Practice Direction HC 142 came into force on 1 September 2026. It reaches finance directors, HR leads and expert witnesses, and it asks for evidence most organisations cannot produce.
Practice Direction HC 142 came into operation on 1 September 2026. It was issued on 29 July by the President of the High Court, Mr Justice David Barniville, and it governs the use of generative AI in documents filed in civil proceedings.
The legal press covered it as a story about barristers and invented case citations. That is the smallest part of it.
The part that matters to a board is the sanction. A practitioner who puts unverified AI-generated material before the court can be ordered to pay costs personally, referred to the Legal Services Regulatory Authority, and have part of a claim or defence struck out. A named individual, a named cost, attached to a failure to check what an AI tool produced. No Irish AI regulation currently does that with the same precision, and the AI Act's high-risk obligations are still fifteen months away.
What HC 142 actually requires
The Court is explicit that generative AI can improve efficiency and widen access to justice. The obligation is about care rather than prohibition.
Verification is personal and non-delegable. Anyone putting AI-assisted material before the court must verify its accuracy first. The Court names hallucinated authorities and deepfaked material as the risks in mind. "The software produced it" is no defence, and the costs order follows whoever filed it.
Affidavits and witness statements now carry a declaration. From 1 September, a deponent must declare that generative AI was not used to generate the substantive content of the statement. The Practice Direction spells out what that covers: altering, embellishing, strengthening, diluting or rephrasing the evidence. The document must reflect the deponent's own personal knowledge and recollection. Formatting and spellchecking are fine.
It is retrospective in reach. It applies to all civil proceedings, whether instituted before 1 September 2026 or after. A case that opened in 2023 is caught by it today.
The people this catches have never called themselves AI users
Read the declaration requirement again and ask who in your organisation actually swears an affidavit.
The finance director in a debt recovery action. The HR lead whose witness statement goes into a Circuit Court file. The company secretary verifying pleadings. The engineer or actuary whose expert report lands in a construction dispute or a clinical negligence claim. None of them are solicitors. All of them are now signing a personal statement about their own use of AI tools.
Here is the organisational problem. Most Irish organisations of 50 to 2,000 people cannot answer, with evidence, whether a named employee used a generative AI tool to draft a specific document in a specific week. They have an AI policy. They do not have a record. The declaration asks for something a policy does not produce.
The gap widens where an assistant is switched on by default across the tenancy. A drafting tool that suggests a rephrase inside Word is doing precisely what the declaration says the deponent must not have done, invisibly. The signatory is certifying the absence of something they may never have noticed happening.
Privilege is the larger commercial exposure
Barniville flagged a second risk that got less coverage than the fake citations, and it is the one with money attached.
Putting case material into a generative AI system risks the loss of legal professional privilege over it. Where a tool runs on consumer terms, or the provider retains inputs for training or human review, the material has arguably left the privileged circle. That becomes a discovery question at the worst possible moment.
Once a matter is contested or reasonably anticipated, your litigation hold needs to cover AI tools alongside email and shared drives. That means knowing which assistants are enabled, what their terms say about retention, and which have been near the file. Most holds issued in Irish organisations this year did not mention AI at all.
What a court rule tells you about the direction of travel
Everyone has been watching the AI Act calendar. The AI Office of Ireland opened in August. High-risk obligations for stand-alone Annex III systems now sit at 2 December 2027, with the Commission's classification guidelines still in draft. Meanwhile a short court rule moved from issue to enforcement in five weeks, with no transition period and no harmonised standard to wait for.
That pattern will repeat. Bodies that already regulate a profession or a process act on their own timetable, using powers they already hold. The Workplace Relations Commission is a designated market surveillance authority for employment AI. The Central Bank supervises model risk under frameworks that predate the AI Act. Professional indemnity insurers will ask their own questions at renewal.
So organise the governance work around where the organisation already carries accountability. The AI Act phase-in calendar is a poor planning instrument here, because the sharpest consequences are arriving from elsewhere and arriving first.
What to do this quarter
Name the signatories. List everyone who swears affidavits, signs witness statements or produces expert reports for the organisation. Send them HC 142. Most have not heard of it.
Make the declaration provable. For anyone on that list, know which AI tools were available to them, which were enabled by default, and whether drafting assistance was active in the applications they used. A declaration you cannot evidence is a liability sitting under someone's signature.
Put AI tools inside the litigation hold template. Add the question to the standard hold notice now rather than drafting it under pressure in a live dispute.
Check the terms before privilege matters. Establish which deployed tools retain inputs, and on what basis, while that is still an administrative task.
The Acuity AI position
We diagnose before we prescribe, and this is a clean example of why. Almost every organisation we assess has an AI policy that would, on paper, have prevented the problem HC 142 is aimed at. Very few can produce evidence that it held. The declaration does not ask what your policy says. It asks what a named person did.
The work is unglamorous. Establish which AI tools are actually live across the estate, including the ones switched on by default inside existing licences. Identify the document types where an individual's signature carries personal exposure. Instrument those two against each other. It takes days rather than months, and it produces the record a court, a regulator or an insurer will eventually ask for.
Our AI governance work builds that record around observed usage rather than declared policy, and our board-level AI governance programme gives directors the questions that surface this exposure before a costs order does.
A court has now established what unverified AI output costs and who pays. Every other Irish body with supervisory powers has been watching.