Damien Charlotin's Hallucination Tracker, Read Like a Risk Manager

The AI Hallucination Cases database is a public list that Damien Charlotin maintains of court decisions in which a party relied on AI-hallucinated content, usually fake citations. The Fifth Circuit describes him as "a French lawyer and data scientist." On 12 October 2026 the page said it was last updated on 10 October 2026 and listed 2,184 cases identified, 1,492 of them in the United States (damiencharlotin.com/hallucinations). Read as news it is a list of embarrassing footnotes. Read as a risk manager would, it shows who gets caught, which tools show up in the orders, and how courts respond.

Reviewed and corrected on 12 October 2026. Counts marked "our count" come from the database's CSV download, which we pulled on 12 October 2026. This is general information, not legal advice; pull the underlying order before relying on any entry.

TL;DR

• It counts decisions that address hallucinated content, not sanctions. The page's own footnote says it covers decisions where AI use, "whether established or merely alleged," is addressed in more than a passing reference, and that it does not track the wider universe of fake citations.
• US entries involving a lawyer rose fast. By our count of the CSV, they went from 4 in 2023 and 14 in 2024 to 207 in 2025 and 361 so far in 2026. Pro se litigants account for 885 US entries and lawyers for 587.
• The tool is usually not recorded. For most US lawyer entries the tool field says "Implied" or "Unidentified." Where a tool is named, ChatGPT is the most common, but tools sold to lawyers appear too.
• Money varies widely. The largest US lawyer-involved penalty in the CSV is $110,204.38 in Couvrette v. Wisnovsky, and the median dollar figure is $3,000.
• Several courts have said the lawyer's response mattered. The orders we quote below tie the penalty to candor after the problem surfaced.

Part of our legal AI verification and hallucination guide series.

For related coverage, see AI Hallucination Sanctions: Lawyer Cases and How to Avoid Them and How to Verify AI Legal Citations Before You File. Our own curated list of US orders is the sanctions tracker.

What the database is

The page's footnote defines it: it tracks "all documents where the use of AI, whether established or merely alleged, is addressed in more than a passing reference by the court or tribunal." It says it does not cover "mere allegations of hallucinations," only cases where the court or tribunal has "explicitly found (or implied)" that a party relied on hallucinated content, with an exception for some decisions where AI use was alleged but not confirmed, which it calls a judgment call. It also says plainly that it "does not track the (necessarily wider) universe of all fake citations or use of AI in court filings."

Each row has a case name, court, state or country, date, the type of party, the AI tool, the hallucination items, the outcome, any monetary penalty, whether a professional sanction was imposed, and a pointer to the source. The data is licensed CC BY 4.0, free to reuse with attribution. On the date we looked, the page's country filter listed entries from dozens of countries, and the United States had 1,492.

The Fifth Circuit used it in Fletcher v. Experian (18 February 2026), writing that Charlotin "has identified 239 cases of hallucination by lawyers in the United States." Our count of US entries with a lawyer in the party field was 587 on 12 October 2026. The two counts may not be defined in exactly the same way, but it gives a sense of the pace.

What we counted from the CSV

We downloaded the CSV on 12 October 2026 and counted the US rows (State(s) = USA). The "party" field is free text and can name more than one type (for example "Lawyer; Judge"), so we matched the lawyer group by substring and the pro se group by its exact value; the two groups overlap in one row. About 20 US rows name neither (experts, judges, prosecutors and others).

The largest US lawyer-involved figure is $110,204.38 in Couvrette v. Wisnovsky (orders of 12 December 2025 and 23 March 2026, which we break down in the Oregon sanction piece). Dollar figures include fee awards as well as penalties, so they mix different things. One row in the CSV has no date.

Which tools appear

For the 587 US lawyer entries, the tool field is filled in for nearly all of them, but 321 say "Implied" and 145 say "Unidentified." A tool is named in 117 entries. Of those, about 57 mention ChatGPT or a GPT model, 14 mention Claude, 6 mention Copilot, 5 mention Gemini and 3 mention Perplexity, and roughly 30 to 40 name a product sold to lawyers, such as CoCounsel, Westlaw, Lexis, Spellbook or Strongsuit (some entries name more than one tool, and the exact figure depends on which products you count as legal tools). So the fair reading is narrow. ChatGPT is the most common named tool, and tools sold to lawyers are in the data. The most common answer, though, is that nobody knows which tool was used.

The orders show the same thing. In Farris (6th Cir., 3 April 2026), counsel admitted using Westlaw's CoCounsel and the court said the risk remains "even when new tools are sponsored by trusted legal technology providers." In Fletcher, counsel named Casetext and vLex among her sources and the court did not find which produced the errors.

A policy that bans "ChatGPT for research" would miss part of this. The rule that fits the data is that every citation and quotation is checked against a primary source, whatever tool produced it.

What kinds of disputes

The database has an area-of-law field. For US lawyer entries the most common values are contract (135), tort (76), civil rights (68) and employment (61), with 63 blank. We do not know what proportion of filings in those fields involve AI, so this is a description of the database, not a rate. The takeaway for a risk manager is simple: the entries are spread across ordinary commercial and civil work, not only high-profile litigation. The fixed point is the court filing, not the size of the client.

How courts responded

The database is a pointer. These are the orders behind some of its better-known entries, read against the source:

• Mata v. Avianca (S.D.N.Y., 22 June 2023): $5,000 penalty. The court found bad faith on the part of the two individual lawyers "based upon acts of conscious avoidance and false and misleading statements to the Court."
• People v. Crabill (Colorado PDJ, 22 November 2023): suspension of one year and one day, 90 days served and the rest stayed on a two-year probation. The approved summary records that he found the problem before the hearing, did not alert the court, and falsely blamed a legal intern.
• Couvrette v. Wisnovsky (D. Or., 12 December 2025): the court found that deleting fake cases but leaving the statements of law in place was no correction, and called it a coverup.
• Fletcher v. Experian (5th Cir., 18 February 2026): $2,500. The court wrote that had counsel "accepted responsibility and been more forthcoming, it is likely that the court would have imposed lesser sanctions."
• Lnu v. Blanche (9th Cir., 3 June 2026): $2,500 each and a six-month suspension from practice before the court. The court wrote that if the lawyers had disclosed AI use and apologized in their Motion to Correct, "lesser sanctions may have been warranted."

Different courts weighed different things, and the totals also depend on how many fakes there were. But the orders agree on one point a managing partner can act on: what you say and do after a problem is flagged is part of the record the court reads.

The three things to take from it

Workflow controls. Decide which tools staff may use and put a verification step into the filing checklist. ABA Formal Opinion 512 says managerial lawyers "must establish clear policies regarding the law firm's permissible use of GAI," and that the level of verification "will necessarily depend on the GAI tool and the specific task." It does not give a number, so the process is on you.

Know the local rules. Courts and judges differ. The Fifth Circuit says it considered a circuit-wide AI certification rule and declined to adopt one. In Lnu the Ninth Circuit imposed a two-year certification on the sanctioned firm. Our state-by-state guide tracks the bar opinions and court rules.

A response plan. Decide in advance who answers if a judge or opposing counsel says a citation is fake, who checks it in a primary source, and what the first filing says. Lnu gives the instruction: when a lawyer learns of an error, "immediately alert the court and opposing counsel of the error and disclose its source." The verification checklist covers the checking step.

Limits of reading the database

• It is a harvest, not a census. The page says so.
• It includes warnings. A row can record a court's warning, an order to show cause or an admonishment as well as a sanction.
• It mixes party types. Pro se litigants are the largest US group, and they are in the same table as lawyers.
• Tool and outcome fields are incomplete. Many rows record "Implied" or "Unidentified" for the tool, and outcome fields are free text.
• It is not a rate. Nobody knows how many AI-assisted filings were never challenged.

We did not look at firm size, since the database does not record it.

FAQ

What is the Charlotin hallucination database?

It is a public database of court and tribunal decisions worldwide in which a party relied on AI-hallucinated content, usually fabricated citations. Damien Charlotin maintains it at damiencharlotin.com/hallucinations. It is licensed CC BY 4.0.

How many cases are in it?

On 12 October 2026 the page said it was last updated on 10 October 2026 and counted 2,184 cases identified, 1,492 in the United States. The count grows as decisions are added, so check the live page before citing a number.

Is it only about US cases?

No. The page lists entries from many countries, though the United States has the most, at 1,492.

What counts as an entry?

A decision in which the court addressed AI use in more than a passing reference and found or implied that a party relied on hallucinated content. The page says it also includes some decisions where AI use was alleged but not confirmed.

Is every entry a sanction?

No. Entries include warnings, orders to show cause and admonishments. In our CSV count, 116 of 587 US lawyer entries are flagged as having a professional sanction.

How do I use it as a risk manager?

Use it to see which kinds of filings and responses reach a court order, then pull the order and read it. Do not rely on the database summary alone.

What is the biggest lesson from the orders?

Verify every citation and quotation against a primary source, and answer accurately when a court asks. Several orders, including Fletcher and Lnu, say candor after the problem surfaced affected the penalty.

Does the duty to verify still apply if a legal research tool gave me the citation?

Yes. Under Rule 11 and ABA Formal Opinion 512, the signing lawyer is responsible for the filing. In Farris the Sixth Circuit said the risk of errors persists even with tools sponsored by trusted legal technology providers.

For the US statutes, regulations and court rules these laws sit in, see the US primary law API.

Arshita Anand

Co-Founder & CEO · Attorney

Arshita leads product and strategy at Vaquill AI, the US primary law data API. As an attorney, she decides which law we carry, how each section is sourced and cited, and what legal AI and compliance teams need from the data.