The Court Didn’t Build a Detector.
It Made You Declare Instead.

Ireland’s High Court now requires affidavits and witness statements to declare that generative AI did not generate or reshape their substantive evidence. Expert reports carry a qualified declaration and must disclose material GenAI use. Effective September 1, 2026, the requirements reach new court documents in cases already underway — not only newly filed proceedings.

By William Douglas Baker

Founding Editor

On July 29, 2026, Mr. Justice David Barniville, President of Ireland’s High Court, issued Practice Direction HC142, governing the use of generative AI in the documents the court reads. Paragraph 27 is the sentence that will touch the most people. Unless a judge directs otherwise, anyone swearing an affidavit or signing a witness statement must now open the document with a declaration that Generative Artificial Intelligence “has not been used for the purposes of generating the substantive content of this” witness statement or affidavit, “including by way or altering, embellishing, strengthening, diluting or rephrasing the evidence set out,” and that it “reflects my personal knowledge, recollection and evidence.” (“By way or altering” is the Practice Direction’s own wording, not a transcription slip introduced here.) What the court now requires, in the opening paragraphs of affidavits and witness statements, is a formal declaration that generative AI was not used where it counts. It does not prescribe a technical means of proving that declaration true — it places responsibility for its truth on the person making it.

The Practice Direction explains its own reasoning before it gets to the rule. Paragraph 10 lists the known risks of using generative AI in preparing court documents: it may produce “responses and material that appear plausible but are false or misleading,” generate “fictitious cases, legislative provisions, academic references or other authorities,” and manufacture “fabricated, manipulated or ‘deep fake’” material outright. Paragraph 12 puts the verification burden on the human: any material “generated wholly or partly with the assistance of” a GenAI tool must be “appropriately and independently verified” before it is used in proceedings. Paragraph 13 closes the obvious workaround: checking one AI’s output with a second AI “is not sufficient.”

The rule doesn’t treat every court document the same way. A witness or deponent supplies the prescribed declaration without routinely filing a tool log or drafting history — though the court may direct the person to disclose whether GenAI was used and, if so, the nature and extent of that use. An expert report is held differently: Paragraph 29 requires its own qualified declaration, and Paragraphs 30 and 31 additionally require the expert to disclose the nature and extent of any material GenAI use in drafting, analysing, interpreting, or generating substantive content, so the court and the parties can assess the reliability, independence, and evidential weight of the expert’s opinion. Two people can sit in the same courtroom, submit comparably weighty documents, and be governed by two different document-specific declaration and disclosure regimes. HC142 prescribes no routine technical test for determining whether GenAI was used; instead, it combines declarations, human source verification, disclosure when directed, adversarial challenge, correction duties, and sanctions. The provenance question itself is answered by no instrument in that system — it is answered by a person, on the record, under consequences.

What the declaration doesn’t check

The infrastructure already waiting beyond the declaration

HC142 did not arrive alone. Paragraphs 38 through 41 apply the direction to litigants in person — unrepresented parties running their own cases — and point them to guidance the Courts Service had already published for people using AI in court proceedings: Using AI to prepare documents for court, written for anyone using these tools in a proceeding. It complicates the clean version of this story. The institution didn’t only reach for the oldest instrument in the room; it already had in place the one piece of public infrastructure a plain declaration can’t supply on its own — instructions, in advance, to verify sources, be ready to explain the checking process, correct inaccuracies immediately, and accept responsibility for every filing. The declaration carries the visible attestation burden inside the document. The guidance carries much of the verification instruction outside it.

What this doesn’t fix

This record has tracked the same institutional reflex from four different angles already. Edition No. 031 documented the European Union’s answer: force the machine to disclose what it is. Edition No. 034 documented Google DeepMind’s answer: watermark the machine’s own output. Edition No. 041 documented the ceiling on that answer: a watermark can prove a file passed through a processing pipeline, not who is accountable for the words inside it. Edition No. 042 documented the failure mode underneath all of it: the same verification tool built to check a watermark can be run in reverse to strip it, for about four cents a pass. HC142 is where the arc lands. The court did not solve machine provenance; it converted the uncertainty into personal accountability. The rule contains no detector and cites no watermark — what it installs is the oldest instrument in the room: a person’s word, in a declaration the court can question, test, and enforce, but for which it prescribes no technical instrument.

In the clinical vocabulary of this record, HC142 exposes a form of Trust Transfer Failure: the document alone no longer carries enough provenance for the institution to accept its substantive evidence as the person’s own without an added human declaration. The parallel to Answer Engine Authority™ is structural, not evidentiary. AEA™ organizes the identity, corroboration, and authority signals available to an inference system; it does not prove who wrote a particular document. The court can compel an attestation after the fact. An answer engine cannot — which is why authority must be made legible before the recommendation is requested.