Legal AI Verification Has to Go Beyond Checking Citations

Checking whether a case exists is necessary. It is not enough.

A real case can be cited for a proposition it does not support. A quotation can be almost right while omitting a decisive qualification. A summary can faithfully describe three pages of a record and miss the fourth page that changes the answer. An AI system can also invent facts, not just law.

The practical lesson from the latest sanctions decisions is broader than “look up the citations.” The lawyer responsible for the work must be able to trace each important claim back to an authoritative source and make an independent judgment about what that source actually proves.

The risk has moved beyond fictional cases

The early warnings about legal AI focused on nonexistent authorities. Courts are now encountering a wider range of failures: fabricated quotations, distorted holdings, inaccurate descriptions of real cases, and factual assertions that do not appear in the record.

The American Bar Association’s 2026 discussion of generative AI in discovery review describes matters involving invented witness statements and false references to deposition testimony. Its recent review of court enforcement and the duty of candor also emphasizes that responsibility does not disappear because work came from AI, another lawyer, local counsel, or a subordinate.

That matters because existence checking catches only one failure mode.

Six different questions belong in the review

1. Does the authority exist?

Open the case, statute, rule, regulation, or administrative material in a reliable source. Confirm the citation, court, date, and status. Never treat a link generated by an AI system as proof that the linked material is authentic or current.

2. Does it support this proposition?

Read the relevant passage in context. Check whether the language is part of the holding, dicta, a party’s argument, a dissent, or a quoted lower-court opinion. Confirm that the procedural posture and jurisdiction make the authority useful for the proposition being asserted.

A source can be real and still be wrong for the sentence.

3. Is the quotation exact?

Search the source for the quoted language. Compare punctuation, omissions, alterations, and surrounding text. A polished paraphrase placed inside quotation marks is still a false quotation.

4. Is every factual assertion in the record?

For matter-grounded work, identify the document and location supporting each material fact. Distinguish testimony from allegation, and allegation from established fact. Record uncertainty when sources conflict.

This is where a legal AI system should be willing to say “not found in the supplied materials.” A complete-looking answer is not more useful when it silently fills a gap.

5. What material has been omitted?

Ask what could make the proposed conclusion wrong. Look for contrary authority, exceptions, later history, missing evidence, unfavorable testimony, procedural limitations, and facts that have not yet been established.

An answer can contain no false sentence and still be dangerously incomplete.

6. Can the responsible lawyer explain the reasoning?

The final test is not whether the prose sounds professional. It is whether the lawyer can explain why the sources support the conclusion, what remains uncertain, and what judgment was exercised.

ABA Formal Opinion 512 frames generative AI use through existing duties of competence, confidentiality, communication, supervision, candor, and reasonable fees. None of those duties can be discharged by accepting a green “verified” badge without understanding what the verification covered.

A source-to-claim review table

For consequential work, a simple table can make review much faster:

Claim in the draft	Source and location	What the source establishes	Limits or contrary material	Reviewer decision
Material fact	Record document and page	Direct testimony, exhibit, allegation, or inference	Conflicting evidence or missing foundation	Use, qualify, or remove
Legal proposition	Authority and pinpoint	Holding and jurisdiction	Later history, exception, or adverse authority	Rely, distinguish, or reject
Quotation	Exact source location	Verbatim language in context	Omitted language	Keep, correct, or paraphrase
Recommendation	Supporting law and facts	Reasoning connecting both	Assumptions and practical uncertainty	Approve or revise

This does not need to become a permanent exhibit for every routine email. The depth should match the consequence. A filing, formal opinion, dispositive motion, settlement recommendation, or advice affecting a client’s rights deserves a more reproducible record than an internal brainstorming note.

Separate the roles that are easy to confuse

One person or system can perform all of these functions, but the work is safer when they are treated as separate passes:

1. Research: locate potentially useful material.
2. Analysis: connect the material to the question.
3. Challenge: search for what the analysis missed or overstated.
4. Verification: compare claims with primary sources and the record.
5. Synthesis: resolve disagreements and state remaining limits.
6. Approval: the responsible lawyer decides what may be used.

Asking the same model to “double-check” its own answer can help, but agreement with itself is not independent proof. A challenge pass should have an explicit job: find unsupported claims, contrary authority, factual gaps, and alternative interpretations.

What software should show the lawyer

A legal AI product should make the review easier, not merely generate more text. Useful signals include:

• the exact source set supplied to the system;
• a source and location for each material claim;
• a distinction between record fact, allegation, inference, and model suggestion;
• citation existence and proposition-support checks as separate results;
• contrary authorities and unresolved questions;
• disagreement between analytical roles;
• changes made after human review; and
• a record of what the lawyer approved.

The product should also say when a check could not be completed. “Could not verify” and “verified false” are different results. Both are different from silence.

Our interest in this question

Usus is being built around supervised legal work rather than autonomous answers. Our current research and deliberation work separates analytical roles, preserves disagreement, records consent before information is sent to an external model, and performs several checks after synthesis.

It is still in development. We are not treating the existence of multiple roles as proof of better legal work. Before making comparative claims, we need blind evaluations against a strong single-model workflow and clear evidence about which checks actually catch meaningful errors.

The aim is straightforward: legal AI should help a lawyer reach the evidence faster and see the weaknesses earlier. It should not make unsupported confidence easier to publish.

A final pre-use checklist

• I opened every authority on which the work materially relies.
• I confirmed that each authority supports the proposition stated.
• I checked quotations against the original text.
• I tied important factual claims to the record.
• I looked for adverse law, missing facts, and important exceptions.
• I know what the AI system could not verify.
• I can explain the conclusion without relying on the system’s confidence.
• The responsible lawyer, not the software, made the final decision.

That is a more demanding standard than checking citations. It is also much closer to the review lawyers have always owed their clients and courts.