Government and Court Filings: Rules on AI-Assisted Legal Writing

A lawyer in New York once filed a brief citing court decisions that turned out not to exist. A chatbot had invented them, complete with realistic-looking quotes and docket numbers, and nobody checked before the document went to the judge. The case, *Mata v. Avianca*, ended in sanctions for the attorneys, and it kicked off a wave of new rules about what you can and can’t do when AI helps write a legal filing.

If you draft anything that lands in front of a court or a government agency, the question isn’t really “can I use AI?” It’s “what do I owe the court when I do?” The answer is scattered across local rules, individual judges’ standing orders, and bar guidance, so it pays to know how the pieces fit together.

Key takeaways

• There’s no single national rule. Requirements differ by court, by judge, and by agency, so check the specific docket before you file.
• The core duty is verification, not disclosure. Rules like Federal Rule of Civil Procedure 11 already make you responsible for every citation and factual claim.
• Hallucinated citations are the headline danger. Lawyers have been sanctioned for filing AI-invented cases, as happened in *Mata v. Avianca* (2023).
• Many judges’ standing orders ask for a disclosure or a human-verification certification. A few restrict AI in certain filings.
• Drafting with AI is usually fine. Treating AI as a replacement for judgment and fact-checking is what gets people in trouble.

Why there’s no one rule to point to

People want a clean answer here, and there isn’t one. The U.S. legal system is layered. Federal courts, state courts, appellate courts, and administrative agencies each set their own procedures, and within a single courthouse individual judges issue their own standing orders that govern how matters proceed in their courtroom.

So a rule about AI-assisted writing can live in at least four places:

• Federal Rules of Civil or Criminal Procedure — broad duties that apply to everyone, like the certification you make every time you sign a filing.
• Local rules — court-wide requirements specific to a district or circuit.
• Standing orders — an individual judge’s instructions, which can be stricter than the local rules.
• Bar and ethics guidance — professional-responsibility opinions from state bars on competence, candor, and confidentiality.

Because these layers stack, the safest move is always to read the judge’s standing orders first. That’s where AI-specific language tends to show up.

What the standing orders actually say

Starting in 2023, after the fake-citation scandals, a number of federal judges began publishing standing orders about generative AI. They aren’t uniform, but they cluster into a few patterns.

Disclose-and-certify. The most common version asks the filer to state whether generative AI was used in preparing the document and to certify that a human reviewed every citation and legal proposition for accuracy. You’re not banned from using AI; you’re on record that you checked its work.

Restriction. A smaller set of orders limit using generative AI for certain filings, or require advance permission. These are less common and tend to apply to specific document types.

Silence. Plenty of judges have said nothing about AI at all. Silence doesn’t mean “anything goes,” though. The general duties under Rule 11 and the rules of professional conduct still apply with full force.

Here’s the practical wrinkle: an order on Judge A’s docket tells you nothing about Judge B down the hall. If you practice across multiple courts, you can’t memorize one rule and move on. You check each one, every time.

The duty that was always there: verification

Strip away the new AI language and you find an old obligation underneath. When you sign a filing in federal court, Rule 11 says you’re certifying that your legal contentions are warranted and your factual claims have evidentiary support. That certification doesn’t care how you produced the document. A quill pen, a paralegal, a search engine, or a language model — the signer owns the accuracy.

This is why the AI panic in courts isn’t really about AI. It’s about people skipping verification. A model that confidently produces a case citation isn’t lying to you on purpose; it’s pattern-matching what a citation looks like. The volume number might be plausible and the case might be fiction. The rule was never “don’t use tools.” It’s “don’t certify things you haven’t checked.”

A quick mini-scenario

Imagine you’re a litigation associate on a tight deadline. You ask an AI to draft a motion to dismiss and it produces three pages of clean argument supported by four cases that look directly on point. Tempting. Say two of those cases are real and roughly say what the draft claims. One is real but actually stands for the opposite proposition. And one doesn’t exist at all — the citation looks plausible, but there’s no such case behind it.

File that as-is and you’ve potentially violated Rule 11 twice over: a fabricated authority and a misrepresented one. The fix costs you twenty minutes. Pull each case in your research database. Read the holding, not just the headnote. Confirm the quote appears on the page cited. The AI saved you drafting time; verification is the price of using the draft.

Government and agency filings

Courts get the headlines, but agencies have their own expectations. Many federal and state agencies have published internal generative-AI guidance, and the recurring themes are accuracy, data security, and human review before submission. If you’re filing comments in a rulemaking, submitting to a regulatory body, or preparing a government report, two questions matter:

1. Does this agency have a published AI policy? Some do, and a few ask submitters to attest to how a document was prepared.
2. Am I about to feed confidential or privileged material into a third-party tool? Putting client or case data into a consumer chatbot can create confidentiality problems independent of any court rule.

That second point trips people up more than the filing rules do. A disclosure obligation is something you can satisfy with a sentence. A confidentiality breach is harder to undo.

What about AI detectors?

A reasonable worry: will a court or agency scan my filing with an AI detector and penalize me for AI-sounding prose? In practice, that’s not where the risk lives. Detectors estimate likelihood — AI detectors only output probabilities, not proof — and they generate false positives often enough that no serious tribunal would hang a sanction on a detector score. (OpenAI pulled its own AI Text Classifier in July 2023 because it just wasn’t accurate enough.)

So don’t reframe this as a game of disguising your writing. The court doesn’t care whether your brief reads like a machine helped. It cares whether the cases are real and the facts are supported. Spend your energy on verification, not camouflage. If you want to browse more articles on AI writing and detection, there’s plenty on how detection actually works and why it’s a shaky basis for accusations.

A practical checklist before you file

• Read the judge’s standing orders and the local rules. Look specifically for AI, generative AI, or certification language.
• Verify every citation by hand. Open each case. Confirm it exists, says what you claim, and supports the proposition.
• Check quotes against the source page. Don’t trust a quotation a model handed you.
• Watch confidentiality. Don’t paste privileged or sensitive material into tools that may store or train on it.
• Keep a record. If an order requires disclosure, include the certification language and keep notes on what you reviewed.
• Use AI for first drafts and structure, not final judgment. The model can organize an argument; you decide whether the argument is correct.

Frequently asked questions

Do I have to tell the court I used AI? It depends on the court and the judge. There’s no universal federal rule as of 2026; some judges require a signed AI-use certification, others say nothing. Read the standing orders.

Is it illegal to draft a brief with ChatGPT? No. What creates liability is filing something you didn’t verify. Rule 11 holds the signer responsible no matter how the document was produced.

What if the tool invents a fake citation? That’s the classic failure mode. In *Mata v. Avianca* (2023), the lawyers were sanctioned and fined for it. Pull and read every cited source before filing.

Do agencies have their own AI rules? Many do, and they’re evolving. They tend to stress data security, accuracy, and human review. Check the specific agency’s published policy.

The bottom line

AI-assisted legal writing isn’t forbidden, and it isn’t a loophole. It’s a tool whose use is governed by duties that predate it — accuracy, candor, confidentiality, and competence — plus a growing patchwork of court-specific orders that mostly ask you to confirm a human checked the work. The lawyers who got sanctioned didn’t get caught for using AI. They got caught for not reading what they signed.

If you draft with AI and want a faster way to make a first pass read naturally before you do the real work of verifying it, you can try PaperBleach — or see what PaperBleach actually does first. Just remember the tool’s job ends where your professional judgment begins. The citations are still yours to check.