Do you have to disclose AI use in US court filings?

Do you have to disclose AI use in court filings? No federal statute says so, and neither do the Federal Rules of Civil Procedure. The answer sits lower down: in a district’s local rules, in your judge’s standing order, in a few state court rules and, from January 1, 2027, in a California statute.

Put these orders side by side, one courtroom at a time, and what strikes me is how little they agree on. Some want a heading, some a certificate, some nothing at all. If your firm needs a written routine around them, our page for law firms is where that starts.

Is there a federal rule on AI disclosure in court filings?

No. The Federal Rules of Civil Procedure say nothing about generative AI. Rule 11 already makes whoever signs a paper vouch for its law and its facts, and many federal judges who write about AI simply remind lawyers of that duty instead of adding a form.

The federal rule people mention most, proposed Federal Rule of Evidence 707, deals with something else. It would set a reliability test for machine output offered as evidence without an expert witness. It says nothing about who or what drafted your brief, and it is not law.

I wouldn’t plan anything around Rule 707 yet. The minutes in the committee’s October agenda book record a second version still being shaped and the Justice Department saying its concerns persist. The per-judge check belongs in your firm’s AI policy, because the highlighted rows are where the obligations actually sit.

Federal judge standing orders on generative AI, court by court

Most requirements live in single courtrooms, not in district-wide rules. Two judges in the same courthouse can ask for different things, and an order can appear after your case is assigned. The panels below sum up the orders I opened; the judge’s own page is the one that binds you.

They are built very differently. Some want the tool named, some want the drafted portions identified, and one wants a certificate with every filing, whether AI was used or not.

Several of these reach any tool that writes text from a prompt, including an assistant built into software you already license. If that assistant is Microsoft’s, read what Copilot does with what it reads before you sign a certificate that mentions it.

When filing without a disclosure counts as a certification

Two Texas courts turn silence into a statement. In the Northern District, a brief without the heading certifies under LR 7.2(f)(3) that no part of it was prepared using generative AI. Magistrate Judge Love in Tyler presumes the same about a filing that comes without his certificate.

The definition is broad. The Northern District’s covers any computer tool, whatever it is called, that generates new content from a prompt. If an associate let a built-in assistant rewrite one paragraph, the brief may need the heading. I would ask everyone who touched the draft, since the signer may not know.

Asking the team has a side benefit: you learn which accounts the drafts went through. That matters later, because what OpenAI keeps from a prompt depends on the plan behind the login.

AI hallucinated citations: sanctions and the “independently verified” certification

The phrase “independently verified the existence of all cited authorities” comes from the Tenth Circuit’s proposed Rule 46.5(B)(4). Every signature would certify that a human reviewed any AI-assisted document before filing, checking authorities, quotations, citations, analysis, and compliance with all relevant rules and ethical obligations.

The court’s memo set October 18, 2026 as the comment deadline and November 30 for posting the final rules, which take effect January 1, 2027. The redline also rewrites the sanctions paragraph, opening with waiver of the affected issue and adding dismissal or affirmance of the appeal.

States are moving the same way, mostly without naming AI in the duty. Since June 15, 2026, Florida signers represent that cited authorities exist and are accurately cited. Indiana’s Trial Rule 11 has required certifying that authority is authentic since August 1, 2026, and Connecticut’s new section 4-9 makes filers verify what the tool produced.

None of these asks which tool you used. Verifying a citation means opening the case, and no assistant does that for you, whether it runs in a browser or on the firm’s own hardware. One named person per filing should own the check; “the team” never opens a case.

State court rules on AI: Florida, Connecticut, Indiana, Illinois and California

Most state courts are moving away from disclosure. Florida went furthest: administrative order AOSC26-12 says its courts may not impose their own AI disclosure or certification requirements, whether by local order, court policy or judicial practice. Amended Rule 2.515 does that job for the whole state.

Illinois got there first. Its supreme court’s policy, effective January 1, 2025, says disclosure of AI use should not be required in a pleading, while requiring review of AI content and protection of confidential information, PII and PHI. How PHI and PII differ matters once medical records sit in the file.

California went the other way, by statute. Governor Newsom signed SB 574 on September 30, 2026, as Chapter 858, and it takes effect January 1, 2027. New Business and Professions Code section 6068.1 requires an attorney who uses generative AI to disclose that use to the court for all documents submitted to it.

The same section bars entering confidential, personal identifying or other nonpublic information into a system whose access isn’t limited to the attorney and people bound to keep it confidential. Amended Code of Civil Procedure section 128.7 adds that no filed paper may carry a citation the responsible attorney hasn’t personally verified.

Section 6068.1 says nothing about the form the disclosure takes, and that gap is what I would watch most closely in California before January. The bill’s own summary leaves the court disclosure out, so read the section itself, not a digest of it.

Certifying that no confidential information went into the AI tool

A smaller group of orders asks something you can’t fix afterward. Judges Vaden and Choe-Groves at the Court of International Trade, Judge Crone in Beaumont and Judge Gonzalez in Del Rio require a certification that using the tool disclosed no confidential information to an unauthorized party.

The trade court adds business proprietary information; the Texas judges add privileged information. Citations can be checked on Thursday for a draft written on Tuesday, but a prompt can’t be called back, and deleting the chat history removes it from your screen first and from the provider later.

The fifth box is the easiest to miss: a Word file can still carry comments and tracked changes nobody accepted, and the tool receives them along with the visible text.

The account matters as much as the product, since consumer and business plans run on different terms. For one major provider, what Anthropic does with Claude inputs sets them side by side.

Opponents can come at the same prompts from another angle. In United States v. Heppner, Judge Rakoff held on February 17, 2026 that a defendant’s own exchanges with a public AI platform were protected by neither privilege nor work product; our guide to privilege and redaction follows that reasoning.

Protective orders and keeping a record of AI prompts

Magistrate Judge Kang wrote the most practical paragraph so far. Because the operator of a public AI tool may retain what you submit, counsel who use one must fully comply with any applicable protective order and keep records, “such as by keeping records of all prompts,” to show compliance if asked.

A protective order lists who may see designated material: counsel, experts, the court. An AI provider is rarely named. I treat “may we paste this?” as a question about the order first and about the tool second, because the order is what a judge will read.

If a spreadsheet went up with the prompt, file the masked CSV you actually sent next to the log entry, dated, so the record shows what the assistant received.

When the assistant is the firm’s own software, the prompts often already sit in an AI gateway’s request log. That log then deserves the same care as the matter file.

If that software calls the OpenAI API, a privacy proxy on the firm’s own servers can replace personal data with labels before each request leaves the network.

Do lawyers have to disclose AI use to the client as well?

The court and the client are different audiences. ABA Formal Opinion 512, of July 29, 2024, mentions court disclosure briefly: some courts require it, and a footnote tells lawyers to consult the applicable local rules. Its long analysis concerns the client, including informed consent before confidential information goes into many self-learning tools.

It also says boilerplate in an engagement letter is not enough for that consent. The client side has its own guide with sample letter language. Neither one substitutes for the other, in my view: a court certificate tells the client nothing, and a client’s consent satisfies no judge.

If client files already went into a consumer chatbot, the question shifts from consent to notice, which is what our guide on client data and breach duties works through.

What a redaction step covers, and what stays the signer’s call

Nonimo is a small program that runs on each attorney’s computer, Mac or Windows. When the attorney highlights a passage and presses its key, it replaces the names and numbers it recognizes with placeholders such as [PERSON_1], sends only that version, and puts the original values back in the answer.

Each swap shows on screen, so the attorney can accept or undo it. The mapping that restores the names never leaves that computer; our security page explains how.

The marked words are the attorney’s decision, not the software’s. Whether a figure is business proprietary, whether a paragraph falls under a protective order, whether the argument gives the case away: that judgment comes before pressing anything. A fuller list of what to remove goes well past names.

Swapping names for placeholders also isn’t deidentification in any statutory sense, so check the three terms compared before anyone calls a prompt “anonymous” in a certificate.

Before you sign the brief: one routine per filing

The disclosure question has a different answer in each courtroom, and you have to look it up per case. The citation question is converging everywhere on the same duty. The confidentiality question appears in only a handful of orders, yet it is the one you have to settle before the first prompt.

Judges revise these pages without notice, and this guide is general information, not legal advice; if an order here reads differently when you file, the order wins. Where the firm wants client identifiers out of prompts without changing tools, Nonimo does that swap on each attorney’s computer before anything is sent.

Sources

• Magistrate Judge Peter H. Kang, Civil Standing Order, N.D. Cal., July 16, 2025: identification of AI-drafted text, protective orders, records of prompts.
• Advisory Committee on Evidence Rules, agenda book for October 15, 2026: proposed Rule 707 withdrawn as published and under further study.
• Northern District of Texas, Local Civil Rules effective September 1, 2026, capture of October 9, 2026: LR 7.2(f)(1) to (3).
• Supreme Court of Florida, No. SC2026-0673, May 28, 2026: amended Rule 2.515(d)(2) and its sanctions, effective June 15, 2026.
• Supreme Court of Florida, AOSC26-12: no local AI disclosure or certification requirements.
• Connecticut Law Journal, June 23, 2026: new Practice Book section 4-9 and revised section 4-2.
• Indiana Supreme Court, Order 26S-MS-8, July 14, 2026: Trial Rules 11, 11.1 and 11.2, effective August 1, 2026.
• Illinois Supreme Court Policy on Artificial Intelligence, effective January 1, 2025: no disclosure in pleadings.
• ABA Formal Opinion 512, July 29, 2024: candor to the tribunal, footnote 47, boilerplate consent.
• Office of the Governor of California, September 30, 2026: SB 574 signed.
• California Legislative Information, SB 574 as chaptered, Chapter 858, Statutes of 2026: Business and Professions Code section 6068.1 and Code of Civil Procedure section 128.7(b)(2).
• Not linked here, each on its court’s site: Magistrate Judge Love’s order of April 9, 2025 and Judge Crone’s Standing Order MC-11 (E.D. Tex.); Judge Gonzalez’s order of June 25, 2026 (W.D. Tex.); the AI orders of Judges Vaden and Choe-Groves (Court of International Trade); District of Kansas Standing Order 26-01; the Tenth Circuit’s 2027 rules memo and redline. United States v. Heppner, No. 25 Cr. 503 (S.D.N.Y. Feb. 17, 2026), ECF No. 27.

Checked on November 7, 2026.