TL;DR: Fake AI citations have led to real sanctions, starting with Mata v. Avianca in 2023 and a growing line of cases since, and some judges now require lawyers to disclose AI use. Fabrication happens because ungrounded language models predict plausible text rather than retrieve real sources. The fix is practical: use grounded tools that cite primary sources, never cite a case you have not read, verify quotes verbatim, and check good law.
In 2023, two New York lawyers became a cautionary tale when they filed a brief citing cases that did not exist. Their chatbot had invented the authorities, complete with quotations and citations, and they had not checked. The court was not amused, and the episode - Mata v. Avianca - became the reference point for every conversation about AI and legal ethics since. The lesson is not that AI is too dangerous to use. It is that a specific, avoidable failure mode has a specific, learnable set of safeguards. This guide lays them out.
Why do fake AI citations keep getting lawyers sanctioned?
Mata v. Avianca was decided in the Southern District of New York in 2023, and the court sanctioned the lawyers $5,000 for submitting fabricated citations and then standing by them when challenged. What made it memorable was not the size of the penalty but the detail: the AI had produced fake case names, fake reporter citations, and fake quotations, and a follow-up prompt asking the tool to confirm the cases were real produced false reassurance. It was a complete, self-contained fabrication that any check against the primary source would have caught.
The problem did not end there. Since 2023 a growing line of sanctions orders, in multiple jurisdictions, has dealt with lawyers and self-represented litigants who filed AI-hallucinated authority, sometimes with larger penalties and referrals to disciplinary bodies. In response, some judges have issued standing orders requiring lawyers to disclose whether AI was used in a filing and to certify that any AI-assisted research was checked. You can look up the underlying decisions on CourtListener, and the federal judiciary publishes court information and rules at uscourts.gov. For a deeper account of the original case, see our breakdown of AI hallucinations and Mata v. Avianca.
Why does AI fabricate citations in the first place?
To prevent fabrication you have to understand why it happens, and the answer is in how large language models work. A general-purpose model generates text by predicting the most plausible next words given everything before them. It is not looking anything up. When you ask for authority on a point, it produces text shaped like a citation - a plausible case name, a plausible reporter, a plausible quote - because that is the pattern it learned, not because it retrieved a real document. The output is fluent and confident precisely because fluency, not accuracy, is what the model optimises for.
This is why a fabricated citation looks so convincing. It has the right form: parties, a volume and reporter, a year, a pinpoint, and a quotation in judicial cadence. The model has seen millions of real citations and can imitate the shape perfectly while inventing the substance. Asking the same model whether its citations are real does not help, because that question just prompts another round of plausible-sounding text. The only reliable test is external: does the cited source actually exist, and does it say what the model claims?
Grounded (RAG) tools vs ungrounded chatbots: what is the difference?
The single most important choice you make is the kind of tool you use. An ungrounded chatbot answers from its trained parameters; a grounded tool retrieves real documents first and then answers from them. That retrieval step - often built with a technique called retrieval-augmented generation, or RAG - is what ties the answer to sources that exist. Instead of predicting what a relevant case might say, a grounded system searches a real corpus, pulls the actual passages, and generates an answer anchored to them, with citations you can open.
The practical difference is night and day. With an ungrounded tool, a citation is a prediction you have to disprove. With a grounded tool, a citation is a pointer to a document you can read. Grounding does not eliminate the need to verify - retrieval can surface the wrong passage, and you still confirm relevance - but it changes the base rate of fabrication from a real hazard to a near-non-issue. For more on the broader practice, see our guide to verifying AI legal research.
One caution: not every tool that shows citations is truly grounded. Some ungrounded systems generate a fluent answer first and then attach citations after the fact, which can be just as fabricated as the prose they decorate. The test is whether you can click a citation and land on the actual source passage the answer was built from. If the tool cannot take you to an openable primary document - a real judgment, statute, or archived page - treat its citations with the same suspicion you would give an unsourced summary. Genuine grounding shows its work; citation theatre only looks like it does.
What workflow stops a fake citation before it reaches a filing?
Knowing the failure modes, you can build a workflow that catches each one. The table maps the ways AI citations go wrong to the safeguard that stops them - none of which takes more than a few minutes per authority.
| Failure mode | Why it happens | Safeguard |
|---|---|---|
| Cited case does not exist | An ungrounded model predicted a plausible name | Use a grounded tool that links to the primary source, then open it |
| Real case, wrong holding | The model summarised from memory, not the text | Read the cited passage and confirm it supports your point |
| Quote is paraphrased or invented | The model reconstructed rather than copied | Verify every quotation verbatim against the source |
| Case is no longer good law | The model has no current-status awareness | Check subsequent history with a citator before citing |
| Right case, wrong pincite | The model guessed the page or paragraph | Confirm the page and paragraph in the actual judgment |
The thread running through every row is the same discipline: never let a citation into a filing that you have not personally traced to a real, current source. Grounded tools make that fast by handing you the document; the verification is a confirmation rather than a fresh search. Skip it, and you are trusting a system that was designed to sound right, not to be right.
A do-and-don't checklist for AI-assisted citations
If the workflow distils to a single habit, it is this: treat AI as a research assistant whose every output you check, never as an authority you quote. The checklist below is the short version you can pin above your desk.
| Do | Don't |
|---|---|
| Use tools that cite to primary sources you can open | Rely on a chatbot's unsourced summary of the law |
| Read every case in the original before you cite it | Cite a case from a list of authorities alone |
| Verify quotations word-for-word | Trust a quotation because it sounds right |
| Check good law with a citator | Assume a case still stands |
| Keep a human in the loop and disclose AI use where required | File AI output unread |
Most of the sanctioned cases would never have happened if the lawyer had honoured even one column of that table. The failures were not sophisticated; they were citations filed unread. Keeping a human in the loop on every cited authority is the whole game.
What are judges and the duty of candor now requiring?
Beyond the practical risk, there is a professional duty in play. The duty of candor to the tribunal - ABA Model Rule 3.3 - requires that a lawyer not knowingly make a false statement of law to a court, and a fabricated citation is exactly that. Courts have made clear that not knowing your citation was fake is no defence when you failed to check; the obligation to verify is part of the duty. The ABA Model Rules of Professional Conduct set the baseline, and individual courts are layering specific AI requirements on top.
Those requirements increasingly take the form of standing orders and certifications. Some judges ask you to disclose whether AI was used to prepare a filing; others require a certification that a human checked any AI-assisted research; many simply remind practitioners that existing rules already cover the situation. Whatever the local rule, the safe assumption is that you are personally certifying every citation you file - so make that true before you sign. Filing fabricated authority can also feed a malpractice claim, which we cover in AI and malpractice risk for lawyers.
How does Judicio make fabricated citations almost impossible?
Judicio is built around the property that prevents fabrication: grounding with verifiable citations. Every finding, answer, and date in Legal Research cites the exact page and the quoted passage it relied on, and the citation labels are deterministic - generated from the source, never written by the model. That means a citation is always a pointer to a real document region you can open and read, not a prediction you have to disprove. The fabrication failure mode that snared the lawyers in Mata v. Avianca is designed out.
Two more features close the loop. Every web source is archived as a permanent PDF at the moment it is retrieved, so the authority you cited cannot quietly change or vanish - your citations do not rot. And you can export an evidence pack containing every source, cited to the page, so a senior, an opponent, or the bench can trace exactly where each proposition came from. None of this removes your duty to read the cases; it makes honouring that duty fast enough to do every time. The same discipline applies across Document Review and the Review Matrix, where every answer is cited to the page.
How do you put this into practice?
Putting this into practice takes three moves. Switch your default research tool to one that grounds answers in primary sources and cites them. Adopt the verification habit - read the case, confirm the quote, check good law - on every authority, without exception. And keep a record of your sources, so you can show your work if a court or a partner asks. Do those three things and the fabricated-citation risk effectively disappears, not because AI stopped predicting plausible text, but because you stopped filing it unchecked.
You can try grounded, citation-first research on your own matters with Judicio's 7-day free trial - 500 credits, no credit card required - and see what page-and-passage citations and archived sources feel like in practice. For a walkthrough geared to your court's AI rules, contact us.
Judicio provides legal research and drafting tools, not legal advice. Verify every citation against the primary source and apply your own professional judgment before filing.