The citations look perfect. The case names are plausible. The reporters check out. None of it is real.
This isn't a hypothetical risk. It's already gone wrong in open court.
In June 2023, attorney Steven Schwartz filed a brief in Mata v. Avianca in the Southern District of New York. The brief cited six cases. Every single one was fabricated by ChatGPT. The case names sounded real. The reporter citations were formatted correctly. But none of the cases existed.
When the court asked Schwartz to produce copies of the opinions, he went back to ChatGPT and asked it to confirm they were real. ChatGPT said yes. He filed that confirmation with the court. The judge was not impressed. Schwartz and his colleague were sanctioned $5,000 and publicly reprimanded.
It happened again in Massachusetts. In Park v. Kim, an attorney submitted a brief with AI-generated citations that didn't hold up to scrutiny. The court flagged them.
These aren't isolated incidents. They're the ones that made the news. How many fabricated citations have been filed and never caught?
Key point: The attorney in Mata v. Avianca didn't know the cases were fake. He asked ChatGPT to verify its own work, and it confirmed the citations were real. You cannot use AI to check AI. You need an independent verification source.
ChatGPT is not searching a legal database. It has never read a case on Westlaw or CourtListener. It's a language model. It predicts what text should come next based on patterns it learned during training.
When you ask it for a case supporting your argument, it generates a statistically plausible combination of words. A realistic-sounding party name. A real reporter abbreviation like "So. 3d" or "F. Supp. 2d." A volume and page number that fall within the normal range for that reporter.
The result looks exactly like a real citation. Martinez v. Florida Department of Revenue, 347 So. 3d 218 (Fla. 4th DCA 2022). That's not a real case. But it looks like one. The reporter is real. The court is real. The volume number is plausible. The party name is believable.
ChatGPT doesn't know the difference between generating a real citation and a fake one. It's doing the same thing either way: predicting the next token. Sometimes it lands on something real. Often it doesn't.
Think of it this way: If you asked someone who had never been to Florida to make up a plausible-sounding Florida address, they might say "1847 Palm Beach Boulevard, Tampa, FL 33602." The street name sounds right. The ZIP code is real. But the address doesn't exist. That's what ChatGPT does with case citations.
The dangerous part isn't that ChatGPT makes things up. It's that what it makes up looks completely legitimate.
The citation format is correct. The reporter abbreviation is a real reporter. The volume number is within the published range. Sometimes the case name is close to a real case, just with a different year or a slightly different party name.
A surface-level check passes. Does this look like a proper citation? Yes. Is the reporter real? Yes. Is the court a real court? Yes. You'd have to actually search the case in a legal database to discover it doesn't exist. And if you're relying on ChatGPT because you don't have access to a legal database, you're stuck in a loop.
Even attorneys get tripped up. Schwartz was a practicing lawyer for 30 years. The citations looked right to him. He didn't verify them because they didn't raise any red flags on the surface.
Attorneys at least have access to Westlaw, LexisNexis, or a law library. They have the tools to verify. When they don't verify, that's negligence, and courts sanction them for it.
Pro se litigants don't have those tools. Many of them are turning to ChatGPT precisely because they can't afford an attorney. They're representing themselves in eviction proceedings, custody disputes, small claims cases. They need legal citations to support their arguments. ChatGPT gives them citations that sound authoritative.
They have no way to know those citations are fake. They can't afford Westlaw ($100+/month). They may not know about free alternatives like CourtListener or Google Scholar. And they trust ChatGPT because it presents information with total confidence, never hedging, never saying "I'm not sure about this one."
The result? Pro se litigants file motions with fabricated citations. Judges notice. The litigant's credibility is destroyed. Their case suffers. And they never understood the risk they were taking.
The credibility problem: Once a judge catches one fake citation in your filing, every other citation in your document is suspect. Even the real ones. You don't just lose the argument that citation supported. You lose the court's trust entirely.
The rule is simple: never file a citation you haven't independently verified. Not "verified by asking ChatGPT if it's real." Independently verified using a source that actually searches legal databases.
Free options for case law:
CourtListener (courtlistener.com) is free and searchable. It has millions of court opinions. If you search for a case name and it doesn't come up, that's a strong signal the case doesn't exist.
Google Scholar (scholar.google.com) has a "Case law" search option. It covers federal and state courts. It's not comprehensive, but it catches most published opinions.
The faster option: Upload your entire document to AI Detector Pro and let it check every citation at once. It searches CourtListener, Google Scholar, and multiple web search engines for each case citation in your document. You get results for all of them in one pass, not one at a time.
Most people stop at "does this case exist?" That's important, but it's only half the problem.
ADP's Verification Report checks existence. It searches CourtListener's database of millions of opinions, falls back to Google Scholar, and runs multi-engine web searches. If a case doesn't exist, you'll know.
But ADP's Deep Analysis goes further. It uses two independent AI models (Anthropic Claude and OpenAI GPT) to assess whether the case actually supports the legal argument you're citing it for. The models analyze independently. They never see each other's output.
When both models agree, you get a consensus assessment. When they disagree, the citation gets automatically flagged for human review. No single AI is trusted by itself. If one model says the citation supports your argument and the other says it doesn't, that disagreement is surfaced to you with both explanations so you can make the call.
That's the key difference. ChatGPT gives you one AI's confident-sounding answer with no way to check it. ADP gives you two independent assessments, shows you where they agree and disagree, and flags anything uncertain for closer review.
Bottom line: If ChatGPT generated your citations, or if you're not 100% sure every case in your filing is real and relevant, run it through ADP before you submit it. It takes less time than explaining fabricated citations to a judge.
