“Verifiable AI” is an attractive phrase, but it can easily overpromise. In litigation, verification is not a single checkbox. A lawyer may need to verify the identity of the source, the accuracy of an extract, the relationship between the source and the proposition, the scope of the material searched, contrary evidence and the professional conclusion drawn from all of it.
The more useful goal is therefore not an AI system that declares itself verified. It is a workflow that makes important propositions inspectable, challengeable and reviewable.
1. Start with a proposition, not an answer
Complex litigation questions are rarely resolved by one sentence. A proposition such as “the employer decided to dismiss before the consultation” contains several components: who made the decision, what counts as the decision, when it was made, which documents support that date, what contrary evidence exists and which legal issue the timing affects.
Breaking the answer into propositions makes verification tractable. It also reveals where the AI is quoting a source, summarising it or making an inference.
2. Preserve source identity and location
A useful result should lead the reviewer back to the actual document and, where possible, the relevant page, paragraph, message or passage. That reduces the cost of checking and helps prevent a document from being cited for a proposition it does not support.
Source identity does not prove correctness. It establishes an inspection route.
3. Distinguish provenance from evidential weight
A genuine source may still be weak evidence. It may be hearsay, retrospective, conditional, superseded or concerned with a different period. A contemporaneous email and a later witness recollection are both sources, but they do not necessarily carry the same evidential significance.
This is why source-linked systems should not collapse provenance into a confidence score. The professional still has to assess context and weight.
4. Treat retrieval adequacy as a separate question
An answer can accurately summarise every retrieved passage and still be incomplete if the decisive document was not retrieved. Retrieval adequacy, corpus completeness, generation accuracy, evidential assessment and legal conclusion are different stages.
Research on citation-generating and retrieval-augmented systems reinforces this distinction. ALCE evaluates correctness separately from citation quality, while the 2024 Magesh study showed that specialist legal retrieval systems in the versions tested could reduce but not remove hallucination. Those historical figures should not be treated as current product benchmarks; the methodological lesson is that retrieval does not end the verification problem.
5. Make contradiction visible
Contentious work requires the system to show material that qualifies or damages the current proposition. Practice Direction 57AD’s treatment of adverse material offers a useful procedural analogy: evidence that points the other way matters.
A verification workflow should therefore ask not only “what supports this?” but “what contradicts it?”, “what narrows it?” and “what expected material has not been found?”
6. Preserve uncertainty
Dates may be exact, approximate or inferred. A document may suggest knowledge without proving it. A search may not find an attachment. “Unable to determine from the available material” can be a professionally superior result to an invented conclusion.
7. Keep professional review distinct from machine assistance
The SRA’s August 2026 warning notice emphasises that regulated individuals remain accountable for their work regardless of how it was prepared. That makes it useful to distinguish an AI-assisted finding from the professional decision that follows it.
A review state can record whether the source was checked, whether the proposition was accepted, qualified or rejected, and which version of the matter analysis is approved for use.
A practical verification stack
- Corpus: what material was available?
- Retrieval: what was actually surfaced?
- Provenance: which source and passage support the finding?
- Interpretation: what is stated and what is inferred?
- Challenge: what qualifies, contradicts or remains missing?
- Professional decision: what does the lawyer accept and why?
8. Verification should be proportionate to consequence
A minor internal categorisation does not require the same scrutiny as a proposition that will determine the pleaded case. The verification burden should rise with the consequence of error, the sensitivity of the material and the degree of inference involved.
This supports a tiered approach: orientation may use lightweight review; issue analysis may require passage-level checking; court-facing or advice-critical propositions may require authoritative sources, contradiction review and explicit professional approval.
9. Verifiability should survive handover
A result is not truly reviewable if only the original user understands how it was produced. Matter knowledge needs to remain intelligible when work passes to another fee earner or to counsel. That is why persistent source relationships and review state matter.
Conclusion
Verifiable litigation AI should not mean that software certifies its own answer. It should mean that the lawyer can reconstruct the evidential path, challenge what the system has done and keep the professional conclusion visible.
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.