Key Takeaway
Matthew Lee and Tomás Keys discuss the current position with suspected and confirmed AI hallucinations in legal work in Ireland. There are now 8 known incidents.
Publication date: 7 April 2026
Ad/Marketing Communication
This legal article/report forms part of ongoing legal commentary on the use of artificial intelligence within the justice system. It supports work in teaching, lecturing, and writing about AI and the law and is published to promote the Barristers’ practice. Not legal advice. Not Direct/Public Access. All instructions via clerks at Doughty Street Chambers. This legal article concerns AI hallucination cases ireland.
Introduction
This week has been relatively quiet in the UK courts when it comes to AI hallucinations. The same, however, cannot be said for the courts and tribunals of Ireland. I am very grateful to everyone who has sent me the decisions, including one that sets out some important principles. Those developments prompted me to get back in touch with Tomás Keys, with whom I wrote a previous article around Halloween, so that we could update our earlier post on the four confirmed or suspected AI-generated incidents in the courts and tribunals.
Since then, there have been two further instances in the Workplace Relations Commission, as well as two judgments delivered by the Court of Appeal. Three of those four instances were the subject of written decisions or judgments in March 2026 alone.
Following on from the guidance published by the Workplace Relations Commission, the Labour Court, which hears appeals from certain decisions of the WRC, has also published guidance to users on the use of AI. In the second judgment of the Court of Appeal discussed below, the President of the Court of Appeal identified principles on the use of AI regarding the preparation of written submissions in the Courts of Ireland.
Goncalves v. Valshan Unlimited (ADJ-00056185)
In this case, the complainant, who was a spa therapist, lodged 5 separate complaints against her employer in the space of 3 days in relation to discrimination on family and gender grounds, equal pay, and penalisation in December 2024.
She had previously succeeded in a claim for discrimination against the same employer (ADJ-00048914) and was awarded €7,000 in compensation.
The adjudicator noted that counsel for the Respondent employer made the following:
“…strident submissions that the documents filed by the complainant prehearing were generated by AI and were not admissible in the case. These documents contained clear mistakes and raised credibility issues on the complainant’s side.
He summarised that the respondent had “been bombarded by repeated dossiers which now stood at 1,000s of pages without any details on key claims, such as the last complaint of penalisation”.
In their decision, the adjudicator noted that they had invited the applicant to net the issues; the complainant replied, “Violations exist and cannot be reduced.” The adjudicator noted that the complainant had a clear difficulty in participating in the process and that she exhibited “high-level hostility” towards the respondent’s representatives.
Before disposing of the five complaints on the merits, the adjudicator highlighted the difficulties with the materials used as follows:
“Unfortunately, for me at least, the tangential scattergun, smorgasbord approaches and references coupled with euphemisms formed outside of Europe approaches of this tool were a poor second to human oversight, I have come to expect and anticipate in actual statement and representation during a case.”
Ferko v. Beyond Reach Limited (ADJ-00060622)
The complainant was a litigant in person, and the respondent employer, which was a company, was represented by its managing director. The complaint arose from an allegation that there had been a delay of a year in providing the complainant with a contract of employment. The evidence from the respondent was that the contract had been provided to the complainant within days of commencing employment as part of the induction procedures for all new employees.
The adjudicator found that the matter turned on conflicting oral evidence. In considering the totality of the evidence, the adjudicator preferred the evidence of the Respondent:
“The Respondent’s evidence regarding its induction procedures and the provision of written contracts to new employees was clear, consistent and credible. The account provided was coherent and aligned with ordinary employment practices.
By contrast, the Complainant accepted that the complaint submitted to the WRC had been prepared by his sister and generated using artificial intelligence and that he had not fully reviewed or verified the contents prior to submission. As a consequence, he was unable to clearly explain aspects of the complaint advanced in his own name. While assistance in preparing complaints is not improper, a complainant remains responsible for the accuracy of allegations made on his behalf. In the circumstances, this materially reduced the weight that could be attached to the Complainant’s evidence.”
Von Geitz v. Kelly & Robertson [2026] IECA 29.
On 16 March 2026, Allen J. delivered a written judgment in the Court of Appeal in relation to the appeal taken by Mr Von Geitz against orders of the High Court striking out his proceedings on the basis that they were bound to fail and amounted to an abuse of process, and an Isaac Wunder Order (an order that the Irish courts can give restraining repeat litigants from issuing further proceedings without first getting leave of the court). In a somewhat unusual feature of the case, the Appellant appealed against an order to which he had consented in the High Court.
Allen J. stated at paragraph 153 as follows:
“It is established that litigants in person can expect to be afforded a certain amount of latitude in the conduct of proceedings but equally, that the same rules apply whether the litigant is represented or not. No more than a represented litigant, an unrepresented litigant is not entitled to appeal against an order to which he has consented on the grounds that, on reflection, he considers that he might have had good grounds to oppose it.”
Allen J. was critical of the submission of incorrect citations and non-existent cases during the course of the appeal. He stated:
“The plaintiff’s written submissions are littered with propositions of law in quotation marks which are unsupported by authority; reference to authorities which have nothing to do with the asserted propositions of law; wrong citations; and a few non-existent cases. If, as was suggested at the hearing of the appeal, the non-existent cases were hallucinations generated by AI, it was nevertheless the responsibility of the plaintiff – as it is of every litigant – to check whatever may have been thrown up by whatever tool he used to ensure in the short term that his opponents were not sent on a wild goose chase and ultimately that the Court was not presented with rubbish.”
Guerin v. O’Doherty [2026] IECA 48
While it was suspected that AI had been used to generate materials in the Von Geitz case, the use of AI by a litigant in person was confirmed in this matter.
The appeal concerned the High Court’s refusal of the defendant’s motion to strike out the proceedings:
“on the grounds that they are malicious, vexatious, frivolous, based on fraud, perjury and deception, are a grave attack on press freedom, have no cause of action and are bound to fail.”
The Plaintiff was represented by a solicitor and counsel, while the Defendant/appellant was a litigant in person.
Costello P noted the use of AI by the Defendant but did not criticise her for doing so, as there was no guidance available to litigants in person in relation to the use of AI when preparing submissions. She also set out some general principles on the use of AI.
“The defendant used AI to prepare the written submissions, which she filed in support of her appeal. Unfortunately, and perhaps as was to be expected, given the nature of large word model Artificial Intelligence systems, the submissions included references to authorities which simply did not exist. These were hallucinations, so-called, generated by the AI system. There were no such cases, and they were not authority for the propositions which they purported to establish. This is an inherent and well-known risk of using AI to write legal submissions. The defendant did not apparently verify the existence of the authorities she cited, or that the cases relied upon actually supported the propositions advanced. Neither did she notify the solicitors for the plaintiff that she had prepared her submissions with the assistance of AI. Counsel for the plaintiff informed the Court that this added to their work in requiring them fruitlessly to attempt to locate the hallucinated authorities and thereby needlessly added to the costs of the plaintiff in the appeal.
Parties, whether represented or not, have an obligation not to mislead the court, which includes the obligation not to rely upon or advance submissions based upon “fake” authorities or propositions which have no basis in law. In addition, lawyers are subject to professional and ethical obligations in relation to the use of AI when practicing their profession which do not apply to a litigant in person. I do not propose to address these in this judgment as they do not arise in this case.
I am concerned that parties, including litigants in person, should use AI appropriately and should be given guidance as to how they may properly use AI in their litigation. I would therefore set out the following principles of general application:
(i) Parties are entitled to use AI to assist in carrying out research in respect of their case, provided that they do so responsibly and do not, even inadvertently, mislead the court by advancing propositions or relying upon supposed authorities which in fact have no foundation at all and are simply hallucinations.
(ii) In all cases where they do so, they should expressly inform both the other parties and the court of their use of AI in this regard.
(iii) A self-represented party is responsible for the ultimate written or oral work in their case, just as much as the lawyers representing parties are.
(iv) It is important, therefore, that any party who uses AI as part of their research independently verifies the accuracy of their submissions and the authorities cited as supposedly establishing the propositions advanced.
(v) No authority should be cited by a party who has not actually verified that it is a genuine judgment of the court and that it is – or at least arguably is – authority for the proposition contended for.
It is not acceptable for parties uncritically to provide submissions which purport to rely on cases or propositions, the existence of which they have not verified. It leads to completely wasted time and costs. It casts an unfair burden on the opposing party in their preparation of their response to the submission or in preparing the books of appeal. It potentially brings the administration of justice into disrepute and may result in misleading the court. Parties should be aware that the court has a variety of sanctions open to it where parties use AI in breach of these guidelines and where their use has the potential to mislead the court.
For the avoidance of doubt, I wish to make it clear that I do not believe that the defendant intended to mislead the court or that she was actually aware of the fact that some of the cases cited in her written submissions were hallucinations. While it would have been preferable if she had informed the plaintiff’s solicitors of her use of AI to produce her written submissions, and thus to the likelihood that the authorities they could not readily locate were hallucinations, I would not draw any adverse conclusion against her. At the time submissions were prepared in this appeal, no guidance was available to litigants in person in relation to their obligations to the other parties to the proceedings and the court as regards the use of AI -generated material in proceedings.”
Guidance on the use of AI published by the Labour Court.
Following on from the guidance published by the WRC in October 2025 as to the use of AI when preparing cases in the WRC, the Labour Court, which hears appeals from certain WRC decisions, published guidance on the use of AI when preparing submissions or other documentation to the Labour Court. That guidance can be obtained by clicking the link on the heading above.
Conclusion
With the identification of principles by the President of the Court of Appeal on the use of AI in Irish courts and the guidelines published by the Labour Court in March 2026, following those issued by the WRC in October 2025, we hope these developments will help reduce the incidence of hallucinated case law and citations in Irish courts and tribunals. At the same time, the wider international picture does not appear to be slowing. On the contrary, the number of reported instances continues to grow, which means this is an area we will need to keep under regular review and update accordingly.
If you have found it helpful, you may wish to subscribe to Matthew’s AI law articles via the Substack below. Please do continue sending decisions that you think should be tracked, whatever the jurisdiction. One of the clearest lessons from this work is that lawyers, judges and court users in different systems are often confronting closely related problems. Sharing decisions, experiences and reflections across practice areas and across jurisdictions remains one of the most constructive ways to respond.
Final Words from Popular LLMs
This is the point at which we invite an LLM to provide an independent comment on the preceding article. At times the observations are slightly off the mark, at other times they are perceptive, and occasionally they produce something entirely unexpected. Whatever the outcome, I find it a useful and revealing exercise to see what emerges.
Chat GPT (Open AI)
“From a legal-risk perspective, the key issue is not AI use in the abstract but responsibility for what is filed or argued in a party’s name. The Irish decisions reviewed here show a consistent theme: parties must verify authorities for themselves, fabricated or misdescribed citations can waste opponent and court time, and AI-assisted material can draw judicial criticism, costs consequences or sanctions risk even where the court does not infer any intention to mislead. The safest approach is to verify every citation, attribute serious allegations precisely to the court/tribunal or the party making them, and distinguish clearly between confirmed AI use, suspected AI assistance, and proven hallucinated authorities. Not legal advice”
Gemini (Google)
“The recent guidance from the Irish Court of Appeal and the Labour Court represents a necessary and welcome evolution in the intersection of artificial intelligence and legal practice. As LLMs become more accessible, the temptation for litigants in person to rely on them for complex legal drafting is understandable, yet fraught with the risk of ‘hallucinations’—as these cases starkly illustrate. The principles set out by Costello P. rightly strike a pragmatic balance: they do not prohibit the use of AI tools for research, but they firmly reinforce the indispensable human elements of independent verification, transparency, and ultimate accountability. For AI to serve as a meaningful asset in the administration of justice rather than a disruptive liability, it must be treated as a collaborative tool that augments, rather than replaces, rigorous legal scrutiny.”
Claude (Anthropic)
“The cases documented here reflect a pattern that is both predictable and preventable. Large language models generate text that is fluent and confident by design — they do not flag uncertainty in the way a careful researcher would. When used for legal work without verification, this creates a particular hazard: hallucinated citations look indistinguishable from real ones until someone tries to find them. The principles set out by the President of the Court of Appeal are a sound and proportionate response — not a prohibition on AI use, but a clear statement that the responsibility for accuracy remains with the party submitting the work. The emerging guidance from the WRC and the Labour Court points in the same direction. AI can be a genuinely useful research aid in legal contexts, but it functions best as a starting point for human scrutiny, not a substitute for it.”
