AI-Written Legal Briefs: Sanctions, Rules and Checks

Can I use a chatbot to draft a brief without getting sanctioned? Yes. The qualification matters far more than the answer, because the sanctions you have read about were never really about the tool.

Courts have not banned AI generated legal briefs. What they sanction is filing a document containing citations to cases that do not exist, and failing the verification duty every signature on a filing already carries. The technology changed how fast fabricated authority can be produced. It did not change who answers for it.

What courts actually sanction, and it is not the software

Strip the headlines down and the same fact pattern appears every time. A filing quotes authority. Opposing counsel or chambers cannot find the authority. The cases turn out not to exist, or to exist while saying something different from what the brief claims. The filer either did not check or checked in a way that could not have worked.

Judges have responded to that pattern with the tools they already had. Certification obligations attached to signing a court filing, the inherent power to police proceedings, and professional conduct rules on competence and candor were all sitting there long before 2023. The sanction is for the false representation to the court, and the remedy has ranged from orders to show cause and public opinions through fee awards and referrals to disciplinary bodies. The specific consequence depends entirely on the court, the jurisdiction and how the filer responded once caught, which is the part worth studying: courts have consistently treated candor after discovery as mattering as much as the original error.

None of this required a new rule about AI. That is the point most commentary misses. If you had submitted a brief in 2015 citing a case you invented, the outcome would have looked much the same. The tool did not create a new offense; it lowered the effort required to commit an old one, and it made the resulting text look far more credible than a human fabricator could have managed.

How a fabricated citation gets into a filing

Understanding the mechanism is what makes the verification step feel necessary rather than bureaucratic.

A language model generates text token by token, selecting what is statistically likely given everything before it. A citation is a highly structured string: volume, reporter, page, court, year. The model has seen millions of them and has learned the shape perfectly. What it has not got is a lookup against a case database. So when the argument calls for authority on a proposition, the model produces something that has the exact form of a citation, with a plausible party name, a plausible reporter, and a year that fits the doctrinal timeline.

The output is not a lie and not a garbled fact. It is a confident reconstruction of something the system never had access to, which is why the resulting citations tend to look better than real ones. Real case names are often ugly. Generated ones are clean.

Put a person in it. A third-year associate at a 30-lawyer firm has a summary judgment reply due at midnight and needs authority on a narrow evidentiary point in an unfamiliar jurisdiction. The model returns four cases. Three are real. The fourth is not, and it is the cleanest and most precisely on-point of the set, because nothing constrained a generated case to be inconvenient. She has forty minutes and verifies three of four. From inside that moment, the check feels like it worked.

Two adjacent failures are subtler and more dangerous, because they survive a superficial check. The first is a real case cited for a proposition it does not support. The citation verifies, the pin cite exists, and the holding is simply not what the brief says it is. The second is a real case that has been reversed, superseded or limited, which a model trained on a snapshot has no way to know.

Disclosure and certification rules as of mid-2026

The rule environment is fragmented and moving, so the honest summary is a shape rather than a list.

Some judges have issued standing orders addressing generative AI, and these vary widely. Some require a certification that a human verified every citation. Some require disclosure that AI was used in preparing a filing. Some do neither and rely on existing certification duties. There is no national registry, no uniform text, and no substitute for reading the standing orders of the specific judge before whom you are appearing. That last sentence is the entire practical takeaway, and it changes month to month.

Outside the courtroom, the EU AI Act's Article 50 transparency obligations become applicable on August 2, 2026, requiring AI-generated content to be marked in machine-readable form and deepfakes to be disclosed. How those obligations interact with professional legal work product is not something anyone can state with confidence yet.

Drafting task	AI suitability	Non-negotiable human check
Summarizing a document you supply	High	Read the source for what was dropped
First-draft structure and headings	High	Argument order serves your theory
Restating undisputed procedural history	Medium	Every date against the docket
Drafting an argument section	Medium	Every authority in a live citator
Finding supporting authority	Low	Assume nothing exists until verified
Quoting a holding	Very low	Read the opinion, not the summary

Why detectors are the wrong tool for AI generated legal briefs

We build detection software, and this is a section arguing against using it for something. Consider that a measure of how badly the fit works.

Legal writing is the most templated professional prose in existence. IRAC and CREAC structure, standardized citation formats, boilerplate procedural recitations, and jurisdiction-specific phrasing that every practitioner copies because deviation looks amateurish. Detectors score statistical predictability, and a well-drafted brief is predictable by design and by tradition. The genre sits permanently near the top of the false positive range.

The research supports the pessimism. In the RAID benchmark (Dugan et al., ACL 2024), built on more than 10 million documents and 12 adversarial attacks, commercial detectors degraded sharply under paraphrase and homoglyph attacks. A 2025 NBER working paper by Jabarian and Imas at the University of Chicago found that only one tested commercial detector met a strict 0.5% false-positive policy cap, with per-detection costs of $0.02 to $0.06. And OpenAI retired its own classifier in July 2023 for low accuracy, publishing that it caught 26% of AI text while false-flagging 9% of human text.

Run the arithmetic for a firm. Screening 500 documents a month at $0.02 to $0.06 costs $10 to $30, so cost is not the obstacle. Reliability is. A tool with a 5% false positive rate produces about 25 wrongly flagged documents a month, which is 25 conversations in which a partner asks an associate to prove they wrote something. That is a morale and time cost with no corresponding benefit, because a true positive tells you nothing actionable either. Knowing a draft was AI-assisted does not tell you whether its citations are real, and a fully human-written brief with a fabricated cite is exactly as sanctionable.

There is a narrow legitimate use: screening inbound documents of unknown provenance, where you want a signal about how much scrutiny to apply. For that, AI Detector 360's free AI detector gives a sentence-level heatmap with an explicit confidence level and a downloadable PDF report, and our ChatGPT detector is tuned for GPT-class output specifically. Read the confidence label rather than the headline number, and treat the whole thing as a triage hint. How we calibrate confidence is published on our methodology page, and the honest error picture is in how often detectors get it wrong.

A verification workflow that holds

Take the strongest counter-argument first, because it is right: blanket prohibitions push AI use underground, where it happens without supervision or logging, and the profession genuinely benefits from tools that compress hours of document review into minutes. A policy that pretends nobody will use these tools produces exactly the invisible, unverified drafting it was meant to prevent.

So build a workflow instead of a prohibition.

1. Approve specific tools under specific terms, and write down which ones. Confidentiality posture, not capability, decides this.
2. Verify every citation in a real citator. Not in the model, not in a search engine snippet. Pull the case.
3. Read every quoted passage in the original. Quotation accuracy is where misdescription hides.
4. Check currency. Reversed, superseded and depublished authority is invisible to a model trained on a snapshot.
5. Confirm the docket facts against the docket, including dates, party names and procedural posture.
6. Read the judge's standing orders for disclosure or certification requirements before filing.
7. Log who verified what. A one-line record per filing turns an unprovable claim into a documented process if anyone asks later.

Step 2 deserves emphasis because of a specific temptation. When a citation looks unfamiliar, the fastest move is to ask the model whether the case is real. Do not. It will tell you yes, for the same reason it produced the citation in the first place, and our piece on why a chatbot cannot vouch for its own output explains the mechanism in detail.

A certification line your firm can adopt today

For internal use, before a filing leaves the building:

I have verified in a current citator that every authority cited in this filing exists, remains good law for the proposition asserted, and is quoted accurately from the original source. I have confirmed all docket facts against the docket. Where generative AI tools assisted in drafting, the approved tool was [name] under [terms], and no client-confidential information was entered into an unapproved system.

Adapt the bracketed parts, attach it to the filing checklist, and require a name against it. The value is not legal magic. It is that a person has to affirmatively say the sentence, which is remarkably effective at making the verification actually happen.

What nobody can verify yet

Whether disclosure requirements will consolidate or fade is unknown. There is a reasonable argument that AI-specific standing orders are transitional, and that once verification norms settle, existing certification duties will absorb the whole problem the way they absorbed word processors and search engines. There is an equally reasonable argument that formal disclosure becomes permanent. Nobody can tell you which, and anyone selling certainty on it is guessing.

Nor can anyone tell you how well detection will handle frontier models. As newer model generations ship, the gap between detector training data and current output widens, and we cover that moving target in whether detectors can catch GPT-5-class writing. AI Detector 360 is not exempt from that drift, which is why every scan carries an explicit confidence level instead of a single confident number.

What is not uncertain is the duty. Every filing carries your signature, and your signature is a representation that you checked. The tool that produced the first draft has never been the question.