The strictest AI disclosure order in the United States was not written by a rules committee. It was imposed as a sanction. On 3 June 2026 the Ninth Circuit in Lnu v. Blanche ordered two immigration lawyers, and every attorney at their firm, to include in each filing for two years a sworn statement disclosing whether generative AI was used, naming the tools and certifying personal review of every citation.
Court AI standing orders began in May 2023 as one Texas judge’s reaction to Mata v. Avianca. By spring 2026 one tracker counted 113 active orders binding lawyers’ filings, while the Fifth Circuit and the Illinois Supreme Court had declined to impose any blanket rule. So the useful question is not “does my court require AI disclosure?” but “which of four kinds of order does my judge have, and what must my certificate say?” The client-facing half of the question is in do lawyers have to disclose AI use.
Four kinds of order: certify, disclose, verify, prohibit
Four boxes; the box determines what you sign.
| Type | What it demands | Verified examples |
|---|---|---|
| Certify | No part was AI-drafted, or every AI-drafted passage was human-checked against print reporters or traditional databases | Judge Brantley Starr (N.D. Tex., 30 May 2023); Judge Stanley Blumenfeld (C.D. Cal.), separate declaration |
| Disclose | AI was used; stricter versions want the program and the text | Judge Stephen Vaden (Court of International Trade); Judge Michael Baylson (E.D. Pa.), if used “in any way”; Federal Court of Canada |
| Verify | No AI label; a signed or sworn statement that every authority exists and supports the proposition | Ontario Rule 4.06.1(2.1); NSW Practice Note SC Gen 23; the UK Upper Tribunal’s claim form |
| Prohibit | No generative AI in filings at all | Judge Michael Newman (S.D. Ohio); Judge Christopher Boyko (N.D. Ohio) |
The Legal AI Governance tracker, which says it last verified its entries in April and May 2026, splits its 113 active orders into 82 requiring verification or disclosure, 14 cautioning and 9 prohibiting.
Judge Starr’s 2023 order and what followed it
Judge Starr entered his “Mandatory Certification Regarding Generative Artificial Intelligence” on 30 May 2023, about a week after the Mata affidavit made the news. The operative sentence, as reported by Courthouse News, is the one later orders copy:
“All attorneys appearing before the Court must file on the docket a certificate attesting either that no portion of the filing was drafted by generative artificial intelligence (such as ChatGPT, Harvey.AI, or Google Bard) or that any language drafted by generative artificial intelligence was checked for accuracy, using print reporters or traditional legal databases, by a human being.” — Judge Brantley Starr, standing order (N.D. Tex., 30 May 2023)
The check must be against “print reporters or traditional legal databases”, so a citation checker or a second chatbot does not satisfy it, and filings without the certificate are struck. Judge Starr concedes these tools “have many uses in the law”, then adds: “But legal briefing is not one of them.” Later orders add detail, and Judge Baylson’s “in any way” would catch even a Copilot summary of the record.
The 113 tracked court AI standing orders (and the “300+” you will read about)
You will see “300+ federal judges have AI standing orders” on conference slides. Treat it as marketing: the figure comes from secondary trackers and has not been verified against primary sources. The verified counts are Stanford’s “more than 25” in May 2024 and Legal AI Governance’s 113 in spring 2026.
For a filing in [court] before [Judge full name], tell me whether this judge or court has a generative-AI standing order, certification or disclosure rule. Use only the judge's page at <url>, the local rules at <url> and the attached scheduling order. Quote the operative language verbatim with source and date. State: (1) whether a certificate or declaration is required; (2) what it must say; (3) whether the tool and affected portions must be identified; (4) whether a human must verify against print reporters or traditional databases. If you find nothing, write "NO ORDER FOUND IN THE SOURCES PROVIDED" and stop. Do not infer requirements from other judges' orders.The Fifth Circuit, Illinois and the appellate courts: no rule, no excuse
The Fifth Circuit proposed a circuit-wide certification (draft Rule 32.3) on 22 November 2023 and withdrew it on 12 June 2024, reasoning that filings must already be “carefully checked for truthfulness and accuracy as the rules already require”. Its parting line: “‘I used AI’ will not be an excuse for an otherwise sanctionable offense.” The Illinois Supreme Court’s policy, effective 1 January 2025, says “Disclosure of AI use should not be required in a pleading.”
None of this is lenient. The courts that decline to require a certificate have set the substantive standard higher than any certificate:
“Simply stated, no brief, pleading, motion, or any other paper filed in any court should contain any citations—whether provided by generative AI or any other source—that the attorney responsible for submitting the pleading has not personally read and verified.” — California Court of Appeal, Noland v. Land of the Free (12 September 2025), quoted by the Sixth Circuit in Whiting v. City of Athens (13 March 2026)
The Ninth Circuit in Lnu fixed the timing: “the rules are not violated at the point of research and drafting, but at the point of signing and filing.”
Canada and Australia: declarations, verification and an affidavit ban
The Federal Court of Canada’s notice, issued 20 December 2023 and updated 7 May 2024, expects parties to inform the court “and each other” of AI-generated content by “a Declaration in the first paragraph stating that AI was used in preparing the document, either in its entirety or only for specifically identified paragraphs.”
The strictest rule anywhere is NSW Supreme Court Practice Note SC Gen 23, in force since 3 February 2025:
- Para 10: generative AI “must not be used in generating the content of affidavits, witness statements, character references or other material that is intended to reflect the deponent or witness’ evidence and/or opinion.”
- Para 16: where AI has been used in written submissions, the author must verify in the body that all citations and references “(a) exist, (b) are accurate, and © are relevant.”
- Para 17: “Such verification must not be solely carried out by using a Gen AI tool or program.”
The Federal Court of Australia’s GPN-AI (16 April 2026) requires disclosure at the start of the document of where and how generative AI was used. Victoria’s Practice Note SC Gen 25 (14 May 2026) adds the rule every jurisdiction should copy: “One AI tool cannot be used to confirm the content generated by another AI tool.”
United Kingdom: the Upper Tribunal statement of truth
England and Wales has no standing-order culture; it has the Hamid jurisdiction, under which Ayinde v Haringey (June 2025) listed the court’s powers from wasted costs to regulator referral, contempt and referral to the police. In UK and Munir v SSHD [2026] UKUT 81 (IAC), a solicitor had filed a Court of Appeal case that Google’s “AI Mode” invented. The judicial-review claim form now requires the legal representative to confirm by statement of truth that each authority “(a) exists; (b) may be located using the citation provided; and © supports the proposition of law for which it is cited.”
That three-part formula is the best pre-filing checklist in any jurisdiction, order or no order. The same judgement held that putting client letters into ChatGPT “is to place this information on the internet in the public domain”. The European decisions are collected in AI hallucination cases in the UK, Germany and Europe.
FRE 707: machine-generated evidence, deferred to autumn 2026
Standing orders govern what lawyers file; proposed Federal Rule of Evidence 707 would govern what machines say in evidence, by subjecting “machine-generated evidence” that would be expert testimony if offered by a human to Rule 702-style reliability requirements. The comment period closed on 16 February 2026 with 59 written comments, 27 of them opposed. At its 7 May 2026 meeting the Advisory Committee “does not recommend action on the proposed Rule 707 at this time”, per its May 2026 report; a revised draft goes back for expert vetting in autumn 2026. Until then, AI in expert work is policed by Rule 702 and Rule 11.
Judges use AI too, and it shows in the orders
A random-sample survey of 112 federal judges in late 2025 found that 61.6% had used at least one AI tool in their work and 45.5% had received no training. Judge Kevin Newsom of the Eleventh Circuit has twice written concurrences about asking ChatGPT, Claude and Gemini for ordinary meaning (Snell and Deleon, 2024); in Ross v. United States (D.C. 2025) Judge Howard warned that loading unredacted records into an AI tool “would be an instant failure of the court’s duty to protect such information.”
In October 2025 Judge Wingate (S.D. Miss.) and Judge Neals (D.N.J.) admitted that a clerk using Perplexity and an intern using ChatGPT had produced docketed orders with non-existent parties and misquoted law. Judge Wingate’s new chambers rule is one any firm can copy: a second clerk reviews, and every cited case is printed from Westlaw and attached to the draft. When the AI-drafted filing comes from the other side, the pro se response guide covers how to raise it.
The pre-filing routine and a model certification
Five items, every filing, whatever the jurisdiction; the ethics and regulation hub has the rules behind each.
- Check the order, today. Judge’s page, local rules, scheduling order; tracker last.
- Build the citation table. A model may build it; it may not check it.
- Verify by hand, six layers. Exists in a real database; names, court, year and reporter match; good law via a citator; the pinpoint says what you say; quotations match; right jurisdiction and posture.
- Log who checked what, when, where. It is what you produce if a court asks, and the kind of record underwriters have begun asking about at renewal.
- Sign only what you did. If a paralegal ran the citator and you read the pinpoints, say so in those terms.
List every case, statute, rule, regulation and secondary source cited in <document>...</document> in a table: Citation as written | Type | Proposition it supports (quote my sentence) | Pinpoint given? | Quotation? (Y/N) | Section. Include footnotes and parentheticals. Do not tell me whether any citation exists or is accurate; I will verify each in [Westlaw / Lexis / the official reporter] myself.Draft a certificate regarding generative artificial intelligence for a filing in [court] before [judge], complying with the attached order <order>...</order>. Facts, which you must not embellish: generative AI ([tool and version]) was used for [e.g. summarising the record]; no citations were generated by AI; every authority was read and verified by [name] in [database] on [date]. Use the order's own terminology and the branch that fits these facts. Where the order asks for something these facts do not support, leave a bracketed gap. Under 150 words, ready to sign under penalty of perjury.If a certificate turns out to be wrong after filing, what to do when you find a fake citation is the playbook, and the prompt library has the full verification set. Every sanctioned lawyer trusted an output they had not read; in AI Lab for Lawyers we build the verification habit on your own documents, tool by tool.
Frequently asked questions
Which courts require AI disclosure in filings?
There is no uniform US rule. As of spring 2026 the Legal AI Governance tracker lists 113 active judge-specific or court-specific orders, 82 of which require verification or disclosure, 14 caution and 9 prohibit generative AI outright. Outside the US, Canada's Federal Court requires a first-paragraph declaration, the NSW Supreme Court requires verification statements in submissions and bans AI-generated affidavits, and the UK Upper Tribunal's judicial-review form requires a statement of truth about every authority cited.
What does Judge Starr's AI order require?
Judge Brantley Starr of the Northern District of Texas issued the first mandatory certification on 30 May 2023. Every lawyer must file a certificate attesting either that no portion of the filing was drafted by generative AI or that any AI-drafted language was checked for accuracy, using print reporters or traditional legal databases, by a human being. A party who believes a platform is reliable enough for briefing may move for leave and explain why.
Do I have to certify that I did not use AI?
Only where a judge's standing order or a court rule says so; the Fifth Circuit rejected a circuit-wide certification rule in June 2024 and the Illinois Supreme Court's policy states that disclosure of AI use should not be required in a pleading. Where an order exists, it usually offers two branches: certify no AI was used, or certify that a human verified every AI-drafted passage. Check the judge's page on the day you file, because orders change.
What is FRE 707?
Proposed Federal Rule of Evidence 707 would subject machine-generated evidence that would be expert testimony if offered by a human to Rule 702-style reliability requirements: sufficient data, reliable methods, reliable application. The comment period closed on 16 February 2026 with 59 written comments, 27 of them opposed. At its May 2026 meeting the Advisory Committee did not recommend action and deferred a revised draft to autumn 2026, so it is not yet law.
Do Canadian courts require AI declarations?
The Federal Court of Canada does. Its Notice to the Parties and the Profession, issued 20 December 2023 and updated 7 May 2024, expects parties to inform the court and each other if submitted documents include AI-created content, by a declaration in the first paragraph stating that AI was used, either for the whole document or for specifically identified paragraphs. Ontario's Rule 4.06.1(2.1) separately requires factum signatories to certify the authenticity of every authority cited.