AI hallucinations and lawyers: sanctions so far and a citation-verification workflow for small firms

Have lawyers really been sanctioned for AI-invented citations?

Yes. Courts have criticised or sanctioned lawyers who filed documents containing case citations, quotations or holdings that a generative AI tool invented and nobody checked. The best-known public record is Damien Charlotin's AI Hallucination Cases database, which says it tracks legal decisions where a court addresses the use of AI, whether established or merely alleged, in more than a passing reference.

Two things matter before you quote any number. First, that database covers all parties, including self-represented litigants, not only lawyers. Second, secondary write-ups of it are snapshots. Last reviewed: 30 October 2026.

How many cases are there, and why do the numbers differ?

We could not open the database itself in preparing this page, so the figures below come from a published audit of it and from commentary, labelled as such. Open the live database for today's count.

The numbers differ because trackers choose different scopes (worldwide or US only), different parties (lawyers or everyone), and different thresholds (court finding versus allegation). A count is only quotable with its definition and its date.

What exactly goes wrong in these filings?

Recurring patterns in the reported cases are consistent with how language models work: a plausible citation that does not exist; a real case that does not say what the brief claims; a genuine quotation attached to the wrong case; and a real case that has since been overturned. Each needs a different check.

Checking only that a case 'exists' catches the first row and none of the others.

What is an 8-step citation-verification workflow?

1. 1

   1. Mark AI-assisted text

   Record at the start that a tool was used and for what: research summary, first draft, cite-check, or nothing. Unflagged use is the root of most failures.

2. 2

   2. Extract every authority

   List every case, statute, rule, quotation and pinpoint in the document, including those in footnotes.

3. 3

   3. Open each source

   Retrieve the authority from a primary or authoritative database, not from the AI's own output or a summary of it.

4. 4

   4. Read the passage

   Confirm the court, date, and that the passage supports the proposition. Paste the relevant text into the log.

5. 5

   5. Match the quotes

   Compare every quotation character by character with the source.

6. 6

   6. Check status

   Use a citator or equivalent to confirm the authority has not been overturned or limited.

7. 7

   7. Second pair of eyes

   A person other than the drafter signs off on the log. Solo practitioners can use a different day and a printed list.

8. 8

   8. Keep the log

   File the completed log with the matter. If you correct or remove a citation, record that as well.

How do I build a verification log in Notion?

A log is one database with one row per checked document or per authority. Suggested properties:

• Matter (relation to your Matters database)
• Document and version
• AI tool used and purpose (select)
• Authorities checked (number) and authorities removed or corrected (number)
• Source database used (select: your research service, official court site, other)
• Checker (person) and second reviewer (person)
• Date checked and result (Verified, Corrected, Removed)
• Notes (paste any discrepancy)

Add a Notion rollup to your Matters view that shows whether a filing has a completed log, and a filter for 'filing due within 7 days, no log'. That turns the policy into something visible before the deadline rather than after.

What do the regulators say?

Regulators have not said 'do not use AI'. They say the lawyer remains responsible. ABA Formal Opinion 512 (July 2024) covers competence, confidentiality, candour and supervision for generative AI. The Solicitors Regulation Authority's August 2026 warning notice, as summarised by law-firm commentary, says solicitors remain accountable for AI output and for work done under their supervision. California's State Bar replaced its 2023 guidance in 2026, and summaries report it says no document should go to a court without attorney review and approval.

None of that is a rule number we have quoted from the source text, so read the instrument that applies where you practise.

• SRA warning notice: what small firms should check
• Supervising AI agents: a checklist
• ABA Model Rules for cloud and AI

Where LawyerOS fits

We make LawyerOS. LawyerOS Pro with AI ($39, one-time) has a Matters database and a Research database, so you can link a verification log to a matter and its research. It does not come with a ready-made verification log: you add the properties above yourself, and the template does not check citations, connect to a citator or give legal advice. LawyerOS Starter ($19) covers matters and deadlines only.

• LawyerOS Starter and Pro with AI: compare and buy
• What LawyerOS includes

Frequently asked questions

What happens to lawyers who file fake AI citations?

Outcomes reported in the trackers range from judicial criticism and orders to explain, to monetary sanctions and referrals to regulators. Severity varies by court and by whether the lawyer was candid once the problem surfaced. Read the decisions, not headlines, and check your own jurisdiction's rules on candour and supervision.

How do I verify AI legal citations?

Open each authority in a primary or authoritative database, read the passage, match every quotation, check that the authority is still good law, and have someone other than the drafter sign off. Record each step in a log. Never treat the AI tool's own output as the source.

Is there a database of AI hallucination cases?

Yes. Damien Charlotin maintains an AI Hallucination Cases database of legal decisions where a court addresses AI use in more than a passing reference. It covers all parties, not only lawyers, and its count changes constantly, so cite the number with its date.

Does using AI for research break the ethics rules?

Not by itself. Regulators such as the ABA and the SRA say existing duties of competence, confidentiality, candour and supervision continue to apply. The risk is unverified output reaching a court or client. Check your jurisdiction's guidance and your insurer's requirements.

Why do AI hallucination statistics disagree?

Trackers define scope differently: worldwide or one country, lawyers or all litigants, court findings or allegations. Secondary articles often copy a number without its date. Quote the original source, the definition used and the as-of date, and treat unsourced totals as unreliable.