Key takeaways
- A public database tracked roughly 1,490 court decisions worldwide involving AI-hallucinated material by May 2026 — more than 1,000 in the United States.
- Every major tool has been implicated: ChatGPT, Claude, Gemini and some legal-specific platforms.
- Penalties have escalated from $5,000 fines to $15,000 per attorney plus fees, and an indefinite suspension.
- Courts punish the failure to verify, not the use of AI. A disciplined checklist removes almost all of the risk.
The scale of the problem
Legal researcher Damien Charlotin maintains a public database of court decisions in which a party relied on AI-hallucinated content and the court responded. As of May 2026 it held roughly 1,490 decisions worldwide, more than 1,000 of them from US courts. According to one tracker, US courts imposed at least $145,000 in sanctions for AI citation errors in the first quarter of 2026 alone.
These are not only self-represented litigants experimenting with chatbots. Licensed attorneys, experienced firms and expert witnesses appear throughout the record — and, as the trackers note, licensed lawyers tend to face steeper consequences.
Why AI invents cases (in plain English)
A large language model does not look up answers. It predicts the most plausible next words based on patterns in its training. A case citation has a very recognisable pattern — party names, a reporter, a volume, a page, a year — so the model can produce something that looks exactly like a real authority without any real authority behind it.
The danger is not that the output is obviously wrong. It is that it is confidently, fluently plausible, often with a quotation that sounds like judicial prose. That is why reading it is not enough. You have to find it.
The cases that matter
You do not need to know all 1,490. These are the decisions lawyers most often cite when explaining the risk, and what each one teaches.
Counsel filed a brief citing six decisions that did not exist, generated by ChatGPT and complete with fabricated quotations. Judge P. Kevin Castel sanctioned the lawyers and their firm $5,000. Lesson: the case that put the whole profession on notice.
An expert declaration was struck after it cited academic articles that did not exist; the drafting had relied on ChatGPT-4o without verification. Lesson: the risk extends to experts and evidence, not only briefs.
Eight of nine cited cases were fake. The court imposed $5,000 in total sanctions and revoked a pro hac vice admission. Lesson: signing counsel share responsibility for what they sign.
A brief contained nearly thirty defective citations; the court fined each attorney $3,000. Lesson: high-profile clients do not make high-quality AI output.
More than two dozen fabricated citations on appeal led to $15,000 per attorney, plus opposing counsel’s full appellate fees and double costs. Lesson: appellate courts are treating this as a serious abuse.
In a divorce appeal, 57 of 63 citations were defective. The lawyer was suspended from practice indefinitely. Lesson: the consequence is no longer just a fine — it can be your licence.
How penalties escalated — and what made them worse
Reading the decisions together, three patterns stand out:
- Penalties rose sharply. From a $5,000 firm-wide fine in 2023 to $15,000 per attorney and suspension in 2026.
- Candour helped; concealment hurt. Lawyers who denied or minimised the problem generally fared worse than those who acknowledged it early.
- “The tool was built for lawyers” was no defence. Decisions have named purpose-built legal tools as well as general chatbots — and, in at least one reported matter, a citation-checking step that missed the fabrications.
The 7-step verification checklist
Use this before anything AI-assisted leaves your desk. Tick items as you go — your progress is saved in this browser.
Build the habit into how you prompt, too: ask the tool to give paragraph or page references for every proposition, and to say “I am not certain” rather than guess. It does not remove the need to check — it makes checking faster.
AI for Lawyers 2026
Prompting methods, professional-responsibility duties and 100 prompts designed to be verified, not blindly trusted.
See it on AmazonIf you’ve already filed something with a fake citation
This is general information, not advice on your situation. But the decisions point in a consistent direction:
- Move quickly. The longer an error sits on the record, the worse it looks.
- Tell the court and opposing counsel candidly, and withdraw or correct the filing.
- Explain what happened without minimising it, and what you have changed.
- Get advice from your firm’s general counsel, professional-responsibility counsel or your bar’s ethics helpline.
Firm-level safeguards
Individual care is necessary but not sufficient. The ABA’s Formal Opinion 512 makes clear that supervisory duties apply to AI use by lawyers and staff. In practice, that means:
- A short written AI-use policy naming approved tools, prohibited uses and verification requirements.
- Training on real firm tasks, not generic demos.
- A second-reader rule for AI-assisted court filings.
- A record of court standing orders on AI disclosure in the jurisdictions where you appear.
For the policy itself, TECHLAWG drafts AI governance documents; for the training, see our organisational programmes. Choosing tools with better grounding helps too — see the best AI tools for lawyers in 2026.
Frequently asked questions
Can lawyers be sanctioned for AI hallucinations?
Yes — fines, revoked admissions, fee awards and, in 2026, indefinite suspension. Courts sanction the failure to verify, not AI use itself.
Is it safer to use a legal-specific AI tool?
Grounded tools lower the risk but do not remove it. Sanctions decisions have involved legal-specific platforms as well as general chatbots. Verify regardless.
Do I have to tell the court I used AI?
It depends on the court. A growing number of judges and courts have standing orders or rules on disclosure. Check the rules for every court you appear in.
Can I bill for time spent verifying AI output?
Under ABA Formal Opinion 512, fees must be reasonable and reflect actual work. Time genuinely spent reviewing and verifying is work; time you would have saved by not using AI poorly is not a reason to bill more.
Make verification a habit, not a hope
Our programmes teach safe AI workflows on real legal scenarios.