Judicial Responses to AI-Generated Fabrications in Legal Practice
By Nicholas Blomfield
The adoption of generative AI (GenAI), ranging from specialist legal platforms such as Harvey and Legora to general-purpose large language models has transformed legal research, drafting and advice. However, alongside these technological advances lies a persistent vulnerability; the tendency of legal probabilistic models to “hallucinate” plausible yet entirely fictitious case law, statutory provisions, and judicial reasoning.
Where fabricated precedents are drafted into court filings, a critical issue arises for the judiciary and regulatory bodies; can a practitioner rely on “AI hallucination” as an exculpatory defence, or may the failure engage professional discipline, personal costs liability, negligence or even contempt?
Recent cases across the UK, Canada, and Australia demonstrate the judicial change in this area. Courts have moved from initial (somewhat reluctant) forbearance toward a strict, non-delegable standard of responsibility. This article examines key case law including: Harber, Ayinde, and Munir v Secretary of State, alongside persuasive international precedents to illustrate why the “hallucination defence” is firmly rejected by the courts.
The UK judiciary’s relatively recent and first direct confrontation with AI-generated legal fabrications occurred in Harber v Commissioners for HMRC [2023] UKFTT 1007 (TC). In an appeal against tax penalties, the self-represented appellant provided fabricated authority citations which she accepted might have been generated using ChatGPT to support her arguments. Upon cross-checking, neither HMRC nor the tribunal could locate the authorities. It transpires that the cases were complete fabrications containing fictitious judge names, quotes, and legal outcomes.
The First-tier Tribunal (FTT) rejected the fictitious authorities and predictably placed no reliance on them, reiterating that judicial decisions must be rooted in authentic law. The tribunal accepted that the appellant had been unaware that the authorities were fabricated, and provided guidance on how genuine decisions could be independently verified.
Harber illustrated two early points: 1. AI-generated fabrications carry zero weight in judicial proceedings; and 2. While Litigants in Person (LiP) might initially receive judicial leniency due to ignorance of Large Language Models (LLM) architecture and possibly legal experience; licensed practitioners should expect no such clemency.
By 2025, the High Court unequivocally set the standard for qualified practitioners, at least in the UK, in Ayinde v London Borough of Haringey; Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin).
In Ayinde, acting counsel presented fabricated authorities in court submissions, in circumstances that raised serious questions about whether generative AI had been used. Unlike in Harber, the court was dealing with admitted members of the legal profession bound by professional practice and ethical obligations. Addressing both the individual failure and the broader systemic implications, the judgment can be distilled into a framework of Responsible Legal Augmentation: 1. Independent verification of authorities – AI cannot be used as a shortcut to bypass verification; 2. Avoid fabricated citations – this amounts to a direct breach of professional conduct; 3. Uphold the Duty to the Court – do not intentionally or unintentionally mislead the Court; 4. Maintain human accountability – the legal professional is entirely responsible for the product; and 5. Exercise professional competence – lawyers must understand the tools they are using. Ayinde effectively dismantled the “defence of hallucination” for legal counsel and firms. Claiming ignorance of an LLM’s limitations or relying on software unreliability does not excuse a failure to verify material submitted to the court and may engage professional, regulatory and costs consequences.
The judicial response intensified in UK and R (on the application of Munir) v Secretary of State for the Home Department (AI hallucinations; supervision; Hamid) [2026] UKUT 81 (IAC). The Upper Tribunal (Immigration and Asylum Chamber) addressed, amongst other things, the intersection between AI hallucinations, duty of honesty in judicial review proceedings, and the tribunal’s supervisory jurisdiction under the “Hamid principles”.
The Upper Tribunal warned that submitting hallucinated or distorted legal authorities in public law claims against the Secretary of State constitutes a serious abuse of court process. Applying its existing supervisory jurisdiction under R (Hamid) v Secretary of State for the Home Department [2012] EWHC 3070 (Admin), the Tribunal signaled that legal representatives who fail to supervise synthetic outputs face formal referrals to regulatory bodies (such as the Bar Standards Board or Solicitors Regulation Authority) and adverse costs consequences.
Beyond hallucination, Munir addressed a secondary and not so apparent risk; the Court ruled that uploading confidential client data into open-source, consumer-grade AI tools such as ChatGPT and Claude places that information into the public domain permanently destroying legal privilege and breaching client confidentiality.
The UK courts’ firm stance aligns with emerging jurisprudential consensus across common law jurisdictions, demonstrating a global standard of care and zero tolerance in relation to breaches of professional obligations.
In Zhang v. Chen, an application in the Supreme Court of British Columbia, counsel unwittingly included non-existent case citations generated by ChatGPT in a family law matter. Upon discovery, counsel promptly withdrew the fake cases and apologised, explaining she was unaware of ChatGPT’s hallucination risks.
Justice Masuhara held that, although using AI for legal research is not inherently improper, citing fictitious authorities is a serious error that causes opposing counsel and the court to expend substantial time and resources investigating and correcting the resulting confusion. While the court declined to award special costs in this case, it noted that such conduct could expose counsel to personal special-costs liability. The matter also attracted scrutiny from the Law Society of British Columbia. Counsel was nevertheless ordered personally to bear the additional ordinary costs caused by the false citations and the delay in correcting them, and to review her other court files for similar material.
In Australia, courts and tribunals have proactively issued practice notes regarding Generative AI. Australian courts have maintained that the duty owed by a legal practitioner to the court is paramount, and non-delegable including to reliance on technology. Submitting hallucinated authorities is treated as a breach of the practitioner’s personal warranty that cited authorities are accurate, and relevant to the case. The absence of an intention to mislead does not remove the duty to verify, although intent remains relevant to dishonesty, contempt and the appropriate sanction. The breach lies in submitted unverified material to the court, violating the duty of competence and honesty.
Delegation without verification (i.e. Human-in-Command) is negligence. High workload or reliance on third-party technology does not displace counsel’s personal duty to the court.
The case law across various jurisdictions delivers a clear message; LLMs are legal assistants, not legal practitioners and ownership, and accountability remains with the professional practitioner.
While tools like Harvey AI, Legora, and general LLMs offer utility in streamlining research and drafting, they operate (at least for now) on probabilistic pattern matching rather than genuine legal understanding. The “defence of hallucination” is futile. As confirmed in Ayinde, Zhang v. Chen, and Munir v Sec State, the responsibility for every word, and every citation filed in court rests with the practitioner and, where applicable, the supervising lawyer and firm responsible for the work.
Legal practice in the age of AI requires responsible augmentation, leveraging technology to enhance capability and efficiency, while maintaining strict, independent human verification to protect the integrity of the judicial process. Synthetic counsel is useful only to the extent that it assists the lawyer. When an LLM hallucinates, accountability rests entirely with the Lawyer-in-Command, not the algorithm.
Nicholas Blomfield is a dual-qualified solicitor (England & Wales and Ireland) with over fourteen years’ experience across leading global institutions, including BlackRock, J.P. Morgan, Barclays and FIS. He most recently served as Acting General Counsel and Deputy General Counsel at Sentora, leading legal, regulatory and transactional work across institutional DeFi.
He specialises in financial services, digital assets and technology,
