analysisAI Governance SRA & UK Regulation EU AI Act AI in Courts

Summer 2026 in legal AI regulation: SRA warning notice, mandatory Irish declaration and a deferred EU Act

Summer 2026 reviewed: the SRA's AI warning notice, Ireland's mandatory AI declaration in civil filings and the EU's deferred high-risk AI Act duties.

Edited and verified by Cognesio LLP2026-09-03

Researched with AI assistance · sources verified by Cognesio LLP · How this was made ↓

In England and Wales, Ireland and the EU, the period from 1 July to 3 September 2026 produced three changes to the rules on generative AI in legal practice. The Solicitors Regulation Authority issued a warning notice on misuse of AI on 17 August; Ireland’s Practice Direction HC 142 came into operation on 1 September; and Regulation (EU) 2026/1744 deferred the AI Act’s high-risk obligations to 2027 and 2028.

Elsewhere the summer produced drafts and interim findings: a draft Indian court regulation with a disclosure declaration, South African guidelines no regulator has adopted, and a US federal evidence rule still under consideration. As of 3 September 2026 the Regulation Tracker records 84 official documents on AI in legal practice. The table lists the documents that changed this summer, plus three June items that preceded it.

Document	Body	Date	What it requires	Status
Misuse of AI: warning notice	Solicitors Regulation Authority (England and Wales)	17 Aug 2026	Solicitors accountable for AI output; cited authorities genuine and verifiable; client data secured and not used for model training without authority; COLPs to secure supervision	In force
Practice Direction HC 142	High Court of Ireland	Signed 29 Jul 2026; in operation 1 Sep 2026	Human verification of AI content in civil court documents; no-substantive-use declarations in affidavits, witness statements and expert reports	In force
Regulation (EU) 2026/1744 (Digital Omnibus on AI)	European Parliament and Council	Adopted 8 Jul; OJ 24 Jul; in force 27 Jul 2026	Annex III high-risk obligations apply from 2 Dec 2027; Annex I from 2 Aug 2028; Article 50 date unchanged	In force
Guidelines on transparency obligations (Article 50)	European Commission	Page last updated 6 Aug 2026	Deployers to inform people exposed to deepfakes and to unreviewed AI text on matters of public interest	Published
Revised Ethics Guidelines on Generative AI	Thaldar, Mbatha, Botes and Esselaar, in De Rebus (South Africa)	July 2026 (DR 11)	Citations confirmed from a primary source; quotations verified verbatim; free tools distinguished from paid	Authors’ proposal; no regulator adoption found
Draft Regulations for Use of AI in Courts, 2026	Supreme Court of India, AI Committee	Notice 3 Jun 2026; comments extended to 15 Jul 2026	Declaration disclosing AI assistance in any pleading or evidence; AI barred from bail, credibility and risk scoring	Draft
Practice Book § 4-9	Connecticut Judicial Branch	Effective 23 Jun 2026	Independent verification of all AI-produced citations, authorities and evidence; filer certification; sanctions up to nonsuit or default	In force
Update on consultation findings, AI in court documents	Civil Justice Council (England and Wales)	30 Jun 2026	Interim finding: no AI-specific formal requirement for professional drafting; final report later in 2026	Interim

What did the SRA’s AI warning notice change?

The SRA’s warning notice on misuse of AI, published on 17 August 2026, creates no new rule. It restates that solicitors are accountable for AI output, that cited authorities must be genuine and verifiable, and that compliance officers must secure supervision of AI use.

The regulator’s press release of the same day gives the number behind the notice: 42 reports of potential AI misuse between July 2025 and July 2026. The press release refers to ongoing investigations; the notice says the SRA has identified a number of cases. Self-referral is now part of the pattern on the UK incident record: in Cork v Smith, Pinsent Masons self-referred to the SRA after the court queried its letters.

The operative sentence is short. “AI has no separate legal personality”, the notice says, and solicitors remain accountable for their work whatever tool prepared it. On the duty to the court, named authorities must be genuine, relevant, carry a verifiable citation and advance the argument being made. On confidentiality, client data must stay in a secure environment, must not train AI models unless explicitly authorised, and must not be kept longer than necessary.

The compliance officer for legal practice gets a named duty: all reasonable steps to ensure compliance with the regulatory arrangements on supervision. The notice cross-refers to the SRA’s effective-supervision guidance, which gained AI sections in June 2026. The closing paragraph of its Standards and Regulations section is what separates a warning notice from guidance: “If you fail to have proper regard to this warning notice, you are at risk of disciplinary action.”

Aileen Armstrong, the SRA’s executive director for strategy and policy, framed the notice as continuity: “the use of AI does not change the professional standards expected of solicitors and law firms.” The version history of UK guidance shows the same standards restated, each time in a heavier document, since the 2022 compliance tips.

What does Ireland’s Practice Direction HC 142 require from 1 September?

Practice Direction HC 142, signed by the President of the High Court, Mr Justice David Barniville, on 29 July 2026 and in operation from 1 September, applies to all documents in civil proceedings. It requires human verification of generative-AI content and a declaration in every affidavit, witness statement and expert report that generative AI was not used to produce the substantive content.

The scope clause, paragraph 3, covers pleadings, submissions, affidavits, witness statements and expert reports, and binds all persons involved in proceedings before the court. Paragraph 13 closes the loophole most verification rules leave open: “It is not sufficient to use the same or other GenAI Tools to confirm that information is accurate.” Paragraph 18(3) requires court users to check that every reference to case law, legislation, textbooks or articles exists and stands for the proposition attributed to it.

The declaration is the novel element. Paragraph 27 prescribes the wording for affidavits and witness statements: “I declare that Generative Artificial Intelligence has not been used for the purposes of generating the substantive content of this [witness statement/affidavit]…”. Paragraph 29 gives expert reports a variant that permits stated exceptions.

Paragraphs 35 to 37 list the consequences. Parties may face adverse costs orders, orders disregarding or rejecting court documents in whole or in part, or the striking out of a claim or defence. Legal practitioners may be referred to the Legal Services Regulatory Authority or their professional body. Ireland now has a court rule that England and Wales, on the Civil Justice Council’s interim reading, does not need. The Irish regulation page and Irish incident record carry the document and the cases that preceded it.

Where does the EU AI Act stand after Regulation 2026/1744?

Regulation (EU) 2026/1744, the Digital Omnibus on AI, was adopted on 8 July 2026, published in the Official Journal on 24 July and entered into force on 27 July. It moves the application date for Annex III high-risk systems to 2 December 2027 and for Annex I product-embedded systems to 2 August 2028. Article 50 transparency duties applied from 2 August 2026 as planned.

The deferral matters to legal practice because Annex III point 8 lists AI systems intended to assist a judicial authority in researching and interpreting facts and the law. Under the original text those obligations would have applied from 2 August 2026. Recital 39 of the amending regulation clarifies the scope of the existing grace period in Article 111(2) for high-risk systems already on the market before the new dates.

What did not move is as important. The amending regulation leaves the Article 50 application date alone, adding only a four-month transitional period, in a new Article 111(4) explained in recital 38, for providers of generative systems already on the market before 2 August 2026 to meet the marking obligation in Article 50(2) by 2 December 2026. Chapter V on general-purpose AI models keeps its original timeline.

The Commission’s guidelines page on Article 50, last updated on 6 August 2026, states that deployers must inform people when they are exposed to deepfakes or to text on matters of public interest published without human review or editorial control. The Code of Practice on AI-generated content, finalised on 10 June 2026, had about 190 signatories by the end of July. Whether any of this reaches a UK firm is the question the extraterritoriality explainer answers; the EU regulation page holds the documents.

What moved in India, South Africa and the US states?

Outside Europe the summer produced drafts rather than rules. India’s Supreme Court AI Committee closed comments on its draft Regulations for Use of AI in Courts, 2026 on 15 July, after a notice of 16 June extended the original 20 June deadline; four South African authors published revised generative-AI ethics guidelines in the July 2026 De Rebus, described by their authors as a contribution rather than a finished pronouncement; and Connecticut’s Practice Book § 4-9 verification rule took effect on 23 June.

The Indian draft is the widest in scope. Regulation 2 applies it to the Supreme Court, the High Courts and every court, tribunal and statutory commission performing adjudicatory functions in India. Regulation 43(3) requires that where a party or legal representative uses an AI tool to prepare any document, pleading or evidence, the AI-assisted character is disclosed at submission by a declaration in a prescribed format. Regulation 43(6) makes the submitter fully responsible for fabricated or inaccurate AI output, with no defence from its AI character. Regulation 20 bars AI from risk scoring, bail eligibility and witness credibility. SafeLegalAI has not located a final, notified version as of 3 September 2026.

The South African text takes the opposite position on disclosure. Its authors, Donrich Thaldar, Siphesihle Mbatha, Marietjie Botes and Paul Esselaar, write that the ethical response to AI is “one of focused responsibility rather than of proactive disclosure.” The verification standard is the strictest sentence of the summer: “Every citation relied on should be confirmed from a primary source; every quotation verified verbatim…” The guidelines distinguish free services that retain inputs for training from paid services with non-training guarantees. SafeLegalAI has found no regulator that has adopted them as of 3 September 2026.

Connecticut’s rule is the shortest. Section 4-9 provides that any person who uses generative AI to create or edit a filed document “shall independently verify all citations, legal authorities or evidence produced by generative AI”, and that failure may draw sanctions “including, without limitation, the entry of a nonsuit or default judgment.” Filing a document is itself a representation of good-faith, diligent efforts to comply. Three verification rules, three drafting choices: Ireland requires a declaration, Connecticut a certification by conduct, South Africa neither.

What is still pending?

Four items remain open as of 3 September 2026. The Civil Justice Council’s final report on AI in court documents is due later in 2026; a possible Federal Rule of Evidence 707 on AI-generated evidence remains under consideration in the United States; the European Commission’s high-risk classification guidelines are still the 19 May draft; and California’s proposed AI amendments to its Rules of Professional Conduct have not been adopted, as far as SafeLegalAI can find.

The CJC’s 30 June update reported a strong consensus that “no additional formal requirements relating specifically to AI are currently necessary” for professional drafting. Its principal open question is witness statements, where AI tools may “reshape, embellish or otherwise influence evidence in ways that may not be immediately apparent.” Publication of the final report is “anticipated later in 2026”.

In the United States, the rules report to the Judicial Conference for September 2026 records that the Advisory Committee on Evidence Rules, meeting on 7 May 2026, discussed “continued consideration of a possible New Rule 707 to address evidence that is the product of artificial intelligence” alongside a possible Rule 901(c) on deepfake evidence. Neither appears among the amendments recommended for approval. That is a status, not a date.

California’s proposal would amend Rules 1.1, 1.4, 1.6, 3.3, 5.1 and 5.3, prompted by a California Supreme Court letter of 22 August 2025. Public comment closed on 4 May 2026. The State Bar’s page scheduled the related changes to its practical guidance, not the rule amendments themselves, for the Board of Trustees in May 2026, and nothing on it records an adopted rule.

The open question is which way the two common-law neighbours settle. Ireland has decided that a signed declaration in every affidavit is the price of AI in litigation. The CJC’s interim view is that existing professional duties suffice. Its final report will be the first document to answer HC 142 directly.