AI hallucinations in court filings have become one of the most visible professional risks of the generative AI era, and three years of fines, referrals and public rebukes have not stopped them. On 17 September 2026 Reuters reported that AI error-ridden court filings are surging despite three years of sanctions, and courts well beyond the United States are now responding with personal costs orders, practising restrictions and, in England, a first strike-off. For partners, general counsel and in-house litigation leads who instruct lawyers across borders, the question is no longer whether a fabricated authority can reach a court file, but what the sanction looks like in each forum and which controls stop it happening in the first place.
What are AI hallucinations in court filings?
An AI hallucination in a court filing is a case, quotation, statute or legal proposition that a generative AI tool has produced with confidence but which does not exist, or does not say what the filing claims it says. The error is not a typo: it is invented material presented to a court as law.
Hallucinated case citations usually take one of three forms. The first is a wholly fictitious authority with a plausible name and neutral citation. The second is a real case given a false citation, date or court. The third, and the hardest to spot, is a real judgment attributed a holding or a quotation that it never contained. India's Supreme Court described all three in a September 2026 ruling, noting that some of the authorities before it were non-existent, some bore fake citations and some were genuine judgments carrying ratios the real decisions did not support.
The Divisional Court in England put the underlying problem bluntly in June 2025. In Ayinde v Haringey and Al-Haroun v Qatar National Bank, it said that freely available generative AI tools such as ChatGPT "are not capable of conducting reliable legal research". They generate text that looks like law; they do not retrieve law.
The scale is now measurable. The AI Hallucination Cases Database maintained by researcher Damien Charlotin, which records decisions where a court or tribunal found that a party relied on hallucinated content, listed 2,077 cases when it was last updated on 23 September 2026. The United States accounts for 1,428 of them, followed by Canada (218), Australia (112), the UK (69), Israel (57) and Brazil (41), with India on 15. The database counts decisions, not every flawed filing, so the true number of documents affected is higher.
Why AI hallucinations in court filings keep rising
The short answer is that adoption of generative AI in legal work has outpaced verification discipline, and sanctions so far have been too scattered to change behaviour across the profession.
The Reuters report by Mike Scarcella catalogued fresh incidents from New Mexico, Oklahoma and Los Angeles in September 2026 alone. It also quoted US District Judge P. Kevin Castel of Manhattan, who issued one of the earliest sanctions in 2023, remarking that a profession already aware that hallucinations happen might have been expected to be more vigilant, according to a Reuters Legal summary of the story.
The Illinois courts have tracked the same trend. A January 2026 article on the Illinois Courts website said that more than 280 court filings had included hallucinated citations since 2023, and that the number of such cases surged sevenfold in 2025, citing a Bloomberg Law analysis.
Three structural reasons explain why lawyers AI hallucinations keep recurring:
- Fluency hides fabrication. A hallucinated authority reads exactly like a real one, complete with party names, a year and a court. Nothing in the output flags it as invented.
- Delegation without review. Many incidents involve a junior lawyer, a paralegal, a former colleague or even the client producing research that the signing lawyer did not independently check.
- Time pressure. Filing deadlines reward speed, and a tool that produces a finished-looking brief in minutes is tempting when the court clock is running.
Regulators are responding unevenly, which is part of a wider contest over who sets the rules for AI. The result, for now, is a patchwork of court-by-court sanctions rather than a single standard.
United States: from Mata v Avianca to five-figure penalties
US courts have moved from modest fines to five-figure penalties per lawyer, fee-shifting and double costs, and the sanction is now routinely imposed on each signatory rather than on the firm alone.
The case that started the pattern was Mata v Avianca in New York in 2023, where a lawyer filed six fictitious cases. When questioned, he explained that he had asked the AI tool whether the cases were real and it had said they were. Judge Castel imposed a US$5,000 fine, as recounted by the Illinois Courts. That answers the recurring question of what happened to the lawyer who used ChatGPT: he was sanctioned, and his case became the profession's cautionary tale.
The penalties have since escalated. In Whiting v City of Athens in March 2026, the US Court of Appeals for the Sixth Circuit found more than two dozen fake or incorrect citations and ordered each of two lawyers to pay US$15,000 in punitive sanctions under appellate Rule 38, plus the appellees' attorney fees and double costs, according to the Sixth Circuit Appellate Blog. Notably, the court could not definitively say whether AI had been used, because counsel did not seriously respond to its show cause order, and it rejected the argument that asking about AI use breached privilege. The sanction attached to the fabricated citations themselves.
Other 2026 decisions show the same direction of travel. EDRM reported that US courts imposed at least US$145,000 in sanctions for AI-generated fake citations in the first quarter of 2026. The same report cited US$15,500 in Couvrette v Wisnovsky in the District of Oregon, where filings contained 15 fake citations and eight fabricated quotations, and a US$10,000 sanction against a lawyer in the Oregon Court of Appeals in March 2026. In February 2026, Reuters reported that a US appeals court ordered a lawyer to pay US$2,500 over AI hallucinations in a brief.
The ethical baseline is set by the American Bar Association's Formal Opinion 512, issued on 29 July 2024. It addresses competence, confidentiality, communication, meritorious claims and candour to the tribunal, supervisory responsibilities and fees, and warns against uncritical reliance on AI output, as summarised in The Bar Examiner.
England and Wales: Ayinde, the SRA and a first strike-off
England and Wales has moved furthest on the regulatory side: a Divisional Court warning in 2025, a formal SRA warning notice in August 2026 and, in September 2026, the first strike-off by the Solicitors Disciplinary Tribunal in an AI case.
The Ayinde judgment
On 6 June 2025, Dame Victoria Sharp P and Johnson J delivered a joint judgment in two cases heard together. In Ayinde, a housing judicial review, a barrister and Haringey Law Centre had cited five non-existent cases. In Al-Haroun, a damages claim against Qatar National Bank, 18 of 45 citations did not exist, and many others did not support the propositions for which they were cited; the solicitor admitted relying on research his client had carried out using public AI tools, without verifying it. The court did not start contempt proceedings, but referred the lawyers involved to the Bar Standards Board and the Solicitors Regulation Authority, and sent the judgment to the Bar Council, the Law Society and the Council of the Inns of Court, according to The Expert Record. English courts have continued to flag the problem since, including in the warning that followed a Pinsent Masons filing error earlier this year.
The SRA warning notice
On 17 August 2026 the SRA published a warning notice on the misuse of AI. It expects solicitors to confirm that every named authority is genuine, relevant and has a verifiable citation, states that reliance on an AI output would not be a suitable defence, and makes clear that supervisors remain responsible for the work they oversee. Firms must maintain effective governance, systems and controls, and must not put client data into public tools without appropriate safeguards. The notice sits alongside the regulator's wider push on standards, including its consultation on tougher complaints-handling rules.
The first SDT strike-off
In a decision dated 3 September 2026, the Solicitors Disciplinary Tribunal struck Abhishek Kumar off the register of Registered Foreign Lawyers. The Tribunal said it was the first time a lawyer's use of AI in legal proceedings had been litigated before it. It found proved allegations of citing non-existent cases and incorrect legal propositions in documents he filed in proceedings in March and April 2026; he admitted using AI tools but denied intending to mislead. A separate allegation concerned employing a person disqualified by immigration status. The route to that outcome, an SRA prosecution before the Tribunal, is the same one used in other conduct cases such as the Post Office Horizon-related referrals.
Australia: practising restrictions and a murder trial delayed
Australian regulators have shown they will restrict a lawyer's right to practise, not just criticise the filing.
On 19 August 2025 the Victorian Legal Services Board varied the practising certificate of a solicitor referred to as Mr Dayal, after he tendered AI-generated authorities with inaccurate citations and summaries to the Federal Circuit and Family Court. According to the Board's statement of 2 September 2025, he may no longer practise as a principal lawyer, handle trust money or run his own practice. He must work as an employee solicitor under supervision for two years, with quarterly reports from both him and his supervisor.
The same month, senior counsel in a Supreme Court of Victoria murder case apologised after submissions were found to contain fabricated quotations from a speech to the state legislature and non-existent case citations. The errors came to light when the judge's associates could not find the cases, and they caused a 24-hour delay. Justice James Elliott said AI use was not acceptable "unless the product of that use is independently and thoroughly verified", CBS News reported.
For clients with Australian litigation, these examples show that the consequences reach the lawyer's licence and the timetable of the case. Instructing counsel through advisers in Australia who can evidence their AI controls is a sensible starting point.
Canada and Singapore: lawyers paying costs personally
Canadian and Singaporean courts have both used personal costs orders, making the individual lawyer, not the client, pay for the waste caused by fake citations.
In Zhang v Chen, 2024 BCSC 285, a British Columbia family case, counsel cited two non-existent cases that ChatGPT had suggested. Opposing counsel's team spent time establishing that the cases did not exist. The court held the lawyer personally liable for the additional expense the fake cases caused and directed her to review her files for other AI-generated material and report back within 30 days, observing that AI is "still no substitute for the professional expertise" required of lawyers, according to CanLII Connects. With Canada now second only to the US in the Charlotin database, the province is also seeing AI accountability questions reach the courts in other forms, including British Columbia's claim against OpenAI.
Singapore followed with a clear statement of principle. In Tajudin bin Gulam Rasul v Suriaya bte Haja Mohideen [2025] SGHCR 33, decided on 29 September 2025, Assistant Registrar Tan Yu Qing ordered the claimants' counsel to pay S$800 in personal costs to the defendant within 14 days for citing a fictitious AI-generated authority. The court deliberately withheld the fake case name so it would not spread. It held that lawyers owe a professional, non-delegable duty to ensure that everything placed before the court exists and is accurate, and it relied on the courts' Guide on the Use of Generative Artificial Intelligence Tools by Court Users. Teams running disputes in the city state can find advisers in Singapore through our directory.
India: the Supreme Court calls fake citations misconduct
India's Supreme Court has taken the hardest doctrinal line: citing unverified AI-generated authorities is professional misconduct, and a decision that rests on them is no decision at all.
The court had voiced concern in March 2026 about the growing "menace" of AI-generated non-existent judgments. On 2 July 2026, in Pooja Ramesh Singh v Jammu and Kashmir Bank Ltd, an insolvency matter concerning Essel Infraprojects, Justices P.S. Narasimha and Alok Aradhe held that it is misconduct for an advocate to cite such judgments without verification and a serious lapse for a judge to rely on them. Orders tainted by fabricated precedents were set aside as "no decision in the eyes of the law", and the court directed the Bar Council of India to form a committee to prescribe guiding principles and disciplinary consequences, LiveLaw reported.
Two months later the stakes became commercial. On 2 September 2026, in Vijay Ghanshyam Gadiya v Union of India, Justices Dipankar Datta and Sheel Nagu set aside a customs penalty of ₹425.27 crore imposed by the Additional Commissioner of Customs, Surat, under section 114 of the Customs Act 1962, because the order relied on non-existent authorities, fake citations and hallucinated ratios. The court said that "assistance can never be substituted for adjudication" and sent the matter back for fresh adjudication by a different officer, according to SCC Online.
The lesson for multinationals is that AI hallucinations can void an adverse decision as well as sanction a lawyer. Companies facing Indian regulators should ask advisers in India to check the authorities in any order against them.
The Philippines: courts warn against overreliance
Philippine courts have so far emphasised correction and warning, placing responsibility squarely on the lawyer who signs.
In a 7 August 2025 resolution, the Sandiganbayan's Seventh Division flagged pleadings in the graft and malversation case against former Technology Resource Center deputy director general Dennis Cunanan. The citations were non-existent or did not contain the quoted passages, and had incorrect docket numbers and dates. The lawyer said AI tools had been used in researching case law. The court accepted her apology but reminded counsel to fact-check AI-generated research under the Code of Professional Responsibility and Accountability, stressing that responsibility fell on her because she signed the pleadings, The Philippine Star reported.
The judiciary's own position is set out in the Supreme Court's Governance Framework on the Use of Human-Centered Augmented Intelligence in the Judiciary, which requires that automated systems assist, but never replace, human judgment. Clients with matters in Manila can locate advisers in the Philippines through Find an Expert.
How AI hallucination sanctions compare across jurisdictions
The sanction varies by forum, but every jurisdiction reviewed here treats verification as the lawyer's personal duty, whatever tool was used.
| Jurisdiction | Leading example | Sanction or outcome | Principle |
|---|---|---|---|
| United States | Mata v Avianca (2023); Whiting v City of Athens (6th Cir., 2026) | US$5,000 fine; US$15,000 per lawyer plus fees and double costs | Sanction attaches to the fake citation, whether or not AI use is proven |
| England and Wales | Ayinde and Al-Haroun (2025); SDT case 12884 (2026) | Referral to regulators; strike-off from the register | Public AI tools cannot do reliable legal research; reliance is no defence |
| Australia | Mr Dayal (2025); Victorian murder trial (2025) | Practising certificate restricted, two years' supervision; 24-hour trial delay | AI output must be independently and thoroughly verified |
| Canada | Zhang v Chen, 2024 BCSC 285 | Personal liability for the extra costs caused | AI is no substitute for professional expertise |
| Singapore | [2025] SGHCR 33 | S$800 personal costs order | Non-delegable duty to ensure authorities exist and are accurate |
| India | Pooja Ramesh Singh; Gadiya (2026) | Orders set aside; ₹425.27 crore penalty quashed; Bar Council directed to act | Citing unverified AI authorities is misconduct; tainted decisions are void |
| Philippines | Sandiganbayan Seventh Division (2025) | Formal reminder; apology accepted | The signing lawyer bears responsibility |
Who carries the liability: signers, supervisors and firms
The lawyer who signs the filing carries the primary liability, but supervisors and firms are increasingly in scope, and clients who supply their own AI research are not a shield.
Three lines of authority point the same way. First, the signatory: the Sandiganbayan placed responsibility on the lawyer "as she signed" the pleadings, and the Sixth Circuit sanctioned each signing lawyer separately. Second, the supervisor: the SRA's notice says supervisors remain responsible for work they oversee, and ABA Formal Opinion 512 expects managers to set clear firm policies and supervisory lawyers to make reasonable efforts to ensure compliance. Third, the source of the research does not matter: in Al-Haroun the fabricated material came from the client, yet it was the solicitor who was referred to the regulator for failing to verify it.
For firms, the exposure is not limited to court sanctions. A referral can lead to regulatory proceedings, a wasted costs application can shift the other side's costs, and a client whose case is damaged may look to the firm's professional indemnity cover. We examined these wider questions in our analysis of who is liable when law firms use AI. Where a matter has already gone wrong, early advice from specialist professional liability advisers is essential.
What law firms must do now: verification, supervision and policy
Firms need a written AI policy, a mandatory verification step for every authority, named supervisory responsibility and training, because regulators now treat the absence of those controls as the failure.
Drawing on the SRA warning notice, ABA Formal Opinion 512, the Illinois ARDC's Guide to Implementing AI and the Singapore courts' guide, a defensible framework has six parts:
- Verify every authority at source. Check each case, statute and quotation in an authoritative database or the official reports, read the judgment itself rather than a summary, and confirm the proposition it is cited for. The SRA expects every named authority to be genuine, relevant and verifiable.
- Keep a verification record. Note who checked each authority, where and when. The Illinois guidance recommends maintaining verification documentation, which also helps if a court asks how a document was prepared.
- Approve tools, and ban the rest for client work. Specify which AI tools may be used for which tasks. Public tools should not receive confidential client data; the SRA warns that doing so can put privileged information into the public domain.
- Name the supervisor. Every AI-assisted filing should have a responsible lawyer who signs off the authorities personally. Supervision cannot be delegated to the tool or to the person who ran it.
- Treat client and third-party research as unverified. Al-Haroun shows that research supplied by a client, a former colleague or a foreign firm must go through the same checks.
- Train, audit and update. ABA Formal Opinion 512 expects firms to train lawyers and staff on tool use, confidentiality and ethical duties. Periodic audits of filed documents catch drift before a court does.
Firms building these controls often combine their knowledge, risk and technology functions; our directory lists specialists in legal sector management and legal and justice innovation. Regulators are not hostile to the technology itself, as the SRA's approval of the AI-driven firm Garfield AI shows; what they require is accountability for the output.
What in-house legal teams should ask of outside counsel
In-house teams should make AI verification a term of instruction, because an adverse decision, a wasted costs order or a delayed hearing lands on the client's matter even when the sanction lands on the lawyer.
The cases above suggest a practical set of questions for panel reviews and new instructions:
- Policy. Does the firm have a written AI policy covering litigation, and which tools are approved for court documents?
- Verification. Who checks each authority before filing, against which sources, and is that check recorded?
- Supervision. Which partner signs off AI-assisted work, and how are juniors and agents in other jurisdictions supervised?
- Confidentiality. Does any client data enter public AI tools, and what contractual and technical safeguards apply?
- Disclosure. Will the firm tell the client if a court raises a concern about citations, and how quickly?
General counsel should also look at their own teams. Where the business drafts material or research that goes to external lawyers, Al-Haroun is a reminder that it must be flagged as unverified. And where an opponent's filing or a regulator's decision cites authorities that cannot be found, the Indian cases show that pressing the point can undo an adverse ruling. Experienced dispute resolution and litigation specialists will test the other side's authorities as a matter of routine.
When to get professional help
Seek specialist advice as soon as a court, opponent or regulator questions the authorities in a filing you signed or instructed on. The early response matters: the Sixth Circuit's heavy sanction in Whiting followed counsel's failure to engage seriously with its show cause order. Advisers in professional liability and regulatory defence can manage the court's show cause process, any regulatory referral and notification to insurers. AI hallucinations in court filings are a solvable problem, but only for firms that verify, supervise and document. To follow further developments, subscribe to the Corporate INTL newsroom.
Frequently asked questions
What is an AI hallucination in law?
An AI hallucination in law is output from a generative AI tool that looks like legal authority but is false: an invented case, a wrong citation, or a real judgment given a quotation or holding it does not contain. Courts treat filing such material as a failure of the lawyer's duty to the court, not a technical glitch.
Is AI allowed to be used in court filings?
Generally, yes. Lawyers are not prohibited from using AI, but they remain responsible for everything they file. Regulators such as the SRA and the American Bar Association expect lawyers to verify every authority, protect client confidentiality and supervise AI-assisted work. Some courts also issue their own guidance, such as Singapore's guide for court users.
Which attorneys have been sanctioned for using AI?
Well-known examples include the lawyers sanctioned in Mata v Avianca, fined US$5,000 in 2023, and two lawyers ordered by the Sixth Circuit in 2026 to pay US$15,000 each. Outside the US, lawyers have faced personal costs orders in Canada and Singapore, practising restrictions in Australia and a strike-off in England.
Did a lawyer get in trouble for using ChatGPT?
Yes. In Mata v Avianca a New York lawyer filed six fictitious cases produced by ChatGPT and was fined US$5,000. In Zhang v Chen in British Columbia, a lawyer who cited two ChatGPT-suggested cases that did not exist was made personally liable for the extra costs they caused.
What are hallucinated case citations?
Hallucinated case citations are references to cases that do not exist, or to real cases with false details, produced by generative AI. They often look entirely plausible, with party names, years and court identifiers, which is why every citation must be checked against an authoritative database or the official law reports before filing.
Can a lawyer be struck off for AI fake citations?
Yes. In September 2026 the Solicitors Disciplinary Tribunal struck a registered foreign lawyer off the register after finding he had cited non-existent cases and incorrect legal propositions generated with AI. It was the first AI misuse case litigated before the Tribunal, and it shows that fake citations can end a career.
Who is responsible if a client supplies AI-generated research?
The lawyer who relies on it. In Al-Haroun v Qatar National Bank, 18 of 45 citations were fictitious and came from research the client had carried out with AI tools. The solicitor, who had not verified them, was referred to the Solicitors Regulation Authority. Client-supplied research must be checked like any other.
Can a court decision based on fake AI citations be overturned?
It can. India's Supreme Court held in July 2026 that decisions tainted by fabricated AI precedents are no decision in law and must be set aside. In September 2026 it quashed a ₹425.27 crore customs penalty because the order relied on non-existent and misattributed authorities, and ordered fresh adjudication.
How can law firms prevent AI hallucinations in court filings?
Adopt a written AI policy, approve specific tools, verify every authority against primary sources, record who checked what, name a supervising lawyer for each filing, treat client and third-party research as unverified, and train staff regularly. These steps reflect the expectations set by the SRA, the ABA and courts in Singapore and elsewhere.
Sources
- Reuters: AI error-ridden court filings surge despite three years of sanctions (17 September 2026)
- Reuters Legal: AI blunders in court pile up
- Damien Charlotin: AI Hallucination Cases Database
- Illinois Courts: Paste in Haste, the fallout of AI hallucinations in court filings
- Sixth Circuit Appellate Blog: Sixth Circuit sanctions attorneys for fake citations
- EDRM: The AI sanction wave, US$145K in Q1 penalties
- Reuters: US appeals court orders lawyer to pay US$2,500 over AI hallucinations in brief
- American Bar Association: ABA issues first ethics guidance on a lawyer's use of AI tools
- The Bar Examiner: ABA Formal Opinion 512
- The Expert Record: Ayinde and Al-Haroun, High Court on fake citations
- Solicitors Regulation Authority: Misuse of AI warning notice
- Solicitors Disciplinary Tribunal: Case 12884
- Victorian Legal Services Board and Commissioner: Statement on the Mr Dayal matter
- CBS News: Lawyer apologises for AI-generated errors in murder case
- CanLII Connects: Zhang v Chen commentary
- Singapore Courts: [2025] SGHCR 33
- The Hindu: Supreme Court voices concern over growing menace of AI-generated judgments
- LiveLaw: Citing AI-generated fake precedents is advocate misconduct
- SCC Online: Supreme Court sets aside ₹425.27 crore customs penalty
- The Philippine Star: Sandigan calls out lawyer who used AI in pleadings
About this article
This analysis was researched and written by The Corporate INTL Newsroom, which covers cross-border legal, regulatory and business developments for lawyers, professional advisers and financiers in over 150 jurisdictions. Every analysis is checked against court judgments, regulator decisions and primary reporting, listed in the sources above. It is general information, not legal advice; for a specific matter, consult a qualified adviser in the relevant jurisdiction. Last reviewed 23 September 2026.
