AI Hallucination Cases: The 1,598-Case Sanctions Tracker

1,598 verified court cases now involve AI-fabricated citations, up from 200 a year ago. Landmark sanctions, the rate curve, and the fix. Updated monthly.

Key facts: AI hallucination cases as of June 2026

• 1,598 AI hallucination cases identified worldwide as of June 9, 2026, per Damien Charlotin's AI Hallucination Cases database.
• The curve: roughly 200 cases in mid-2025, 719 by January 2026, 1,227 by early April 2026, 1,598 by June 9, 2026.
• Between our May 22 audit (1,458 cases) and June 9 (1,598), the database added 140 cases. That is just under 8 per day, up from the 5 to 6 per day reporters measured in April.
• Record penalty: about $109,700 in combined sanctions and fees in Couvrette v. Wisnovsky (D. Oregon), per the ABA Journal (2026).
• US courts imposed over $145,000 in AI-filing penalties in Q1 2026 alone, per ComplexDiscovery (2026).
• New severity ceiling: in Withers v. City of Aberdeen (N.D. Mississippi, June 8, 2026), both sides filed fake citations, the judge canceled the trial and suspended the two lead attorneys from the district for two years.
• Even paid legal research tools hallucinate 17% to 34% of the time, per Stanford RegLab's peer-reviewed study.

The count: 1,598 AI hallucination cases and climbing

In June 2023, Mata v. Avianca made two New York lawyers famous for filing six cases ChatGPT invented. The sanction was $5,000, and the working assumption was that the profession would learn from one loud, embarrassing example. It did not. Three years on, the same failure has its own public database, and the database needs daily updates.

That database is maintained by Damien Charlotin, a research fellow at HEC Paris. Its bar for inclusion is strict: it only lists decisions where a court or tribunal explicitly found, or clearly implied, that a party relied on hallucinated material. Mere accusations do not count. As of its June 9, 2026 update, it had identified 1,598 cases.

We did not take that number on faith. When we first audited this page on May 22, 2026, the count stood at 1,458. Eighteen days later: 1,598. That is 140 new documented cases, just under eight per day, faster than the five-to-six daily pace reported in April 2026. The curve is not flattening. It is steepening.

And these are only the cases a judge caught and wrote up. A fabricated citation that opposing counsel never checks does not make the database. Whatever the true number is, 1,598 is the floor.

The growth curve: 200 to 1,598 in about a year

The May and June figures are HAQQ's own checks against the live database. The implied rate is now just under 8 new cases per day.

Reporting by Scientific American (2026) frames the same curve in one sentence: the database held about 200 cases a year ago and 719 in January. An April 2026 analysis by PlatinumIDS put the count at 1,227, with 811 of those in US courts. Our own two checks extend the line to 1,458 and then 1,598. Every measurement points the same direction.

Lawyers sanctioned for AI fake citations: 8 landmark cases

Eight cases define how courts now treat AI-fabricated citations. Each one below was verified against primary reporting before publication. Together they show the penalty arc: a fine in 2023, fee awards and revoked admissions in 2025, then suspensions and a canceled trial by mid-2026.

Mata v. Avianca: the original (2023)

Two New York lawyers cited six nonexistent decisions generated by ChatGPT in a personal injury suit against Avianca Airlines. The Southern District of New York fined them $5,000. At the time it read as a freak accident. In hindsight it was case one of 1,598.

Wadsworth v. Walmart: the firm's own AI tool failed (2025)

Three Morgan & Morgan attorneys filed motions citing nine cases. Eight did not exist. The citations came from the firm's in-house AI platform, MX2.law, not a consumer chatbot. A Wyoming federal judge revoked the drafting attorney's pro hac vice admission and fined him $3,000, with $1,000 fines for the two signing attorneys, per LawSites (2025). The lesson: 'legal-grade' branding on the tool does not transfer the verification duty away from the lawyer.

Ayinde v Haringey: the UK draws the line (2025)

Freely available generative artificial intelligence tools, trained on a large language model such as ChatGPT are not capable of conducting reliable legal research.

Five fake cases in a homelessness judicial review put a junior barrister before the Divisional Court. The court found the threshold for contempt of court was met, declined (this once) to initiate proceedings, and referred the matter to the Bar Standards Board, according to Burges Salmon's case analysis (2025). The judgment also defined what counts as an authoritative source, which we covered in depth in our When AI Lies to the Court report.

Coomer v. Lindell: 30 bad citations in one brief (2025)

Lawyers for MyPillow CEO Mike Lindell filed a brief in his Colorado defamation case containing nearly 30 defective or fabricated citations, drafted with help from generative AI tools. Judge Nina Y. Wang fined both attorneys $3,000 each, as reported by NPR (2025).

Couvrette v. Wisnovsky: the record penalty (2025-2026)

Across three summary judgment briefs in an Oregon winery dispute, counsel filed 15 AI-generated fake citations and eight fabricated quotations. The result: roughly $109,700 in combined sanctions, fines, and opposing fees across the two lawyers involved, believed to be the largest aggregate penalty yet, per the ABA Journal (2026). The magistrate judge called the matter 'a notorious outlier in both degree and volume.'

Tan Hai Peng v Tan Cheong Joo: the supervisor pays too (2026)

Singapore's High Court imposed personal costs orders on both the junior lawyer who generated hallucinated content and the partner who signed the filing without checking it. The partner's workload defense was rejected. We documented this decision, [2026] SGHC 49, in our primary-source report on AI hallucinations in law; it pairs with a UK Upper Tribunal ruling to form a new rule: failing to check a junior's AI output makes the supervisor more culpable, not less.

Whiting v. City of Athens: appellate courts lose patience (2026)

The Sixth Circuit found counsel had cited over 24 fake citations, misquoted real cases, and misrepresented the record. It ordered two attorneys to pay $15,000 each in punitive sanctions on top of the opposing side's attorney fees and double costs, the stiffest penalty available under Rule 38, per the Sixth Circuit Appellate Blog (2026). Aggravating factor: prior discipline for lack of candor.

Withers v. City of Aberdeen: both sides, canceled trial (June 2026)

The newest landmark, decided June 8, 2026. Filings from both the plaintiff's and the city's lawyers in a Mississippi fee dispute contained nonexistent cases. Judge Sharion Aycock canceled the scheduled trial, suspended the two lead out-of-state attorneys from practicing in the Northern District of Mississippi for two years, revoked pro hac vice admissions, fined every lawyer of record between $1,000 and $3,500, and notified state bar authorities, per Bloomberg Law (2026). Sanctions have moved from money to careers.

The penalty trend: from fines to career damage

Our When AI Lies to the Court report tracked the early escalation: $5,000 in 2023 to $55,597 in a single matter by 2025, an 11x jump in 18 months. The record now sits near $109,700, and US courts imposed over $145,000 in AI-filing penalties in the first quarter of 2026 alone, per ComplexDiscovery (2026).

But the money is no longer the story. The 2026 cases add suspensions, revoked admissions, regulator referrals, a canceled trial, and personal liability for supervising partners. A fine is a bad day. A two-year suspension from a federal district is a different category of consequence.

Why lawyers keep filing fake AI citations

The judges have warned. The sanctions are public. The database is free. So why does the count grow by eight cases a day? Scientific American (2026) put the question in its headline: lawyers know AI hallucinates, and they keep citing fake cases anyway.

Part of the answer is base rates. Stanford RegLab's peer-reviewed study found that even purpose-built paid legal research tools hallucinate 17% to 34% of the time, and general chatbots far more. Our own 300-task benchmark of 10 frontier models found 24% of all answers cited law that did not actually support the proposition. A tool that is wrong one time in four, used by thousands of time-pressed lawyers daily, produces this database on schedule.

Part of it is misplaced delegation. Wadsworth shows an in-house 'legal AI' platform fabricating eight of nine citations. The UK and Singapore supervision rulings show seniors signing junior work unchecked. As we argued in When AI Lies to the Court, the verification duty is absolute and non-delegable: not to the tool, not to the junior, not to the vendor. Courts across the UK, Singapore, Argentina, and the US converged on that principle independently.

How HAQQ grounds citations

Our position has not changed since the first version of this audit: do not ask a model to remember the law. Ask it to retrieve the law, cite the source, and surface its own uncertainty, then put a human in the loop before anything leaves the building. Every output in HAQQ carries a citation back to a primary source, and material work routes to attorney sign-off.

That architecture exists because of pages like this one. A confident fake citation is worse than an honest 'I do not know.' The 1,598 cases above are what happens at scale when systems are built the other way around.

How this tracker stays current

We re-verify this page monthly against the live database and primary reporting. Each update checks four things: the total case count, the record penalty, new landmark decisions (appellate, cross-border, or novel sanction types), and the implied daily rate. Numbers we cannot re-verify get removed, not kept.

• June 11, 2026: rebuilt the one-off audit as a tracker. Count updated 1,458 to 1,598. Added Couvrette, Whiting, Withers, Wadsworth, Coomer, Ayinde, and Tan Hai Peng as landmark entries.
• May 22, 2026: original audit published at 1,458 cases.

FAQ

How many AI hallucination cases are there in 2026?

1,598 documented cases worldwide as of June 9, 2026, per Damien Charlotin's public AI Hallucination Cases database. The count covers only cases where a court explicitly found or implied reliance on hallucinated material, so the true number is higher.

What is the largest sanction for AI fake citations?

About $109,700 in combined sanctions, fines, and opposing fees in Couvrette v. Wisnovsky in the District of Oregon, where counsel filed 15 fake citations and eight fabricated quotes across three briefs. It is believed to be the largest aggregate penalty tied to one matter so far.

Can a lawyer be suspended for citing fake AI cases?

Yes. In Withers v. City of Aberdeen (June 2026), a federal judge in Mississippi suspended two lead attorneys from practicing in the district for two years after both sides filed AI-hallucinated citations, and canceled the trial. Courts also revoke pro hac vice admissions and refer lawyers to bar regulators.

How often does legal AI hallucinate?

Stanford RegLab's peer-reviewed study found purpose-built legal research tools hallucinate 17% to 34% of the time, and general-purpose chatbots far more often on legal questions. Our own 300-task benchmark found 24% of frontier-model answers cited law that did not support the claim.

Do paid legal AI tools also produce fake citations?

Yes. In Wadsworth v. Walmart, eight of nine citations in a filing came from a law firm's own in-house AI platform and did not exist. The Stanford data shows paid research tools hallucinate less than chatbots, but nowhere near zero.

What was the first lawyer sanctioned for ChatGPT fake cases?

Mata v. Avianca in the Southern District of New York, June 2023. Two lawyers were fined $5,000 for citing six nonexistent cases generated by ChatGPT. It remains the reference point for every decision since.

How do lawyers avoid sanctions when using AI?

Verify every citation against a primary source before filing, treat the duty as non-delegable, and prefer tools that ground answers in retrieved sources with verifiable citations rather than generating from memory. Courts in the UK and Singapore have also made supervisors personally liable for unchecked junior AI work, so the check has to happen at signature, not just at drafting.

How often is this tracker updated?

Monthly. We re-verify the case count against the live database, re-check the record penalty, and add new landmark decisions. The changelog above lists each revision.

Key takeaways

• 1,598 documented AI hallucination cases as of June 9, 2026, adding just under 8 per day and accelerating.
• Penalties moved from a $5,000 fine in 2023 to a ~$109,700 record, two-year suspensions, and a canceled trial by mid-2026.
• Supervisors are now personally liable in the UK and Singapore for signing unchecked AI work.
• Even paid legal tools hallucinate 17% to 34% of the time. The fix is architectural: grounded retrieval, verifiable citations, human sign-off.
• This page is re-verified monthly. Last verified June 9, 2026.

Sources & further reading

• When AI Lies to the Court: our primary-source report on AI hallucinations in law
• Nippon Life v. OpenAI: when AI plays lawyer
• Our benchmark: 3 models on 100 real legal questions
• Our 300-task frontier model benchmark (the 24% citation gap)
• Damien Charlotin: AI Hallucination Cases database
• ABA Journal: Oregon federal judge hands down $110K penalty for AI errors
• Bloomberg Law: lawyers on both sides in Mississippi case punished for AI errors
• Sixth Circuit Appellate Blog: sanctions for fake citations in Whiting v. City of Athens
• LawSites: Morgan & Morgan attorneys sanctioned in Wadsworth v. Walmart
• NPR: AI hallucination in Mike Lindell case serves as a stark warning
• Burges Salmon: professional conduct and AI in Ayinde v Haringey
• Scientific American: why lawyers keep citing fake cases invented by AI
• ComplexDiscovery: the AI sanction wave, $145K in Q1 penalties
• PlatinumIDS: inside the AI hallucination crisis hitting courts
• Stanford HAI / RegLab: legal models hallucinate 1 out of 6 queries or more
• Mata v. Avianca, Inc. (Wikipedia)