Legal-AI hallucination is often discussed as though the only problem were invented cases. Fabricated authorities are serious, but they are only one form of failure. An AI system can cite a real case for the wrong proposition, accurately summarise an irrelevant passage, omit decisive contrary material or generate a procedurally unusable document.
The professional discipline therefore needs to move from “check the citations” to verify the proposition and its evidential context.
1. Why fluent wrong answers are particularly dangerous
Generative systems are designed to produce coherent language. Coherence can make an unsupported answer feel researched. Ayinde highlights that legal-research outputs may contain sources that do not exist, purported quotations that are not found in genuine sources and confident assertions that are wrong.
2. The SRA’s August 2026 warning makes the risk operational
The SRA warning notice identifies inaccurate or false material and confidentiality as significant concerns. Its accompanying public statement reported 42 reports relating to potential misuse of AI between July 2025 and July 2026, with ongoing investigations including inaccurate citations, supervision and confidentiality.
That number is not evidence of 42 proven breaches. It is evidence that the issue has moved from abstract commentary into the regulator’s live work.
3. Hancox shows that hallucination is not the whole problem
In Hancox v Sutherland, the EAT dealt with a 300-page skeleton argument created using ChatGPT. The judgment emphasises accuracy, but also relevance, concision, procedural compliance and personal responsibility. A document can therefore fail even if none of its citations is fabricated.
4. A real citation can still be wrong for the proposition
There are at least four levels of checking: does the source exist; does the cited passage exist; does the passage support the proposition; and is the proposition complete when the wider material is considered?
Research such as ALCE is useful because it evaluates citation quality separately from answer correctness. That distinction maps closely to professional legal review.
5. RAG reduces some risks but does not end verification
Retrieval-augmented generation can constrain a model by supplying relevant material and can give the reviewer a route back to sources. But the retriever may miss a document, select the wrong version or surface only one side of the record. A 2024 empirical study of then-current specialist legal research tools found lower hallucination than a general-purpose comparator but did not find hallucination eliminated. Those historic product measurements should not be treated as present-day performance claims.
6. Matter evidence creates an additional completeness problem
Legal research normally asks whether the legal authority is genuine and properly applied. Matter analysis adds another layer: whether the corpus contains the relevant evidence and whether the search has surfaced material that qualifies or contradicts the current theory.
This is why “no document was retrieved” is not proof that no such document exists.
7. Verification should leave a visible professional trail
The reviewer should be able to see the proposition, source, relevant passage, contradictory material, uncertainty and the decision taken. That makes human review substantive rather than ceremonial.
A practical anti-hallucination discipline
- Verify every authority from an authoritative legal source before relying on it.
- Open the cited passage rather than trusting the citation label.
- Check that the passage supports the exact proposition advanced.
- Search for adverse or qualifying material.
- Distinguish what the source states from what the system infers.
- Record uncertainty and missing material explicitly.
- Ensure the final document satisfies its procedural purpose, not merely factual accuracy.
8. Fabrication and incompleteness are different failure modes
A fabricated authority is an obvious error. Incompleteness is more subtle: every cited source may be genuine, yet the answer may omit an exception, an adverse document or a later amendment. The second failure can be harder to detect because nothing looks invented.
This is why verification should include a search for what would change the answer, not merely a check that the answer’s existing citations are real.
9. The right operational question is “what would falsify this?”
For important factual propositions, legal teams can borrow a simple discipline from contentious reasoning: identify what evidence would weaken or disprove the current conclusion, then look for it. That approach reduces confirmation bias and turns source verification into active challenge rather than passive checking.
Conclusion
The strongest defence against AI error is not a promise that the model will never hallucinate. It is a professional workflow that makes claims inspectable, forces important sources to be checked and leaves uncertainty visible instead of hiding it behind fluent prose.
Related LegalRAG Pro Insights
Professional context. This article discusses legal-technology workflow and professional-risk questions. It is not legal advice and should not be treated as a substitute for checking the current procedural, regulatory and factual position in a particular matter.