Do you have to disclose AI use in court filings? What you sign

Do you have to disclose AI use in court filings? In most American courts, nothing requires it. In a growing number, something does, and it may be a local rule for the whole district, an order from the judge you drew, or since June 2026 in Florida, the meaning of your signature itself. There is no national rule, so the answer sits in the local rules and in each judge’s standing orders.

What those rules ask comes in three shapes. Most ask you to say that you used a generative AI tool, or to certify that a human checked what it wrote. A few ask something else: that using the tool did not put confidential information in front of anyone not entitled to see it. That third question is the one you cannot repair after the fact, because by the time you sign, the prompt has already gone.

This guide walks all three with the rule text in front of you, then turns to the part no certificate covers: what should not have left the firm before anyone drafted a word.

Do you have to disclose AI use in court filings? The short answer

Check two places before every filing: the court’s local rules and the assigned judge’s individual practices or standing orders. The first binds the whole district. The second binds only that courtroom, and it is where most AI requirements still live. A rule in one courtroom tells you nothing about the one next door.

Then read what the requirement asks you to attest, because the kinds in use today are not the same promise.

What the court asks	What you do	Where it is in force
Disclose the use	Say so on the filing itself	Northern District of Texas, Local Civil Rule 7.2(f)
Certify the citations	Represent that every authority exists and is accurately cited	Florida, Rule 2.515(d)(2), since June 15, 2026
Certify confidentiality	Attest that the tool received nothing it should not have	Judges Vaden and Choe-Groves, Court of International Trade
Keep records	Be able to show your prompts if the court asks	Magistrate Judge Kang, Northern District of California

The rows are not alternatives. Some orders stack two or three of them, and a judge can add one after the case is underway. Who in the firm checks the assigned judge’s page before a filing goes out is a question worth answering in writing, in the same document as your law firm AI policy.

The disclosure rules: a heading on page one, or a notice with the brief

A disclosure rule asks one factual question: did you use the tool? It does not ask whether the output was good. The two versions below are the clearest in federal court today, and both are more demanding than they look on a first read.

In the Northern District of Texas, leaving the heading off is a certification

Local Civil Rule 7.2(f)(1) says that a brief prepared using generative artificial intelligence “must disclose this fact on the first page under the heading” Use of Generative Artificial Intelligence. If the presiding judge so directs, the filer must also identify the specific parts prepared that way.

The definition in (f)(2) is broad on purpose. It covers any computer tool, “whether referred to as” generative AI “or by another name”, capable of generating new content in response to a prompt. A drafting feature inside software you already license counts if it writes text for you.

The paragraph to read twice is (f)(3). A party who files a brief without the disclosure “certifies that no part of the brief was prepared using generative artificial intelligence.” Silence is not neutral here. It is a representation to the court, and it is made by every brief filed in the district.

We compared the district’s archived rule books to date it. The edition effective September 1, 2023 has no such paragraph. The edition effective September 3, 2024 has it, word for word as it reads in the edition effective September 1, 2026. So the rule has now run for two years, which is long enough that nobody filing there can plead surprise.

Magistrate Judge Kang wants the tool named and the prompts kept

The civil standing order of Magistrate Judge Peter H. Kang in the Northern District of California, dated July 16, 2025, goes further than a heading. Any document whose text was created or drafted with an AI tool must be identified as such in its title or caption, in a table before the body, or in a separate notice filed with it.

Counsel must also maintain records sufficient to identify which portions came from the tool. And a separate paragraph, headed AI and Confidentiality, reaches past the filing to the work behind it. That paragraph comes back in the section on protective orders below.

How far privilege survives a trip through a chatbot is a separate question from what the court wants to be told. We work that one through in our guide to redaction and privilege.

The accuracy certifications, and why they are moving into the signature

The first AI orders, in the late spring of 2023, answered one failure: briefs citing cases that do not exist. So many of those that followed ask you to certify that a person checked the citations. The trend of the last twelve months is that courts are dropping the separate certificate and building the promise into the signature you already give under Rule 11 and its state equivalents.

Judge Starr’s certificate started it in 2023

Judge Brantley Starr of the Northern District of Texas posted his Mandatory Certification Regarding Generative Artificial Intelligence at the end of May 2023. Attorneys had to file a certificate attesting either that no portion of a filing was drafted by generative AI or that any such language “was checked for accuracy, using print reporters or traditional legal databases, by a human being.”

It was about hallucinations and bias, and it said nothing about confidentiality. When the Internet Archive captured his page on August 27, 2026, the certificate was no longer there. The page instead has litigants represent that they will abide by the district’s Local Civil Rules, which by then included 7.2(f).

Florida replaced its circuit orders with one line in Rule 2.515

On January 15, 2026, the chief judge of Florida’s Eleventh Judicial Circuit, which is Miami-Dade, issued Administrative Order 26-04. It required disclosure “on the face of the filing” of any generative AI use and a certification that all factual assertions, legal authority and citations had been independently reviewed and verified.

Four months later the Florida Supreme Court moved statewide, on its own motion. Its opinion of May 28, 2026 amends Rule 2.515(d)(2) so that the signer of any document filed in a Florida court represents that “the legal authorities identified exist and are accurately cited,” and courts may sanction a filing inconsistent with that. The change took effect on June 15, 2026.

Miami-Dade then withdrew its own. In May 2026 the circuit had already replaced AO 26-04 with a narrower order, AO 26-15, and AO 26-18 rescinded that one from June 15, 2026, saying the amended rule made it unnecessary.

When	What changed in Florida
January 15, 2026	Miami-Dade’s AO 26-04 requires disclosure and a verification certificate
May 2026	AO 26-15 replaces it with a narrower order
May 28, 2026	The Florida Supreme Court amends Rule 2.515(d)(2)
June 15, 2026	The amendment takes effect, and AO 26-18 rescinds AO 26-15

Florida Supreme Court opinion of May 28, 2026, and Eleventh Judicial Circuit administrative orders.

The court’s commentary says it adopted the amendment “principally to create a statewide, uniform replacement for varied circuit court administrative orders imposing disclosure and certification requirements.” Notice what the replacement no longer asks. It does not ask which tool you used, and it does not ask anything about confidentiality. If you practice in a Florida circuit that had its own order, check whether it has been withdrawn.

The Tenth Circuit wants it in the signature from 2027

On September 18, 2026 the Tenth Circuit published proposed revisions to its rules. A new paragraph, Rule 46.5(B)(4), would add to what every signature certifies: that any document prepared with the assistance of generative AI “has been reviewed” before filing by a human who has independently verified the existence of all cited authorities, the accuracy of all quotations, citations and legal analysis, and “the document’s compliance with all relevant rules and ethical obligations.”

46.5(B)(4)the proposed Tenth Circuit paragraph

Oct 18, 2026last day for public comments

2027revised rules take effect on January 1

Comments go by email to the clerk until October 18, 2026, and the final rules are due on the court’s website by November 30. That last clause of the proposal is broad. Read literally, a signer certifies compliance with every relevant ethical obligation, and a careful signer will read that as reaching the confidentiality rules too, though the proposal does not say so expressly.

Inside the same circuit, the District of Kansas already took a softer route. Its Standing Order 26-01 of January 28, 2026 reminds litigants that they must verify everything an AI tool drafted, and keeps the court’s discretion to require, case by case, a sworn statement naming the tool and the portions it produced. If you are drafting the firm’s position anyway, there is room for it in the AI policy template we publish.

The certificate that asks what left the firm

Two judges at the Court of International Trade ask a question the others do not. Trade cases run on business proprietary information that the parties are allowed to see only under the court’s rules, so a judge there has a concrete reason to care where a prompt goes, not only whether the citations in the brief are real.

Judge Vaden: name the program and certify nothing confidential went in

Judge Stephen Alexander Vaden’s Order on Artificial Intelligence is dated June 8, 2023. It starts from the court’s own rules: confidential information in a brief must be bracketed under USCIT Rule 5(g), and an attorney may receive business proprietary information only after certifying and being granted access under Rule 73.2(c)(2).

The order then explains its worry in plain terms. Users “may include confidential information in their prompts, which in turn may result in the corporate owner of the program retaining access to the confidential information.” Any submission containing text drafted with a generative AI program must therefore come with two things.

1. A disclosure notice that identifies the program used and the specific portions of text drafted with it.
2. A certification that the use of the program “has not resulted in the disclosure of any confidential or business proprietary information to any unauthorized party.”

After the notice is filed, any party may ask the court for whatever relief it thinks the facts warrant. So the notice is not only for the judge. It tells your opponent exactly which program saw your draft.

Judge Choe-Groves added a third line in 2025

Judge Jennifer Choe-Groves’s standing order of July 3, 2025 keeps both of those and adds one more. The notice must also say the manner in which the tool was used, and counsel must certify that the text and legal research were reviewed for factual and legal accuracy. Her third item is the confidentiality line, worded as no “unauthorized disclosure of any confidential or business proprietary information.” So her order asks for three statements where his asked for two.

3

The difference between the accuracy line and the confidentiality line is what this guide is about. You can make the first one true after the fact: you check the citations, fix what is wrong, and sign. You cannot make the second one true after the fact.

If the paragraph with the bracketed figures went into a consumer chatbot on Tuesday, no amount of checking on Thursday changes what the provider received. Our guide to what OpenAI does with a prompt covers what the consumer products say they keep.

Do you have to disclose AI use in court filings if no court asks? The duty is still there

Most courts in the country have no AI order at all, and some have decided on purpose not to have one. That settles the disclosure question. It does not settle the confidentiality question, which was never the court’s to create.

Illinois says do not require disclosure, and still names what must not leak

The Illinois Supreme Court’s policy on artificial intelligence, effective January 1, 2025, is the clearest example. It says AI use by attorneys “should not be discouraged” and that “disclosure of AI use should not be required in a pleading.” Attorneys remain accountable for their final work product.

The same page then draws a line on data. “AI applications must not compromise sensitive information, such as confidential communications, personal identifying information (PII), protected health information (PHI)” and several other categories. So even in the state that told its courts not to demand a disclosure, the confidentiality expectation is written down. In an employment case the two labels part ways, since the line between PHI and PII depends on who holds the medical note.

Opinion 512 wants consent before the prompt, not a boilerplate clause

ABA Formal Opinion 512, dated July 29, 2024, touches the court question only briefly. It notes that “some courts have responded by requiring lawyers to disclose their use” of generative AI, and it tells lawyers to consult the applicable court’s local rules. The duty it spells out at length runs to the client. How that duty reads on paper, with notice and consent kept apart, is in our sample clause for engagement letters.

For self-learning tools, it says client informed consent is required before information relating to the representation goes in, and that “merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use GAI is not sufficient.” The letter still has a role: for telling clients how the firm uses AI, the opinion calls the engagement agreement “a logical place.” The rest of the opinion, and how it sits next to privilege and Rule 1.6, is in our guide to redaction and privilege.

Protective orders do not come with an AI exception

A protective order decides who may receive designated material, and it was drafted with people in mind: counsel, experts, the court. An AI provider is not usually on that list. Magistrate Judge Kang’s order says so directly. Because the owner of a publicly available AI tool “may have access to, ownership of, or otherwise retain information input or submitted” to it, counsel choosing to use one “shall fully comply with any applicable protective order.”

The same paragraph asks counsel to keep records sufficient to show that compliance, “such as by keeping records of all prompts.” That is the most concrete statement in any federal order so far of what a firm should be able to produce if a judge asks how a document was made.

What Heppner shows about a prompt that leaves the building

The confidentiality certificates are about what a court might ask. United States v. Heppner is about what an opponent can obtain. On February 17, 2026, Judge Jed S. Rakoff, sitting in Manhattan, explained a ruling he had made from the bench a week earlier, which the memorandum itself describes as answering a question of first impression.

31

The defendant, facing securities and wire fraud charges, had used Claude on his own after receiving a grand jury subpoena, to prepare reports outlining his defense. Agents seized approximately 31 documents of that kind, and his lawyers claimed privilege over them. The court held that neither the attorney-client privilege nor the work-product doctrine protected them.

Three reasons, and the one that matters for firms

The first reason was simple: Claude is not an attorney. The second reason travels further. The provider’s written privacy policy said it collects inputs and outputs and may disclose data to third parties, including in connection with litigation, so the defendant “could have had no reasonable expectation of confidentiality.” The third was that counsel had not directed him to use the tool.

This was a client acting alone, not a lawyer drafting a brief, and the opinion says so. But the second reason is about the terms of the product, not about who typed. What each of the major providers says it keeps, and for how long, is in our guides to Claude and Copilot on US terms.

Before you sign: a checklist for AI in court filings

None of the rules above is hard to follow once you know it applies. The failures come from not looking, or from treating the certificate as a formality at the end instead of a description of how the work was done. This is the order we would run it in for AI in court filings, whichever court you are in.

1. Find the rule for this judge. Local rules first, then the assigned judge’s page, then any administrative order of the circuit or division. Do it per case, not per firm.
2. Decide whether the brief needs a disclosure. In the Northern District of Texas, the absence of the heading is itself a certification that no AI was used.
3. Verify every authority by hand. That is what the Florida signature, the Tenth Circuit proposal and most individual orders now make you promise.
4. Read the protective order against the tool. If designated material would go to a provider not named in the order, the answer is no, whatever the judge’s AI order says.
5. Check the account, not just the product. A consumer login and a contracted workspace are different terms, and Heppner turned on the terms.
6. Keep a record of what went in. Magistrate Judge Kang asks for the prompts, and no other court can be ruled out from asking.
7. Get the client’s consent in words, not boilerplate. Opinion 512 says a general clause in the engagement letter is not enough.

Items four and five are the ones a certificate cannot fix later, which is why they come before the draft and not before the signature. The notification side of the same mistake is worked through in our guide on client files, chatbots and breach duties.

What should not have left the firm before you sign

If a court asks you to certify that the AI tool received nothing confidential, the honest way to sign is to know what went in. The list has several kinds of material on it, and software can find only a few of them.

What it is	Where the duty comes from	Can a masking step catch it?
Client names, SSNs, birth dates, account numbers	Model Rule 1.6	Largely, if the tool recognizes the format
Figures marked as business proprietary	USCIT Rule 5(g) and the court’s access orders	No. A number is not a name
Anything designated under a protective order	The order itself	No. The designation is not in the text
Litigation strategy and the facts that reveal it	Privilege and work product	No. The story survives the placeholders
Sealed material	The sealing order	No, unless it is only identifiers

The first row is the only one where a tool can do most of the work, and even there it helps to know what the tool does not see. The other rows are judgments about what a paragraph is, not about which characters are in it.

A placeholder hides a name, not a strategy

Swap every name in a draft for [PERSON_1] and the argument is still there: the timeline, the weak point in the other side’s case, the expert you have not yet disclosed. Model Rule 1.6 covers anything tied to the representation, wherever it came from, and a matter can still be recognized from its facts once every name has been stripped out.

So masking belongs in the workflow as the step that keeps identifiers out of a prompt that was going to be sent anyway. It is not the step that decides whether a document should be sent at all. Masking, pseudonymizing and truly deidentifying are three different things, and our comparison of the three terms sets them side by side.

Where Nonimo fits in this, and where it stops

Nonimo is a small program that runs on each attorney’s computer, Mac or Windows. When the attorney highlights a passage and presses the key, it replaces the names and numbers it recognizes with placeholders such as [PERSON_1] so that the version the chatbot receives carries only the placeholders, and restores the real values when the answer comes back.

Most American numbers have no check digit that arithmetic could confirm, so when one is swapped, the swap is marked for the attorney to accept or undo.

The working conversation is kept on that machine in encrypted form, since that is what lets the answer get its names back, as the security page sets out.

How it fits a firm’s written policy, with a real before and after on invented intake notes, is on our page for law firms.

What it cannot do is the part the Court of International Trade asks you to certify. It cannot tell whether a figure is business proprietary, whether a paragraph is under a protective order, or whether a strategy is privileged. Those are the attorney’s calls, made before the key is pressed.

The short version for whoever signs the brief

The disclosure question has an answer in every courtroom, and it is written down somewhere: the local rules, the judge’s page, or a statewide rule of procedure. Find it per case. In the Northern District of Texas, remember that silence certifies.

The accuracy question is converging on one answer. Florida and, from 2027 if the proposal holds, the Tenth Circuit fold it into the signature, so every brief carries the promise whether or not AI touched it.

The confidentiality question is the one that matters most and is asked least. Only a few courts put it on paper, but Rule 1.6, protective orders and now Heppner put it everywhere. Answer it before the prompt, because afterward there is nothing left to decide. If you want a firm policy that names the safeguards you actually use, start there.

Sources

• Local Civil Rules of the Northern District of Texas, edition effective September 1, 2026, LR 7.2(f)(1) to (3): the heading, the definition and the certification by silence. The capture of September 6, 2026 is the copy we read; earlier captures date the rule to the edition effective September 3, 2024.
• Magistrate Judge Peter H. Kang, Civil Standing Order, Northern District of California, July 16, 2025, section on AI: identification in the caption, records of portions, compliance with protective orders and records of prompts.
• Judge Brantley Starr, judge specific requirements, as captured on May 30, 2023: the Mandatory Certification Regarding Generative Artificial Intelligence. The capture of August 27, 2026 no longer carries it.
• Eleventh Judicial Circuit of Florida, Administrative Order 26-04, January 15, 2026: disclosure on the face of the filing and the model certification.
• Eleventh Judicial Circuit of Florida, Administrative Order 26-18, June 2026: rescinds AO 26-15, which had vacated and superseded AO 26-04, effective June 15, 2026, citing the amended Rule 2.515. The circuit’s list of administrative orders no longer carries either of those two AI orders.
• Supreme Court of Florida, No. SC2026-0673, May 28, 2026: amended Rule 2.515(d)(2), effective June 15, 2026, and the commentary describing it as a statewide replacement for circuit orders.
• Tenth Circuit, proposed 2027 rule revisions, September 18, 2026, with the memorandum and the full redline: proposed Rule 46.5(B)(4), comments until October 18, 2026, effective January 1, 2027.
• District of Kansas, Standing Order 26-01, January 28, 2026: verification duty and the court’s discretion to require a sworn statement.
• Judge Stephen Alexander Vaden, Order on Artificial Intelligence, Court of International Trade, June 8, 2023: the disclosure notice and the confidentiality certification, with USCIT Rules 5(g) and 73.2(c)(2).
• Judge Jennifer Choe-Groves, Standing Order on the Use of Artificial Intelligence, Court of International Trade, July 3, 2025: the three statements.
• Illinois Supreme Court Policy on Artificial Intelligence, effective January 1, 2025: no disclosure in pleadings, and the list of sensitive information AI must not compromise.
• ABA Formal Opinion 512, July 29, 2024: informed consent before input into self-learning tools, boilerplate not sufficient, and the note on court disclosure requirements.
• United States v. Heppner, 25 Cr. 503 (JSR), Memorandum, S.D.N.Y., February 17, 2026: the 31 AI documents and the three reasons privilege failed.