Where two worlds collide

Odysseus was warned: ‘You and your crew may still reach home, suffering all the way, if you only have the power to curb their wild desire and curb your own.’ This is equally valid when it comes to navigating the use of AI in legal practice. Fiona Campbell and Shauna Gillan traverse the wine-dark sea

The calendar now sits at September 2026. We have recently seen the beauty of a solar eclipse and, on an entirely unrelated note, can safely say that artificial intelligence (AI) is beginning to eclipse some traditional methods of legal practice – for some practitioners considerably more than others.

Put simply, many lawyers are already using AI-assisted tools across a wide range of tasks, including document search and classification, comparison, summarisation, structuring information, and producing first drafts.

Generative AI (GenAI) has extended those possibilities considerably: a user can now ask a system in natural language to produce legal research, draft submissions, summarise authorities, or suggest an answer to a legal problem within seconds. 

Against that backdrop, AI must still be used in a manner consistent with the existing professional obligations imposed on lawyers by the Law Society and Bar Council.

More recent Irish jurisprudence and guidance show the position developing quickly.

The Court of Appeal (Costello P) has addressed the use of AI-generated authorities in Guerin v O’Doherty; different courts and tribunals have begun to issue guidance of their own; and, on 29 July 2026, High Court President David Barniville issued Practice Direction HC142 on the responsible use of GenAI in court documents, which took effect on 1 September.

This was followed by CA16 from the Court of Appeal, to similar effect.

The practice directions state that, in light of the use of GenAI tools in legal proceedings, it is appropriate to highlight the associated risks, including inaccuracy, and to address the duties and obligations of parties and their legal representatives when such tools are used.

The emerging position is not one of prohibition, but rather an application of familiar professional duties to a technology capable of producing unusually convincing errors.

The recurring themes addressed by the regulators and the courts return to human responsibility, verification, competence and supervision, confidentiality and privilege – and the integrity of evidence and material placed before a court.

What do we mean by AI?

‘Artificial intelligence’ describes a broad range of technologies. For lawyers, an important distinction is between ‘discriminative AI’ and ‘generative AI’.

Discriminative AI identifies, classifies or predicts on the basis of existing data. Technology-assisted review/predictive coding in discovery is a familiar legal example: a system may classify documents by relevance or identify patterns within a document-set without itself creating substantive new content.

Generative AI produces new content in response to a prompt.

Large language models can generate text, summaries, translations, chronologies, draft clauses, research answers, or suggested submissions. They do so by identifying statistical patterns and predicting likely outputs.

That is quite different from a lawyer understanding a legal proposition and determining that it is correct.

The Bar of Ireland’s Ethical Toolkit describes discriminative AI as classifying or distinguishing data, while generative AI produces new content. It also cautions that GenAI outputs are predictions rather than the product of human understanding.

This means that the model uses probability to predict the most likely next word or sequence of words based on patterns in its training data, rather than understanding whether the resulting answer is legally correct.

Obviously, quite different technologies like these inevitably pose different risks; a document-classification system may misclassify material or reflect bias in its data, whereas a generative system may go further and create a judgment that does not exist (known as ‘hallucination’ for open-source data and ‘fabricated output’ for enterprise-level licensed data), misstate a statute, invent a quotation, or produce a plausible legal argument with no proper foundation.

An AI output will never assume the lawyer’s professional responsibility for deciding whether that output can safely be used. With that in mind, best practice for AI-use in legal work would be to follow a model like this:

Prompt → Output → Human review → Verification → Decision

Professional guidance

The Law Society (Guidelines for the Use of Generative Artificial Intelligence by the Legal Profession in Ireland, December 2025) and the Bar (Ethical Toolkit: Ethical Use and Use Cases for Artificial Intelligence in Legal Practice, April 2025) have each published guidance on the use of GenAI in legal practice.

As with any guidance, these publications are intended to assist practitioners in applying their existing professional duties when using AI – relied upon in tandem with the professional standards by which solicitors and barristers are already regulated.

There is considerable overlap in the principles identified by both bodies: practitioners remain responsible for the accuracy and reliability of their work and must protect confidential and privileged information, exercise appropriate competence and supervision, and independently check AI-assisted legal research or drafting before relying upon it or submitting it externally.

The Bar expresses the practical approach as:

Prompt – Review – Verify – Repeat

The Bar’s guidance goes further in relation to material that is intended to represent another person’s own evidence or opinion.

It advises that GenAI should not be used to draft affidavits, statements of truth, or witness statements, because those documents must contain the evidence of the person giving it.

It takes a similarly restrictive approach to expert reports, save for limited circumstances where AI itself forms part of the subject matter and its use is appropriately identified.

Those considerations are also reflected in the aforementioned new practice directions.

Other jurisdictions

Since ChatGPT was released on 30 November 2022, courts across a number of common-law jurisdictions have had to grapple with an expanding range of problems arising from GenAI use in legal proceedings.

The best-known starting point was Mata v Avianca in 2023, where a New York lawyer used ChatGPT for legal research, and submissions were filed containing non-existent cases and fabricated judicial opinions, ultimately resulting in sanctions.

Other examples followed quickly: in Canada, Zhang v Chen concerned reliance on fictitious authorities generated by ChatGPT; in Kohls v Ellison in Minnesota, an expert used GenAI in preparing a declaration that contained citations to non-existent academic material; and in Australia, Valu v Minister for Immigration and Multicultural Affairs (No 2) concerned a lawyer who incorporated AI-generated authorities into submissions without checking them and was referred to the professional regulator.

The jurisprudence has, therefore, developed beyond the straightforward problem of invented cases and now extends to questions of evidence, professional supervision, confidentiality, privilege and the extent to which lawyers can properly rely on AI-assisted work produced by others.

One particularly interesting development arose in United States v Heppner in February 2026.

Bradley Heppner, who was the subject of a criminal investigation, had used Anthropic’s Claude to prepare material addressing his anticipated defence, facts, and legal issues, without having been directed by his lawyers to do so.

The Southern District of New York held that his exchanges with Claude were neither protected by the US jurisprudence of attorney/client privilege nor by the work-product doctrine.

The court also held that, to the extent Heppner had entered information that was already privileged into Claude, that privilege had been waived by disclosure to Claude and Anthropic.

(Note that this decision turned on the particular circumstances, including that the exchanges were undertaken independently of counsel and the applicable data and privacy arrangements, so it should not be read as establishing that every interaction with an externally hosted AI system automatically waives privilege. It does, however, demonstrate that courts are now being asked to consider the consequences of AI use for privilege itself, rather than merely the accuracy of what AI produces. We will not discuss here the differences between privileged status in open source data LLMs (for example, open source Chat GPT) and licensed enterprise-level systems, for which the terms may define separate privilege and confidentiality considerations.)

‘Atrocious conduct’

More recently, and despite that backdrop of judicial warnings, the conduct considered in Miller v Regions Bank in May 2026 went considerably further.

Alabama attorney Gregory Harp filed a brief containing four false quotations attributed to genuine authorities.

When the US District Court subsequently ordered production of his ChatGPT history, Harp deleted his account and took further steps that prevented the relevant history from being produced. 

The court drew an adverse inference as to what the deleted material would have shown and said that it had never expected to encounter “such atrocious conduct” from an officer of the court.

Importantly, the court made clear that the sanction was not for the mere use of GenAI. It was concerned with the false legal material, the absence of reasonable verification, lack of candour, and the destruction of evidence after the problem had emerged.

Harp was publicly reprimanded, disqualified from further participation in the case, referred to the relevant licensing authorities, and suspended from practice in that federal district for six months.

Closer to us geographically and in legal tradition, England and Wales has developed its own line of authority.

The leading decision in 2025 was R (Ayinde) v London Borough of Haringey and Al-Haroun v Qatar National Bank QPSC.

The two matters were heard together under what is known as the ‘Hamid jurisdiction’, through which the courts of England and Wales exercise their inherent supervisor powers in relation to the duties lawyers owe to the court and the integrity of the court process.

In Ayinde, judicial review grounds settled and signed by a pupil barrister cited five cases that did not exist. The court did not find as a fact that she had used GenAI, and identified two possible explanations, one of which was that GenAI had been used and its output had not been checked.

In Al-Haroun, the scale was greater: 45 authorities had been cited, 18 of which did not exist, while a number of the remainder did not support the propositions or quotations attributed to them.

One fictitious authority was purportedly a decision of Dias J – the very judge before whom the material was placed.

The solicitor had relied on research carried out by his lay client, who accepted using publicly available AI tools among his research sources, without independently verifying the resulting authorities.

The court’s message was direct: lawyers remained responsible for material placed before a court, whether they produced it themselves, delegated the work to a junior, or relied on research supplied by somebody else.

AI-assisted research must be checked against authoritative sources.

The court considered the possible use of contempt powers, made or considered regulatory referrals, and stressed that the profession could expect a serious response where these obligations were not met.

Confidentiality and privilege

The English cases have continued since Ayinde. In MS (Professional conduct; AI generated documents), a barrister ultimately accepted having used ChatGPT in work that relied on a fictitious authority and was referred to the Bar Standards Board.

In UK and R, the Upper Tribunal dealt with false authorities, inadequate supervision of junior legal staff, and the use of ChatGPT with client material.

The tribunal took a particularly strong position on confidentiality and privilege, stating that uploading confidential material into an open-source AI tool, such as ChatGPT, placed it in the public domain, breached confidentiality, and waived privilege.

As with our brief privilege comment regarding the US Heppner matter above, there is a lot more to this decision that would need, in substance, its own article to discuss.

Two further English decisions in 2026 demonstrate that the problem is not confined to invented case law. In Cork v Smith, a junior lawyer at international law firm Pinsent Masons used AI in legal research, and an invented quotation from the Insolvency Rules subsequently found its way into professional correspondence.

Neither the junior lawyer nor those supervising the work had checked the statutory source.

More recently again, in Tobosaru and Tofan (8 July), formal submissions prepared for the Crown Prosecution Service contained two non-existent authorities. The CPS explained that the problem appeared to have arisen through AI use combined with a failure of human verification.

The errors had not been identified by those preparing or reviewing the submissions, or by counsel, but by the opposing party.

The court accepted the apology and remedial action taken, but recorded the episode as a further warning about the need for proper verification.

Without repeating the absolute obvious, the common theme across all of these when they come before a court remains the same – independent human verification is key.

Irish case law

Irish courts began sounding warnings before any detailed procedural framework existed. In Coulston & Ors v Elliott & Anor, Nolan J encountered submissions advanced by a litigant-in-person who could not explain the new legal argument contained in them and said that a friend had prepared the material.

The judge did not make a finding that AI had, in fact, been used. He considered it highly likely either that somebody had purported to act as a lawyer or that a generative AI programme had been used.

In warning of the latter possibility, he observed that GenAI could sound persuasive while being legally flawed.

Reddan v An Bord Pleanála provides another caution against overstating the evidence of AI use.

An unfamiliar expression used by a litigant-in-person prompted Nolan J to say that it sounded as though it had come from an AI source and had the hallmarks of ChatGPT or a similar tool.

Again, that was an inference rather than a finding that GenAI had been used.

More fundamentally, the court was concerned with serious allegations being advanced without an adequate evidential or legal basis.

The position became considerably clearer with the Court of Appeal decision in Guerin v O’Doherty (March 2026).

The proceedings concerned a civil defamation claim brought by James Guerin, a brother of the late journalist Veronica Guerin, against Gemma O’Doherty in respect of online and print publications made in July 2019. 

A first trial before a judge and jury in November 2023 ended when the jury was unable to agree a verdict, and a retrial was directed.

O’Doherty subsequently sought to have the proceedings struck out; when that application was dismissed by the High Court, she appealed to the Court of Appeal.

There, O’Doherty, who was acting as a litigant-in-person, had used AI to assist in preparing written submissions, which included authorities that did not exist.

Costello P described those authorities as hallucinations generated by the AI system.

The court nevertheless did not suggest that AI-assisted legal research was prohibited.

Parties may use AI for research, but they remain responsible for the resulting submissions and must independently establish that an authority is genuine and, at least, arguably supports the proposition for which it is cited.

The court also said that parties using AI for legal research should expressly inform the other parties and the court. It warned that sanctions may be available where misuse of AI has the potential to mislead the court.

On the facts, however, the court did not find that the defendant intended to mislead it and took account of the absence at that point of prior guidance directed to litigants-in-person.

That should not be taken to mean that the same latitude will necessarily be available in future cases.

The judgement is better treated as a warning shot: where parties now have the benefit of published judicial and professional guidance, a repetition of the same failures of verification may attract a materially different response. 

The Court of Appeal has acted upon its experience in Guerin and, on 11 August 2026, issued Practice Direction CA16, operative from 1 September 2026. This is designed to align with HC142 (see para 1), ensuring that its provisions continue to apply to any appeal from the High Court to the Court of Appeal.

Practice Direction HC142

The first of the High Court’s formal procedural framework for GenAI use in civil proceedings came into operation on 1 September 2026. The scope of HC142 applies to:

• All documents used in civil proceedings, including pleadings, submissions, affidavits, witness statements, and expert reports – collectively defined as ‘court documents’, 
• All persons involved in proceedings before the High Court, including parties, legal practitioners, witnesses, experts, and any third party required by the court to produce documents,
• All High Court proceedings, whether instituted before or after commencement of the practice direction, and
• All court documents prepared, produced, or generated after the date of commencement.

Unsurprisingly, like the other court-issued and regulatory guidance we have seen to date concerning GenAI, HC142’s central requirement is independent human verification of all source material – the verification standard differing from the type of court document, for example, the content of pleadings and submissions must be accurate; witness evidence must remain with the witness for witness statements and affidavits, meaning that any AI rewrites or finessing must not alter or drift from the witness’s own words; expert reports must represent the expert’s own opinion and reasoning, and the use of GenAI in drafting the report, analysing or interpreting information, or generating substantive content must be disclosed in accordance with the practice direction.

The consequences of misuse are also made explicit. AI-related errors may lead to corrective steps, costs consequences, rejection or disregard of material, strikeout in an appropriate case and, for legal practitioners, possible referral to the relevant professional or regulatory body and personal costs orders.

Use and disclosure

There is currently no single rule requiring disclosure of every use of AI across Irish legal practice.

In Guerin, the Court of Appeal said that parties using AI to assist with legal research should expressly inform both the court and the other parties, whereas the Law Society’s guidance states that solicitors are under no general positive obligation to disclose GenAI use to clients or other solicitors merely because it has been used, unless specifically asked or another transparency obligation applies.

The position also varies by forum: the Labour Court states that a party using AI may wish to reference its use; the International Protection Appeals Tribunal requires disclosure where submissions have been prepared using AI tools, even in part; and the Workplace Relations Commission asks parties to be open about AI use where appropriate.

HC142 adopts a more detailed approach for the High Court: GenAI use in preparing a court document does not, of itself, require disclosure, although the court may require a party to disclose and explain the nature and extent of that use, and specific requirements apply to affidavits, witness statements, and expert reports.

The position is, therefore, document-specific and forum-specific rather than a blanket disclosure obligation. 

As a practical matter, practitioners should keep a proportionate record sufficient to identify any material AI use and the checks undertaken, particularly for court-facing work, so that they can explain that use if required.

What this means in practice

• For private practitioners and barristers, the practical position follows from the guidance and authorities discussed above: understand the tool, protect confidential and privileged material, independently verify substantive AI-assisted work, and remain professionally responsible for the result. Where work falls within HC142 or another applicable court or tribunal requirement, that more specific framework must, of course, also be followed.
• The use-cases for in-house lawyers are likely to be considerably broader and will vary between organisations. AI might be used to summarise or compare documents, produce first drafts, assist with contract review or legal research, support internal investigations, or help the wider business with functionsranging from HR and procurement to compliance and product development. Those activities will not necessarily be subject to an express procedural framework and, therefore, a proportionate, risk-based approach and validation log provides a useful practical standard.

The obvious starting point should be the use-case: what is the AI being asked to do, what information is being supplied to it, what system is being used, and where will the output ultimately go?

A low-risk internal administrative task calls for a different level of scrutiny from legal advice, regulatory analysis, litigation work, or a decision affecting an individual.

As the significance of the output increases, so too should the level of human review, verification, and oversight.

The same approach applies to the tool itself. Consumer, enterprise, and legal-specific AI systems may have materially different arrangements for data access, retention, security, and use for model training. In-house teams should, therefore, understand the environment in which information is being processed, particularly before supplying confidential, privileged, personal, or commercially sensitive material.

Ultimately, the best practice is familiar: use AI proportionately to the task, verify what needs to be relied upon, preserve professional judgement, and be able to explain how a material output was produced and checked. Maintain a log or record of your decisions and results.

Changing landscape

Ireland enacted the Regulation of Artificial Intelligence Act 2026 in July, establishing the AI Office of Ireland as the central coordinating authority within Ireland’s distributed regulatory framework.

Other competent authorities include the Data Protection Commission and, for relevant workplace uses, the WRC.

The act is subject to commencement provisions and came into force, only in part, on 31 July 2026.

At EU level, the AI Omnibus entered into force on 27 July 2026 and revised aspects of the AI Act timetable, including the high-risk regime, while certain transparency obligations have applied since 2 August 2026.

For in-house teams, AI governance, therefore, extends beyond technology policy and may engage data protection, employment, consumer, contractual, and other sectors-pecific obligations, depending on the use-case.

The principal principle

AI is a tool, not a decision-maker. It may make legal work faster, but professional judgement, responsibility, and verification remain human. An apparently convincing submission is still only as good as the lawyer who checks it.

There is, perhaps, an apt illustration in Stanley Kubrick’s 2001: A Space Odyssey. HAL 9000, the film’s highly capable AI-system, plays a game of chess against astronaut Frank Poole and, in doing so, gives an incorrect description of one of its moves.

The error is easy to miss because HAL delivers it with complete confidence and apparent authority, and only someone following the game closely would recognise it.

The lesson for lawyers is equally simple: confidence is not verification. Never trust HAL to make the final call.

Fiona Campbell is an Irish and UK-qualified solicitor and former Irish barrister. She leads Legal AI Innovation at Lexis Nexis UK. Shauna Gillan is an Irish and UK-qualified barrister and deputy chair of the International Protection Appeals Tribunal.