How do we verify AI legal output (no fake citations)? — CLO AI Chambers

When it comes up

When AI is used for research, drafting submissions, advice memos or board papers, and whenever outside counsel file AI-assisted work for you.

Your options

Spot-check

Fits when: Internal, low-stakes outputs used by the lawyer who reads the source anyway (for example, a contract summary for their own use).

Watch out: Not enough for anything relied on externally or by the business.

Full lawyer review

Fits when: Advice to the business, board materials, counterparty documents.

Watch out: Reviewers can trust fluent text; give them a checklist.

Full review plus recorded cite-check

Fits when: Anything filed with a court or regulator, or quoting legal authority to a third party.

Watch out: Takes time; build it into deadlines rather than skipping it.

Independent second check

Fits when: High-stakes filings, novel points of law, or work by a team new to AI.

Watch out: Costly; reserve for the highest-risk outputs.

How to decide

Set review levels by destination: internal use → spot-check; business or board → full review; court or regulator → full review plus recorded cite-check.
Define the cite-check: each authority exists, is cited correctly, is quoted verbatim, supports the proposition and is still good law.
Name the checker and record the check.
Check court rules for AI certification or disclosure.
Require the same of outside counsel in your guidelines.

Trade-offs

More verification reduces time savings; less verification risks sanctions, reputational harm and bad decisions. Matching review to destination keeps most of the value where the risk is low.

Real examples

In Mata v. Avianca (S.D.N.Y., June 2023), Judge P. Kevin Castel sanctioned two lawyers and their firm, with a $5,000 penalty, for filing ChatGPT-generated fake case citations and standing by them after they were questioned. Source: Wikipedia — Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023) ↗ · Reed Smith — ChatGPT artificial intelligence tool generates fake opinions (Mata v. Avianca, S.D.N.Y. June 22, 2023) ↗
In Ayinde v London Borough of Haringey [2025] EWHC 1383 (Admin), the High Court of England and Wales dealt with fictitious citations in two cases, made wasted-costs orders and referred lawyers to their regulators. Source: Ayinde v London Borough of Haringey; Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin) — National Archives ↗
In July 2025 three Butler Snow lawyers were disqualified from Johnson v. Dunn (N.D. Ala.) after filings with AI-fabricated citations, despite the firm's internal AI warnings. Source: Bloomberg Law — Butler Snow attorneys disqualified over fake AI case citations (Johnson v. Dunn, N.D. Ala., July 2025) ↗

What to check first

Where will this output go (internal, business, board, court or regulator)?
Does it contain legal authorities or quotations that must be cite-checked?
Who is the named reviewer, and is the check recorded?
Do the court's or regulator's rules require AI disclosure or certification?

Learn and use

Sources

Last reviewed 2026-10-01.