In 2023, a New York attorney submitted a brief citing six cases that didn't exist. ChatGPT made them up. The judge sanctioned him and his firm. Three years later, attorneys are still getting disciplined for citing AI hallucinations. Here's why ChatGPT can't safely do legal research — and how Ark solves it.
Roberto Mata sued Avianca Airlines after a metal serving cart struck his knee on a flight. His attorney, Steven Schwartz of Levidow, Levidow & Oberman, filed a brief opposing Avianca's motion to dismiss. The brief cited six federal cases, including Varghese v. China Southern Airlines, Shaboon v. Egypt Air, and Miller v. United Airlines.
Avianca's lawyers tried to find the cases. They couldn't. Neither could the judge. They didn't exist.
Schwartz had asked ChatGPT to do the research. ChatGPT confidently produced quotations, case names, citations, and even fake docket numbers. Schwartz copy-pasted them into his brief without verifying.
Judge P. Kevin Castel sanctioned Schwartz and his firm $5,000 and required them to notify each judge falsely identified as the author of the fake opinions. The case became the defining cautionary tale for lawyers using generative AI.
ChatGPT is a general-purpose language model. It predicts the next most likely word based on patterns from its training data. When you ask it for a case citation, it doesn't search a legal database — it generates text that looks like a case citation.
A fake citation like Varghese v. China Southern Airlines, 925 F.3d 1339 (11th Cir. 2019) has all the features of a real citation:
It looks exactly like a real case. It just isn't one. ChatGPT generated it the same way it generates any other fluent English — by pattern-matching on what citations usually look like.
Even in 2026, after multiple model upgrades, ChatGPT continues to fabricate citations. OpenAI has added warnings, but the underlying architecture doesn't retrieve from a verified legal database. It generates.
Ark is built on a different architecture: retrieval-augmented generation with strict citation binding. Here's the pipeline:
Because the AI can only cite cases that exist in the retrieval step, it cannot fabricate them. If Ark doesn't find a relevant case, it tells you. It doesn't make one up.
| Feature | Ark | ChatGPT |
|---|---|---|
| Legal database | 7.4M cases, 26.6M salient points | None (training data only) |
| Citation verification | Every citation grounded in retrieved case | Unverified — hallucinations common |
| Clickable citations | Yes — link to actual case text | No |
| State-specific search | Yes, per-state indexes | No |
| Court rules coverage | 142K+ rules, all 50 states | Inconsistent, often outdated |
| Statute coverage | All 50 states + DC, current | Cutoff-dependent, not authoritative |
| Sanctions risk | Low — every cite is real | High — documented case law |
| Cost for legal use | $29/month | $20/month (but not safe for filings) |
Don't risk sanctions. Use AI legal research that cites real cases, not hallucinations. Try Ark free — no credit card required.
Start Free TrialMata wasn't an isolated incident. Since 2023, courts have sanctioned lawyers repeatedly for filing briefs with AI-generated fake citations:
The pattern is consistent: an attorney uses ChatGPT (or a similar general-purpose LLM), copies citations without verification, opposing counsel catches the fabrication, court sanctions follow. The fix is simple — use a tool that can't hallucinate.
Yes, in theory. In practice, verification means looking up every single citation ChatGPT produces — which takes longer than doing the research the traditional way. You're also unlikely to catch every hallucination, because fake citations look exactly like real ones. Tools like Ark that verify citations automatically give you AI speed without the verification burden.
Regular ChatGPT generates text based on patterns in its training data. Retrieval-augmented AI (like Ark) first searches a real database for relevant documents, then generates answers using only those retrieved documents as context. The AI cannot cite what it doesn't retrieve. This eliminates the category of error that caused Mata v. Avianca.
Yes. Any general-purpose LLM — Claude, Gemini, Llama, ChatGPT — will fabricate legal citations when asked. The fix isn't switching models; it's adding retrieval grounding. Ark uses Claude Opus 4.7 as its language model, but Opus never generates free-form citations — it only cites what Ark's retrieval pipeline provides.
Some judges in the 5th Circuit, Eastern District of Texas, and Northern District of Illinois have issued standing orders requiring disclosure of generative AI use in filings. More jurisdictions are likely to follow. When you use Ark, you can disclose AI assistance and still rely on your citations because every one is a real case.
Submitting fabricated citations without verification is a serious violation of Rule 11 (federal) and analogous state rules requiring attorneys to certify the accuracy of filings. Multiple bar associations have issued ethics opinions classifying AI-hallucination-based filings as potential malpractice. Using verified tools like Ark substantially reduces this risk.
ChatGPT is a remarkable general-purpose tool. It's not a legal research tool. Using it for case research is like using a stopped clock to set your watch — it looks right, but the accuracy is coincidental.
Use ChatGPT for: brainstorming, drafting non-citation prose, summarizing documents you provide it, rephrasing passages, generating first drafts.
Don't use ChatGPT for: finding cases, finding statutes, finding court rules, generating citations, anything that gets filed with a court.
Use Ark for: AI-powered legal research where every citation is a verified, clickable, real case. $29/month. No hallucinations.
Stop risking sanctions. Ark gives you the speed of AI with the safety of verified citations.
Get Started FreeRelated articles: Ark Legal AI home • Pricing plans • Ark vs Westlaw • Ark vs Casetext / CoCounsel