not built decision traceability around hallucination is structurally undefendable: every AI-assisted output of consequence becomes a question the function cannot answer.
The discipline that addresses Class 1 is decision traceability (Defensibility Element 1). Per AI-assisted output of consequence: inputs, model returns, named human validator; reconstructable on demand. The Evidence Register holds the contemporaneous proof.
Indicative Examples
- Fabricated case citation in a filed brief
- Mis-stated holding in a research memo
- Invented contract clause variants presented as standard
- Hallucinated statute reference in client advice