The law is increasingly requiring verification.
The missing layer is evidence that it happened.

What changed on 1 September

On 1 September 2026, Practice Direction HC 142 of the High Court of Ireland came into operation. It was signed by the President of the High Court, Mr Justice David Barniville, on 29 July 2026, and it applies to pleadings, submissions, affidavits, witness statements and expert reports in civil proceedings, whether the proceedings began before or after that date.

Courts have warned about AI for three years. HC 142 is different in kind: it writes the duty down in advance, names who carries it, and prices the failure.

What the duty actually is

• Para 12Any person responsible for a court document must ensure that material generated wholly or partly with generative AI has been appropriately and independently verified before it is used.
• Para 13“It is not sufficient to use the same or other GenAI Tools to confirm accuracy.” The check has to be a person’s.
• Para 18Verification must be under human control and supervision: fact-check, proofread, edit to the case, and confirm that every authority exists and stands for the proposition attributed to it.
• Para 21When the Court directs it, a party must disclose whether generative AI was used and the nature and extent of that use. All persons are expected to be in a position to answer.
• Paras 35, 36Adverse costs. Rejection of the document. Strike-out of a claim or defence. For a practitioner, referral to the Legal Services Regulatory Authority and costs orders against them personally.

Independent verification. Human control. The ability to explain the AI use. Three duties, each of which assumes the check happened.

What one unverified output costs

Two days after HC 142 came into operation, the Supreme Court of India answered the question from the other direction. In Gadiya v. Union of India, a customs officer had imposed a penalty of Rs 425,27,99,100 for mis-declaring a consignment of natural diamonds as lab-grown. The appellant said several judgments and articles cited in the order had been generated with AI. The Court checked them itself.

“Upon verifying such judgments and articles, it does appear that the second respondent has relied upon case laws that are either non-existent or have fake citations. A further investigation has also revealed that some of the case laws relied upon by the second respondent, which do exist, do not lay down the ratio deduced therefrom and it appears to be a hallucination of AI.”

The Court did not reach the merits. It applied its own rule from Pooja Ramesh Singh v. Jammu and Kashmir Bank, decided 2 July 2026, which it quoted in full:

“It is a misconduct on the part of an advocate to cite such judgments without verification. Equally, it is a serious lapse if a judge relies on such a fake or hallucinated artificial intelligence-generated material as precedents in support of the determination. We have no hesitation in declaring that such a decision is no decision in the eyes of the law, irrespective of whether such material had a direct or indirect bearing on the decision-making.”

The penalty order and the High Court order upholding it were set aside. The matter went back to a different officer. The Court left it to the appointing authority to act against the author of the order. No showing that the fabricated authorities affected the outcome was required, because the rule is that even an iota of hallucinated material entering the decision-making process is enough.

That is the price of one unverified output, stated by a supreme court: not a fine on the person who used the tool, but the voiding of the decision itself.

The gap between the two

Ireland says the check must happen and must be a person’s. India says a decision that rested on an unchecked output does not exist. Neither instrument says how anyone would know, on the day it matters, whether the check took place.

A declaration is a sentence a person signs. A verification is work a person did.

The declaration identifies the signer. It does not itself evidence which AI was used, what was reviewed, against which sources, or when. It cannot be checked by the other side or the court without trusting the person who made it. The sanctions record is a list of matters in which that trust was tested and failed: in Booker v. Kroger, counsel told the court he had not used AI, then acknowledged in his own response that he had, and the sanction was doubled for the denial.

Verification is a process. Evidence that the process occurred is a separate object, and the law has begun to require the first without saying anything about the second.

The question that decides it

Who authorized this action, and can an outsider check it?

For a filing, the action is the verification. Who did it, on what exact document, at what time, and can the other side or the court confirm that without taking the signer’s word for it. The same question applies to every consequential act of an AI system, inside the courtroom or far from it. A duty to verify, on its own, produces a promise. A record that can be independently checked produces evidence.

What the record shows

Regulayer creates the verifiable record behind that answer. A signed, tamper-evident record shows which AI was used, what a person reviewed, and when, sealed to the exact document that was filed, and checkable free by anyone holding the file, without an account and without contacting us.

HC 142 also names two risks that bear on where that record should be made. Paragraph 10 lists confidentiality and legal privilege among the known risks of generative AI tools. Paragraph 32 says that where material is subject to a court order or an undertaking restricting its use, including discovery material, generative AI tools should not be employed on it where that would involve or risk disclosure contrary to the order. The Regulayer record is designed so that creating and verifying it does not require the underlying document to be sent to Regulayer. The work stays where it was done.

Not your court

Ireland’s High Court is not your court. India’s Supreme Court may not govern your organisation. That is not the point. Within two days, one court wrote down a human-verification duty for AI-assisted work, and another set aside a consequential decision after fabricated AI material entered it.

The underlying problem travels. A claim gets approved. A batch gets released. Money moves. A filing is made. Who authorized the action? What did a person actually review? And when something goes wrong, can an outsider check the record?

That question does not stop at a border. The courts reached it first because their failures are public and their remedies are written down. Why the control point is moving closer to the action, and Regulayer for institutions deploying AI.

Sources

• Courts Service of Ireland, Practice Direction HC 142, full text
• Supreme Court of India, Gadiya v. Union of India, 2026 INSC 947, order of 2 September 2026 (copy bearing the Court’s neutral citation)
• Supreme Court of India, Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd., 2026 INSC 668, judgment of 2 July 2026 (copy bearing the Court’s neutral citation)
• Supreme Court of India

Related

• Ireland, Practice Direction HC 142
• Gadiya v. Union of India · India, the Supreme Court’s rule and Kerala’s policy
• The sanctions record · Privilege and AI · When the law changes

Checked 3 September 2026 against the full text of HC 142 and the two Supreme Court of India judgments named above. Information, not legal advice, and not advice on the law of any jurisdiction.

Common questions

What does Ireland’s Practice Direction HC 142 require?

From 1 September 2026, any person responsible for a court document in Irish High Court civil proceedings must ensure that material generated wholly or partly with generative AI has been appropriately and independently verified before it is used. Paragraph 13 says it is not sufficient to use the same or another generative AI tool to confirm accuracy. Paragraph 18 says the verification must be under human control and supervision. Failure carries adverse costs, rejection of the document, strike-out, and for practitioners referral to the Legal Services Regulatory Authority and personal costs orders.

What did India’s Supreme Court decide about precedents produced by AI?

In Pooja Ramesh Singh v. Jammu and Kashmir Bank, decided 2 July 2026, the Supreme Court declared zero tolerance for producing, citing or using precedents generated by AI without verification, held that citing such material without verification is misconduct by an advocate, and held that a decision resting on fake or hallucinated material is no decision in the eyes of the law, irrespective of whether the material bore on the outcome. On 2 September 2026, in Gadiya v. Union of India, it applied that rule to set aside a customs penalty of Rs 425,27,99,100 because the officer’s order relied on non-existent or misattributed authorities.

Why is a declaration that the work was checked not enough?

A declaration is a sentence a person signs. A verification is work a person did. The declaration identifies the signer. It does not itself evidence which AI was used, what was reviewed, against which sources, or when, and it cannot be checked by the other side or the court without trusting the person who made it. That distance is where the sanctions record lives.