Treat every citation an AI tool produces as unverified until a named lawyer has opened the source and confirmed it says what the draft claims. Put that in a written rule, prefer tools that link to sources, train the team, and log who checked each citation before anything is filed. This is not legal advice.
#Key takeaways
- Fake citations come from unverified drafts, not from one careless person; the fix is a checking rule everyone follows.
- Courts in several jurisdictions have sanctioned or referred lawyers who filed citations that did not exist.
- Even specialist legal research tools produce wrong or misgrounded answers, so a paid tool is not a substitute for checking.
- The partner or senior lawyer who signs is responsible for the juniors' research, AI-assisted or not.
- The training that matters is short, practical and on your own research tasks, not a general AI course.
#How do fake citations end up in a filing?
The pattern is familiar. A junior lawyer under pressure asks a chatbot for authorities, gets a tidy list with case names, courts and quotes, and pastes it into a draft. Nobody opens the cases, because they look real. The senior lawyer reviews the argument, not the footnotes. The other side or the judge tries to find the cases and cannot.
In Mata v. Avianca, a new york federal court case, a lawyer used ChatGPT for research and filed cases that did not exist (CNBC, may 2023). It was not the last.
#What are courts doing about AI citations?
In england, the high court's Ayinde judgment dealt with submissions containing fictitious authorities and signalled that fake citations are likely to lead at least to a regulator referral and a wasted costs order (Linklaters, june 2025).
Closer to home, the DIFC Courts' guidance on large language models asks parties to declare AI-generated content early and to verify its accuracy (DIFC Courts, december 2023). Check the rules of each court you appear before.
#Are paid legal research tools safe from hallucination?
Safer, not safe. A Stanford study found that leading legal research tools still gave incorrect or misgrounded answers in more than one in six benchmark queries; misgrounded means the cited source does not support the statement (Stanford HAI, may 2024). A real case cited for a proposition it does not support is harder to catch than an invented one.
#What should a citation verification rule say?
- No authority goes into a draft until someone has opened it in a primary source or an official reporter.
- The checker confirms the case exists, the citation is correct, the quote is verbatim and the case supports the point made.
- The checker's name is recorded against each authority in the file.
- Public chatbots are not used for research on client matters, and client facts are never pasted into them.
- The signing lawyer confirms the check was done before filing, and the court's AI disclosure rules are followed.
This works with no new software. A shared checklist in the matter file and a line in the filing sign-off are enough to start.
#What training actually prevents this?
A general AI awareness session does not. What works is short, hands-on practice on your own research tasks: asking a tool for authorities, then finding the errors in its answer. People remember the fake case they caught themselves. The session should cover the verification rule, the court rules that apply to your team, and what not to paste into which tool.
Only 39% of AI users said their company had given them AI training (Microsoft and LinkedIn Work Trend Index, may 2024). In a legal team, that gap shows up in the footnotes.
#Where does Stennir fit?
Stennir's AI training runs the AI operator program for legal teams as role-specific cohorts on your real research and drafting work, in Arabic and English, led by practitioners, with usage measured before and after and an adoption report at the end. AI consulting writes the governance around it: which tools are approved for which tasks, who verifies, and what gets logged.
If research support is worth building, AI implementation can set up a tool that only returns authorities from sources you subscribe to, with a link to each. See also how to train staff who already use ChatGPT with company data.
We are
Stennir trains legal teams to verify AI output on their own work and writes the rules for who checks what.
We aren't
Stennir is not a legal research provider and does not check citations or give legal opinions for you.
#What should a managing partner or general counsel do next?
Write the verification rule as one page and add a citation check line to your filing sign-off. Then train the people who draft. See how compliance and legal teams use AI on AI for compliance, take the free readiness assessment or book a 30-minute call.
#Common questions
Should we ban AI tools for legal research?
A ban usually pushes use onto personal accounts, where you cannot see it. A better rule names approved tools, bans client facts in public chatbots, and requires every authority to be opened and checked by a named person. That keeps the speed of AI drafting and removes the reason citations go unchecked.
Who is responsible for a fake citation drafted by a junior lawyer?
The lawyer who signs the filing, and possibly supervisors too, depending on your jurisdiction's professional rules. Courts have looked at the conduct of senior lawyers as well as the person who drafted. That is why the check belongs in the sign-off process, with a name recorded against it.
Does Stennir run this training in Arabic?
Yes. Stennir runs the AI operator program in Arabic and English, on your team's own research and drafting tasks, with exercises on legal Arabic sources where they apply. Other languages are available on request. Usage is measured before and after, and the adoption report goes to the person who sponsored the training.
What could go wrong even with a verification rule?
People can tick the box without opening the source, or check that a case exists but not that it supports the point. Sampling filed documents for another check, recording names against each authority, and training on misgrounded citations, not only invented ones, closes most of that gap.
