AI Detection for Lawyers: Why ChatGPT Briefs Get Flagged and What’s at Stake in Court

A first-year associate runs a motion through GPTZero before filing, just to be safe. It comes back 92% likely AI. The thing is, she wrote every word herself. Welcome to the strange new problem facing legal writers: the same tools that flag ChatGPT also flag good, careful, human-written law.

This matters more for lawyers than for almost anyone else, because in litigation the stakes aren’t a grade or a bounced freelance gig. They’re sanctions, reputation, and your standing with the court. Let’s untangle what’s actually happening, what a detector can and can’t prove, and how to use AI without ending up in a published opinion for the wrong reasons.

Key takeaways

• Legal prose is naturally low-perplexity and formulaic, which is exactly the pattern AI detectors associate with machine writing, so genuine briefs get flagged too.
• The real courtroom danger isn’t a detector score. It’s fabricated citations, like the ones that drew a $5,000 sanction against two attorneys and their firm in *Mata v. Avianca*.
• Many federal judges now have standing orders requiring attorneys to disclose or certify generative AI use.
• A detector output is a probability, not proof. OpenAI retired its own classifier in July 2023 for being inaccurate.
• You can use AI safely by verifying every cite, keeping a drafting record, and editing drafts so they read like a person wrote them.

Why legal writing trips detectors in the first place

Most AI detectors work by measuring how predictable your text is. They look at two things: perplexity (how surprised a language model is by your next word) and burstiness (how much your sentence length and rhythm vary). People write in fits and starts. A long, winding sentence, then a short one. A model, left to itself, smooths all that out and lands on the safe, average choice every time. If you want the full mechanics, detectors measure perplexity and burstiness to put a number on it.

Now think about what legal writing actually looks like. Defined terms repeated verbatim. Boilerplate recitals. Statutory language quoted exactly. “For the foregoing reasons,” “Plaintiff respectfully submits,” “It is well-settled that.” We’re trained to be consistent, to mirror the language of the controlling authority, to avoid creative phrasing. That discipline produces text with low perplexity and low burstiness.

In other words, the genre is engineered to look like the thing detectors flag. A good brief and a ChatGPT draft share a lot of surface DNA. That’s why a careful, human-written motion can score 90%-plus AI while a chatty blog post written by a bot might sneak under the threshold.

The thing that actually gets lawyers in trouble

Here’s the part that should reset your priorities. No lawyer has been sanctioned because a brief scored high on GPTZero. Lawyers get sanctioned because the brief contained things that weren’t real.

The landmark example is *Mata v. Avianca* (S.D.N.Y. 2023). Two attorneys submitted a brief that cited multiple court decisions to support their argument. The decisions didn’t exist. ChatGPT had invented them, complete with fake quotes and fake internal citations, and the lawyers filed them without checking. When opposing counsel and the judge couldn’t find the cases, the lawyers doubled down and even produced “copies” of the fabricated opinions, which ChatGPT had also made up. Judge P. Kevin Castel sanctioned the two attorneys and their firm $5,000, and the story went national.

Notice what the problem was. Not the writing style. Not whether a tool wrote it. The problem was a fabricated factual record submitted to a court. AI didn’t create a new category of misconduct here. It created a fast, confident way to commit an old one.

So the order of operations for any lawyer using AI is simple. Worry about substance first, style second. A brief that’s 100% accurate and 92% “AI” is fine. A brief that’s 5% “AI” and cites a hallucinated case can end your week.

What judges and bar rules actually require

The rules are still forming, and they vary by jurisdiction, but a few things are clear.

A number of federal judges have issued standing orders on generative AI. Some require attorneys to disclose whether AI was used in preparing a filing. Others go further and require a certification that a human checked any AI-assisted research for accuracy. These aren’t uniform, and they’re not nationwide, which is exactly why you have to read the specific judge’s standing orders and the local rules before you file. A surprise here is avoidable with five minutes of reading.

On the ethics side, the American Bar Association issued Formal Opinion 512 in July 2024, its first formal guidance on generative AI. The throughline is that the old duties still govern: competence, candor to the tribunal, confidentiality, and supervision. You’re responsible for what you file. Using AI doesn’t dilute that one bit.

So where does a detector score fit? It’s not a rule and it’s not evidence of wrongdoing. But practically, if you’re in a jurisdiction with disclosure requirements and your filing reads like raw model output, you’re inviting scrutiny you don’t need. A score isn’t proof of anything, yet it can still start an awkward conversation.

A quick scenario

Picture a solo practitioner with a brief due Friday. She uses ChatGPT to draft the standard-of-review section and the procedural history, then moves on. Three smart moves separate a safe outcome from a *Mata* repeat:

1. She pulls every cited case herself in Westlaw or Lexis and confirms each one says what the brief claims. Two “cases” don’t exist. She cuts them.
2. She checks the judge’s standing orders, finds a disclosure requirement, and includes the required certification.
3. She edits the draft so it sounds like her, not like a template, tightening the robotic rhythm and cutting filler.

Same tool, completely different risk profile. The AI saved her an hour. Her judgment saved her career.

How to read a detection score without panicking

If you run your own work through a detector and it lights up red, take a breath. As we explain in why a detection score is a probability, not proof, these tools are estimators, not lie detectors.

A few facts worth keeping in your back pocket:

• OpenAI launched an AI Text Classifier in early 2023 and then quietly retired it that July, citing a low rate of accuracy. The company that builds the model that writes the text couldn’t reliably detect that text.
• A 2023 Stanford study (Liang et al.) found that several GPT detectors flagged writing by non-native English speakers as AI-generated far more often than native writing, because both tend toward simpler, more uniform phrasing. Careful, formal prose looks “machine-like” to a detector regardless of who wrote it.
• Detectors disagree with each other constantly. The same paragraph can score 15% on one tool and 80% on another.

For a lawyer, the takeaway is twofold. Don’t treat a high score on your own honest work as a confession. And don’t treat a low score as permission to skip verification, because the score says nothing about whether your cases are real.

Using AI in legal work without the risk

You don’t have to choose between never touching AI and gambling with sanctions. A sane middle exists.

Verify everything that’s checkable. Every citation, every quote, every pin cite. Pull the actual authority. This single habit would have prevented the most famous AI legal disasters.

Keep a record of your process. Notes, drafts, research trails. If anyone ever questions a filing, a clear paper trail showing how you built it is your best friend. It’s the same logic behind keeping version history in any field where authorship gets challenged.

Edit for a human voice. Raw model output has tells: the relentlessly even sentences, the stock transitions, the over-polished hedging. Rewriting a draft so it reads like you actually wrote it isn’t about deceiving anyone. It’s about clarity and ownership, the same reason you’d never file a junior associate’s draft without making it your own. An AI humanizer can help break up that robotic rhythm so your brief reads naturally, but the substance is still yours to stand behind.

Know your court. Read the standing orders. Read the local rules. Disclose when required. When unsure, lean toward disclosure.

If you want to go deeper on any of these, you can browse more on AI detection and writing, or see the plans if you’re handling volume.

Frequently asked questions

Can a judge tell if I used ChatGPT to write my brief?

Not with certainty, and neither can a detector. Tools like GPTZero and Originality.ai produce a probability, not a verdict, and legal writing’s formal style tends to score as AI-like even when a human wrote it. What gets lawyers in trouble is substance, like fabricated cases or quotes that don’t check out, which judges catch by reading the brief and the authorities, not by running a detector.

Is it unethical to use AI to draft legal documents?

Using AI as a drafting tool isn’t inherently unethical, and bar guidance like ABA Formal Opinion 512 treats it as technology a competent lawyer supervises. Your duties of competence and candor don’t change. You’re responsible for everything you file, so verify the law and follow disclosure rules. The problem in sanctioned cases was filing AI output without checking it.

Do I have to disclose that I used AI in a court filing?

It depends on the court. There’s no single nationwide rule, but a number of federal judges have issued standing orders requiring attorneys to disclose or certify generative AI use. Check the specific judge’s standing orders and your local rules before filing, and when in doubt, disclose.

Why does my brief get a high AI score even though I wrote it myself?

Because legal writing is formulaic. Defined terms, boilerplate, and statutory language create low perplexity and low burstiness, the same signals detectors read as machine-generated. A human-written motion can score as AI-like for the same reason a non-native speaker’s careful prose does. The score reflects predictability, not authorship.

The bottom line

Detectors will keep flagging legal writing, because good legal writing is, by design, predictable. That’s a style problem, not a misconduct problem. The misconduct problem is filing things you didn’t verify. Get that order right, edit your drafts so they sound like a person, and AI becomes what it should be: a fast first draft you bring your judgment to.

If you’ve got a draft that reads too much like a machine, run it through PaperBleach and make it sound like you again.