Over 1,000 Hallucination Cases and Counting: What Sanctioned Lawyers Reveal About the Competence Gap

Liga leads the legal analysis. Alex frames the strategic implications.

The number is 1,002 and climbing. Damien Charlotin's global database now tracks over a thousand documented instances of AI-generated hallucinated content in court proceedings. 712 judicial decisions worldwide explicitly address these fabrications, 90 percent of them issued in 2025 alone. The rate is roughly four to five new incidents per day. The sanctions are escalating -- from $5,000 fines in 2023 to an $86,000 order in 2025 to regulatory referrals that can end careers entirely.

And the pattern is always, unfailingly, the same.

A lawyer asks an AI tool for case law. The tool produces something plausible -- correct formatting, realistic citation structure, convincing legal reasoning. The lawyer includes it in a filing without checking it against a legal database. Opposing counsel or the court discovers the cases do not exist. Sanctions follow.

Every sanctioned lawyer knew they were supposed to verify their sources. Not one of them claimed to have intentionally fabricated anything. They all believed the AI was right. And this is the part that matters: rules telling lawyers to verify their work have existed for centuries. Adding "including AI-generated work" to those rules has not changed the outcome. The problem is not that professionals lack instructions. The problem is that they lack the skill to recognise what a hallucinated citation looks like, the understanding of why AI produces them, and the practised workflow to catch them before they reach a courtroom.

This is a competence problem. Not a compliance problem. And it is exactly the gap that Article 4 of the EU AI Act was written to address.

The Failure Mode That Never Changes

Liga's analysis of the major cases reveals a failure mode so consistent it could be a diagnostic checklist. Every case follows the same sequence:

Step 1: The query. A lawyer -- often under time pressure, sometimes unfamiliar with the technology -- asks a generative AI tool for case law supporting a legal proposition. The prompt is typically broad: "find me cases that support X."

Step 2: The plausible output. The AI generates citations that look correct. They follow proper formatting conventions. They reference real courts. The legal reasoning sounds authoritative. In some cases, the AI even fabricates procedural histories and quotes from judges.

Step 3: The missing verification. The lawyer reads the output, finds it persuasive, and includes it in a filing. At no point does anyone check whether the cited cases exist in Westlaw, LexisNexis, CanLII, BAILII, or any other primary legal database.

Step 4: The discovery. Opposing counsel, a clerk, or the judge attempts to look up the citations and finds nothing. The fabrication unravels.

This sequence has now played out over a thousand times across at least six countries and dozens of jurisdictions. The scale of the repetition should tell us something important: this is not a problem of rogue individuals. It is a systemic skills gap dressed in the clothing of individual negligence.

The Cases That Defined the Problem

Mata v. Avianca -- Where It Started

In June 2023, Judge P. Kevin Castel of the Southern District of New York fined two lawyers and their firm $5,000 for submitting a motion built on fabricated ChatGPT-generated case citations. The case became global news not because the sanction was large -- it was modest -- but because of what attorney Steven Schwartz said in his testimony: he had been "operating under the false perception that ChatGPT could not possibly be fabricating cases on its own."

That sentence captures the core of the competence gap. Schwartz was not lazy. He was not unethical. He fundamentally did not understand what the tool he was using actually does. He believed he was accessing a legal database. He was generating statistically probable text.

This distinction -- between retrieval and generation -- is the single most important thing any professional using AI needs to understand. And most training programmes still do not teach it.

Morgan & Morgan -- When Big Firms Proved Immune to Size

When Morgan & Morgan -- the 42nd largest law firm in the United States by headcount -- was sanctioned in February 2025, the "this only happens to solo practitioners" narrative collapsed. Attorney Rudwin Ayala used the firm's own in-house AI platform, MX2.law, to generate case law for motions in limine. His prompts were strikingly naive: "add more case law regarding motions in limine." Eight non-existent cases went into the filings.

The sanctions were telling. Judge Kelly H. Rankin fined Ayala $3,000 and revoked his pro hac vice admission. But the supervising attorney, T. Michael Morgan, and local counsel Taly Goody were each fined $1,000 -- despite never having seen the filing before it was submitted. Their signatures were on the document. In the court's view, that was enough.

This case established a critical principle: supervisory liability attaches to the signature, not to involvement in drafting. If your name is on a filing, the verification obligation is yours.

Al-Haroun v. Qatar National Bank -- The Worst Case on Record

The Al-Haroun case, decided by the UK High Court in June 2025, remains the most alarming documented incident. In this GBP 89.4 million damages claim, the claimant's solicitor submitted 45 case-law authorities. Eighteen of them -- 40 percent -- were entirely fabricated. No other sanctioned case comes close to this fabrication rate.

What makes Al-Haroun uniquely instructive is the chain of failure. The client used AI tools to generate the research. The solicitor, Abid Hussain of Primus Solicitors, admitted relying on his client's work without independent verification. The hallucination problem extended beyond direct AI use by the lawyer -- the solicitor failed to verify authorities regardless of their source.

Dame Victoria Sharp, President of the King's Bench Division, heard this case alongside the companion Ayinde v. Haringey case. The Bar Council called the judgment "a wake-up call to the profession." Both practitioners were referred to their professional regulators -- the SRA for Hussain, the BSB for barrister Sarah Forey in the Ayinde matter. The court made clear that regulatory referral would "usually be appropriate" for lawyers who submit unverified citations.

ByoPlanet v. Johansson -- When the Ceiling Broke

In July 2025, a federal judge in the Southern District of Florida imposed $86,000 in sanctions in ByoPlanet International, LLC v. Johansson -- the largest AI hallucination penalty to date and a dramatic escalation from the $1,000-to-$5,000 range that had become the norm.

The circumstances made it inescapable. Plaintiffs' counsel had been put on notice in April 2025 that his AI-generated filings contained hallucinated authorities. He continued filing documents with fabricated citations in seven subsequent submissions across eight related cases. All four matters were dismissed. The attorney was ordered to attach a copy of the sanctions order to every case he files in the Southern District of Florida for the next two years. The judge referred him to the Florida Bar for discipline.

This was not a mistake. It was a pattern -- and the court priced it accordingly.

Gordon Rees -- The Repeat Offender

If any case proves that awareness without competence solves nothing, it is Gordon Rees Scully Mansukhani. This Am Law 100 firm has been publicly implicated in AI hallucination incidents three times in under a year.

The first incident, in a bankruptcy case in Alabama in summer 2025, resulted in the firm paying $20,494 in opposing counsel's fees. Gordon Rees described itself as "profoundly embarrassed" and told the court it had "expended significant financial and human resources" to remediate the problem and prevent future violations.

A second reprimand followed in December 2025. A third accusation landed in February 2026.

Gordon Rees knew about the problem. Gordon Rees spent money on the problem. Gordon Rees kept having the problem. This is what happens when firms treat AI hallucination as a compliance issue -- write a policy, send a memo, move on -- rather than building genuine verification competence into daily workflows.

The US-Europe Divergence: Two Systems, Same Gap

The enforcement approaches on either side of the Atlantic are diverging in ways that matter strategically.

The American Approach: Fines as Cost of Business

US courts sanction under Federal Rule of Civil Procedure 11 and inherent authority. The typical range is $1,000 to $5,000 per lawyer, per incident. Even the ByoPlanet outlier at $86,000 is, for a law firm, an operational expense. Over 200 federal judges have issued standing orders requiring AI disclosure in filings. Some courts have begun requiring lawyers to certify that cited authorities are real.

The American model treats hallucination as an individual disciplinary matter. Pay the fine, attend the CLE, move on. The systemic problem persists because the systemic response is absent.

The European Approach: Career Consequences and Regulatory Architecture

The UK High Court's approach in Al-Haroun and Ayinde was categorically different. No fines were imposed. Instead, both practitioners were referred to their professional regulators. The SRA and BSB can impose outcomes ranging from formal warnings to conditions on practice to suspension or striking off. A regulatory referral is not a cheque you write -- it is a proceeding that can follow a lawyer for the rest of their career.

In Germany, the Regional Court of Darmstadt took yet another approach in November 2025, ruling that a court-appointed expert who relied extensively on AI without disclosing it could have their fee set to zero euros. The message: undisclosed AI use in professional work is not a minor procedural lapse. It goes to the legitimacy of the work product itself.

And overarching all of this, Article 4 of the EU AI Act -- effective since 2 February 2025 -- now requires that organisations deploying AI systems ensure their staff have "sufficient AI literacy." This is a binding obligation across all 27 EU member states. It does not tell lawyers to verify their citations. It tells their employers to ensure they have the competence to do so. The failure is no longer the individual lawyer's alone -- it is the organisation's.

The combined European enforcement exposure -- GDPR penalties up to EUR 20 million or 4% of global turnover, AI Act penalties up to EUR 35 million or 7% of turnover, plus professional regulatory action -- creates a fundamentally different risk environment than a $5,000 court fine.

Jurisdiction	Enforcement Mechanism	Typical Consequence	What It Means
United States	Court sanctions (FRCP Rule 11)	$1,000-$86,000 fine	Operational cost; absorbed and forgotten
United Kingdom	Regulatory referral (SRA/BSB)	Disciplinary proceedings; potential practice restrictions	Career-level consequences; systemic deterrent
Canada	Cost orders + certification rules	Personal cost liability; signed verification	Growing procedural safeguards
Germany	Court rulings on expert legitimacy	Fee set to zero; work product challenged	Professional credibility at stake
EU (27 states)	AI Act + GDPR enforcement	EUR 7.5M-35M or 1-7% turnover	Organisational liability, not just individual

The Competence Analysis: Why Rules Alone Have Not Worked

Here is the uncomfortable truth the case data reveals.

Every professional ethics code in every jurisdiction already requires lawyers to verify the accuracy of citations submitted to courts. This has been true for decades. The obligation did not appear with ChatGPT. It has been black-letter professional conduct law since long before any of these lawyers were admitted to practice.

Every sanctioned lawyer in these cases was bound by those rules. They all knew them. Not a single one claimed ignorance of the duty to verify sources. And they all submitted fabricated citations anyway.

The Ontario lawyer in Ko v. Li said she was "shocked" when the citations could not be found. The New York lawyer in Mata said he believed "ChatGPT could not possibly be fabricating cases." The UK solicitor in Al-Haroun trusted his client's research without verification. The Morgan & Morgan associate prompted his firm's own AI tool to "add more case law" and filed the output raw.

None of these failures was a rules failure. Every one was a skills failure.

These lawyers did not lack a prohibition against submitting fabricated cases. They lacked three specific competencies:

1. Understanding what generative AI actually does. They treated AI as a research database -- a tool that retrieves existing information. It is not. It generates statistically probable text. The distinction is fundamental, and most training programmes gloss over it in a single slide.

2. Recognising the signature of a hallucinated citation. Fabricated AI citations have tells -- subtle but detectable patterns in formatting, case numbers, court designations, and procedural history. A lawyer trained to spot these patterns catches them. A lawyer who has never seen a hallucinated citation before, and has never practised identifying one, does not.

3. Embedding verification as an automatic workflow step. Knowing you should verify is not the same as having verification built into your process so deeply that skipping it feels wrong. Competence is not knowledge of a rule. It is the practised habit of following it.

This is why Article 4 of the EU AI Act exists. Not because the existing rules were unclear. Because telling people to be competent does not make them competent. Article 4 places the obligation on the organisation to ensure AI literacy -- not to write a policy about it, but to build the actual capability. The difference between those two things is the difference between the firms that keep appearing in sanction orders and those that do not.

What Actually Prevents This

The case data points clearly to what works and what does not.

What does not work:

Policies stating that lawyers must verify AI-generated citations. (Every sanctioned firm had one or could point to a professional rule requiring it.)
One-off awareness sessions explaining that AI can hallucinate. (Awareness without practice produces the gap seen in every case.)
AI disclosure requirements and certification forms. (Ontario's Rule 4.06.1 mandates signed certification of citation authenticity. Connecticut's Supreme Court is now dealing with hallucinated citations anyway.)

What does work:

Structured verification training that gives practitioners hands-on experience identifying hallucinated citations -- not a lecture about the risk, but supervised practice catching fabricated outputs before they would reach a filing.
Workflow integration that makes verification a defined step in the document production process, with clear responsibility assignments and documentation.
Understanding AI mechanics at a sufficient depth to know why hallucinations occur, when they are most likely, and what they look like -- so that the instinct to verify becomes informed rather than procedural.
Organisational accountability that treats AI competence as a training obligation, not a memo obligation. Gordon Rees spent "significant financial and human resources" on remediation after its first incident. Two more followed. Money spent on compliance theatre does not build competence.

This is the Twin Ladder thesis in practice. Level 0 -- AI literacy -- means understanding what generative AI does and does not do, and having the practised skill to verify its outputs. Not knowing the rule. Having the skill. Every sanctioned lawyer in this analysis would have been helped by genuine Level 0 training. None of them would have been helped by another policy document.

Key Takeaways

Over 1,000 documented incidents of AI-generated hallucinated content in court filings have been catalogued globally, with the rate accelerating to roughly four to five new cases per day.
The failure mode is identical in every case: lawyer queries AI, receives plausible-looking citation, files without verification, faces sanctions when fabrication is discovered. No sanctioned lawyer claimed to have intentionally submitted false citations.
Sanctions are escalating rapidly. From $5,000 in Mata (2023) to $86,000 in ByoPlanet (2025), with UK regulatory referrals that threaten practice rights entirely.
Repeat offenders prove that awareness is not competence. Gordon Rees has been implicated three times in under a year despite investing in remediation after the first incident.
The US-Europe enforcement divergence is widening. American monetary sanctions remain absorbable business costs. European regulatory referrals, combined with AI Act and GDPR penalties, create career-level and organisation-level consequences.
Every sanctioned lawyer knew the rules. None had the skills. The competence gap is not about what professionals are told to do. It is about whether they have been trained to actually do it.
Article 4 of the EU AI Act shifts the obligation to organisations. Effective since February 2025, it requires employers to ensure staff have "sufficient AI literacy" -- making the skills gap an organisational compliance failure, not just an individual ethical one.
Verification is a trainable skill, not a policy outcome. The cases consistently show that policies, memos, and standing orders do not prevent hallucination incidents. Structured training that builds practised verification capability does.

The global AI hallucination case database is maintained by Damien Charlotin and updated continuously. For the Twin Ladder approach to building verification competence, see our methodology.

MI halucināciju sankcijas: 14 lietas un to mācības