Three years of legal AI hallucinations: the record from 2023 to 2026

The first sanction for an AI-fabricated citation was imposed in New York in June 2023. Three years on, SafeLegalAI tracks 66 verified incidents (of 72 recorded, as of 16 July 2026) across twelve countries: the United States, United Kingdom, Ireland, Canada, Australia, New Zealand, India, Singapore, South Africa, Qatar, Brazil, and Israel. The conduct barely changes from case to case. The consequences have escalated sharply.

This is the narrative companion to the incident tracker: what the record shows once every row is checked against the primary judgment.

The pattern set by Mata v Avianca

Mata v Avianca (S.D.N.Y., June 2023) is the template. A lawyer under pressure asked ChatGPT for authority, received confident fabrications, failed to verify them, and — the aggravating fact courts punish hardest — defended the fake cases after they were questioned. Judge Castel’s $5,000 sanction rested on a point every later court has repeated: existing rules already required a reasonable inquiry, so no new “AI rule” was needed to punish the conduct.

Nebraska’s Supreme Court put the same principle plainly in Prososki v. Regan (2026), holding that filing fictitious authority is sanctionable “whether through generative AI or not.” India’s Supreme Court reached it too, in Pooja Ramesh Singh (2026 INSC 668): a decision into which such material enters is, in its words, “no decision in the eyes of the law.”

The consequences got heavier

Money first. Mata’s $5,000 looked severe in 2023. By late 2025 an Oregon federal court had imposed roughly $110,000 in penalties and fees on the plaintiffs’ lawyers in Couvrette v. Wisnovsky after three briefs cited fifteen non-existent cases and the fabrications were then quietly concealed. In England, the High Court ordered £24,727 in indemnity costs against the struck-off solicitor in Bandla v SRA. Singapore’s High Court made personal costs orders against a lawyer and his supervisor in Tan Hai Peng [2026] SGHC 49.

Then careers. Victoria’s legal regulator barred the solicitor in the Dayal matter from practising as a principal and placed him on two years of supervised practice. Nebraska’s Supreme Court referred the attorney in Prososki to its Counsel for Discipline.

Then the criminal register. In Ontario, Ko v. Li escalated from a contempt show-cause to a referral of a criminal-contempt prosecution to the Attorney General, after the lawyer admitted lying about who wrote a ChatGPT-generated factum. New Zealand’s Supreme Court warned in Jones v Family Court at Whangārei [2026] NZSC 1 that reliance on unverified AI citations may, in serious cases, amount to obstruction of justice.

The layer other trackers miss: what regulators did

Twenty-two of the tracked incidents now carry a regulatory outcome, not just a court disposition. Barristers have been referred to the Bar Standards Board; solicitors to the Solicitors Regulation Authority, the Law Society of Ontario, the Legal Practice Board of Western Australia, and, twice in South Africa, the Legal Practice Council. US counsel have been referred to the California and Oregon bars and to the Second Circuit’s own Grievance Panel. India’s Supreme Court went further and directed the Bar Council of India to frame rules for advocates’ AI use.

These dispositions arrive twelve to eighteen months after the incident. Most are still pending, which is precisely why the record matters: each referral is a future ruling.

It is no longer only litigants in person

The early entries were dominated by self-represented parties and sole practitioners. That has changed. In Cork v Smith [2026] EWHC 1199 (Ch), a Pinsent Masons solicitor quoted to the court the text of an insolvency rule that does not exist, and the firm self-referred to the SRA. Sullivan & Cromwell reportedly disclosed AI-hallucinated citations in a Chapter 15 filing in the Southern District of New York — a self-report this tracker still flags as unverified pending the court record.

The courts themselves are not immune. In India, the Bengaluru bench of the Income Tax Appellate Tribunal had to recall a roughly Rs 669 crore ruling that rested on three non-existent citations; the Bombay High Court quashed a faceless tax assessment built on AI-generated precedents; and the Karnataka High Court referred a trial judge who had relied on two Supreme Court decisions that were never delivered.

What the courts have settled, and what they have not

Three points are now common ground across jurisdictions. The verification duty is non-delegable — Alberta’s Court of Appeal held in Reddy v Saroya that the lawyer whose name is on a filing owns its contents even when a contractor drafts it, and Singapore held a supervising solicitor equally accountable. Disclosure regimes are diverging: Ireland’s Court of Appeal set out five principles in Guerin v O’Doherty requiring litigants to disclose AI use, while New South Wales prohibits generative AI for affidavits outright.

What is not settled is enforcement consistency. The same conduct draws a $1,000 charitable donation in Utah, a $10,000 sanction in California, and a criminal-contempt referral in Ontario. That variance, and the regulatory dispositions still to land, are what the next year of this record will document.

Sources

• SafeLegalAI Global Legal AI Incident Tracker — every case above links to its primary judgment
• Anchor cases: Mata v Avianca · Cork v Smith · Pooja Ramesh Singh · Ko v. Li