The numbers got worse this fortnight, and worse again in the three days since discovering the database. Damien Charlotin's AI Hallucination Cases database, the most comprehensive public record of fabricated AI material appearing in court filings, logged 1,397 cases as of 6 May 2026. As of 9 May it logs 1,420. Twenty-three new entries in seventy-two hours.
The headline rate that gets quoted in most coverage of this trend, two to three cases per day, is already out of date. On 30 April alone Charlotin logged ten separate decisions involving AI fabrications, on 5 May he logged seven, and the 28–29 April window produced nine. The current pace, on the days when it is moving, is closer to five to ten cases per day than two to three.
Indeed, this fortnight alone has produced a steady flow of named matters that are worth noting, the first of which involving Cherry Hill attorney Raja Rajan. On 27 April, Rajan was sanctioned $5,000 by US District Judge Kai Scott for filing a brief with AI-generated fabrications, his second sanction in the same matter, where he had previously been fined $2,500 for the same conduct. Rajan said he could not remember whether he had used Claude, ChatGPT or Grok to write the memo. He had asked a different chatbot to verify the citations and it had told him they were fine, yet many were fake. Judge Scott noted that any first-year law student knows you have to verify the authority cited, and warned that a third strike would mean referral to Pennsylvania's disciplinary board.
On 5 May the Charlotin database logged Jessica Fuller v Hyde School, a Maine federal class action where Massachusetts-based attorney Karen Guagenty had filed a response to a motion to dismiss containing inaccurate citations and a misquoted statement of Maine's human trafficking law. Guagenty initially called the issues ‘clerical errors’, but four months later she admitted the errors were the output of either ChatGPT or Claude, and that she had failed to review the work before filing. Judge Stacey Neumann's order included the following line: ‘Although AI can be a useful aid in research and drafting, its use does not diminish an attorney's nondelegable duties of diligence, candor, and reasonable inquiry.’
The same day, the Indian Supreme Court was logged as having relied on multiple fabricated case citations in Pooja Ramesh Singh v. J&K Bank. The party using AI in that matter was the judge.
These cases are the current shape of the trend, which has been visible in Australia for nearly a year now. The Victorian solicitor sanctioned in August 2025 was the first Australian practitioner stripped of his ability to practice as a principal lawyer over AI fabrications. The Rishi Nathwani matter in the Supreme Court of Victoria in late 2024, where defence submissions in a murder trial contained fabricated quotes from a parliamentary speech and citations to non-existent Supreme Court judgments, was the case that prompted Justice Elliott to remind the profession that AI output must be ‘independently and thoroughly verified’ before it is filed. The Federal Court has since issued GPN-AI, the NSW Supreme Court has issued Practice Note SC Gen 23, the Queensland Supreme Court has issued judicial guidelines, and other states have issued similar.
Two things are worth thinking about as this trend matures.
Reading the case write-ups in sequence is striking because the lawyers caught are not unsophisticated. Rajan is a practising attorney. Guagenty was lead counsel on a class action. Nathwani is a King's Counsel. The Victorian solicitor was running his own practice. What unites them, rather than being technological ignorance, is the moment at which they decided the verification step was the part of the workflow that could be skipped.
Purpose-built tooling can actually address this part of the workflow. Unlike specialised tools for Australian legal research like Habeas, generic AI tools produce fluent, confident, plausible output and offer little to no native ways to check its own claims against authoritative sources. In a busy practice, under time pressure, with a chatbot that confidently confirms its own citations when asked, it is easy to see how lawyers slip up.
Madison Marcus published a piece on this two days ago that flagged a finding from research in Computers in Human Behavior. The study identified what the authors call a "Reverse Dunning-Kruger Effect" in AI use. Across all skill levels, users overestimated the quality of their own AI-assisted work. More strikingly, users with greater AI literacy showed more overconfidence, not less. The instinct that experienced users will be safer with these tools turns out to be wrong in the data. Familiarity makes verification feel less necessary, not more.
If the problem is the verification gap, the answer is not better lawyers using general-purpose AI. The answer is tooling that closes the gap by design.
Q1 2026 alone produced more than US$145,000 in sanctions tied to AI fabrications, including a record US$110,000 penalty in Oregon and the first indefinite license suspension in US history (Nebraska, attorney Greg Lake, after a brief with 57 of 63 citations flagged as defective). The Sixth Circuit has now dismissed an entire matter for AI-driven misconduct, and Sullivan & Cromwell, one of the largest firms in the world, issued a public apology to a federal bankruptcy judge in April for AI-generated fabrications in their filings.
Australian sanctions to date have been smaller in dollar terms but structurally more severe in some respects, the Victorian solicitor's loss of principal-practice rights being the obvious example. The Australian regulators are watching the international curve closely and the Federal Court is expected to issue further practice notes later this year.
The economic argument for taking this seriously goes as follows: a single hallucination caught by opposing counsel can now mean five-figure sanctions, indemnity costs, a referral to the disciplinary board, and a professional record that follows the practitioner for the rest of their career. The expected cost of an unverified AI citation has gone up by an order of magnitude in less than two years.
Where this leaves the profession
‘Use a legal-specific’ tool is no longer sound advice in and of itself, as Charlotin's data complicates that pitch. The 28 April entries include Tekoma Chaney v. Transdev Services, where a lawyer using LexisNexis Protégé was sanctioned $2,500 for fabricated case law, and Hill v. Workday, where a lawyer using Thomson Reuters' CoCounsel was sanctioned $1,001. On 29 April, a New York lawyer using LexisNexis was sanctioned $2,500 for seven fabricated citations in Jimenez-Fogarty v. Fogarty. Specialist legal AI tools are now showing up in the sanctions data alongside the general-purpose chatbots.
Therefore, that easy pitch (’just use a legal-specific tool and you will be fine’) is empirically falsifiable. A 2024 Stanford RegLab study found that even specialist legal AI tools hallucinate at rates above 17%, with one product hallucinating in more than a third of test queries. The presence of LexisNexis and CoCounsel in the recent sanctions data is the live demonstration of that finding. Switching from ChatGPT to a legal-branded tool does not eliminate the verification duty. It changes the failure rate.
What does change with purpose-built tooling, when the tool is honest about its design, is what the verification step looks like. A tool that grounds every output in a citation linked to a verifiable primary source, and that surfaces those sources alongside the answer, makes the verification step part of the workflow rather than a separate task the lawyer has to remember to do. Habeas does exactly that, with clickable citations encoded into structural outputs.
The next twelve months will be defined by which firms have moved their AI use onto tooling that supports the verification duty and by which lawyers actually use the verification features that purpose-built tools provide. Both halves of that sentence matter. The tool is not enough on its own. Neither is care without infrastructure.
If your firm is still relying on general-purpose chatbots and lawyer memory under deadline pressure, the case rate suggests you are running out of time. If your firm has bought a specialist tool and assumed the problem is solved, the LexisNexis and CoCounsel entries in the past fortnight suggest that assumption is also wrong.
The ultimate advice is this: use tools designed for the task, use the verification features they provide, and treat every citation that goes out under your name as if a court is going to check it. Because increasingly, courts are.
To see what purpose-built Australian legal AI where every citation is observable to the lawyer looks like in practice, book a demo with our team, or if you're a sole practitioner, you can try Habeas at app.habeas.ai.
