Gulf courts have started to act on AI-invented citations. In December 2025 an ADGM judge ordered MIO Legal Consultants LLP to pay AED 282,508 in costs. In November 2025 the QFC Civil and Commercial Court said that citing an unchecked AI-provided case "ordinarily amounts to reckless conduct", and found contempt proved because this lawyer's conduct, including a misleading reply to the Registry, was intentional. Both courts put the duty on the lawyer who cites, whatever tool produced the text.
I build AI systems for firms, including the legal research assistant described at the end, and each judgment and rule here is linked so you can read it in full.
The ADGM costs order: Arabyads v Alam
The case is Arabyads Holding Limited v Gulrez Alam Marghoob Alam [2025] ADGMCFI 0032 (ADGMCFI-2025-165). Justice Paul Heath KC gave the costs judgment on 18 December 2025 (judgment). The order reads: "MIO Legal Consultants LLP shall pay the Claimant its costs in the sum of AED 282,508 which have been summarily assessed on the indemnity basis."
The judge found that the inaccurate citations "appear to have arisen out of reckless conduct as to accuracy rather than through a genuine accident", under rule 4(6) of the ADGM Courts Rules of Conduct. That rule says lawyers must not "knowingly or recklessly mislead the Courts by making incorrect or misleading statements of fact or law".
The judge also accepted that MIO "did not intend to deliberately mislead the Court or opposing counsel" (paragraph 52), and did not find MIO "guilty of conduct that was intended to mislead the Court" (paragraph 59). MIO's evidence was that "AI tools were not used for research purposes" (paragraph 53); the judge found that AI had been used (paragraph 56(c)).
Paragraph 41 sets out what a lawyer who researches with AI has to do: "lawyers using AI tools for research purposes should start from the premise that all authorities and/or articles on a particular topic that are revealed by AI research may not necessarily be accurately summarised in the response, or indeed may not exist. That puts the onus on a legal practitioner using AI for research purposes to verify the existence of authorities on which they wish to rely and to confirm that they stand for the propositions for which they are being offered".
On fault, paragraph 40 says it "lies not with the research programme used ... but with the person responsible for conducting the search". Paragraph 39 notes that sanctions "can take the form of reference to a regulatory authority or an award both to compensate and punish". In paragraph 46, the judge adopts the view of the English Divisional Court in Ayinde ([2025] EWHC 1383 (Admin)) that "every person providing legal services within ADGM ... must understand and comply with their professional and ethical obligations and their duties to the Court if using artificial intelligence", which calls for "practical and effective measures" from those "with individual leadership responsibilities".
The QFC contempt finding: Sheppard v Jillion
The case is Jonathan David Sheppard v Jillion LLC [2025] QIC (E) 3 (CTFIC0033/2025), decided by Lord Thomas of Cwmgiedd on 12 November 2025 in the QFC Civil and Commercial Court, whose judgments are cited as QIC (judgment page, sealed judgment). The order reads: "Contempt of court and breach of article 35.2 of the Rules and Procedures of the Civil and Commercial Court of the QFC proved." It adds: "There be no penal sanction."
Paragraph 23 sets the standard: "It is difficult to see how any lawyer could, by the second half of 2025 be under any doubt as to the necessity of checking that a case provided by an AI system is in fact a decision set out in the jurisprudence of the court. Proceeding therefore to refer to such a case without making such a check ordinarily amounts to reckless conduct".
Here the court went further. When the Registry asked for the citations, the lawyer said "he was unable to access judgment copies of the cases on the Court's website" (paragraph 23), and he later admitted that the cases "were provided by Google". The court found that conduct intentional, which contempt requires (paragraphs 22 and 24): "Contravention of article 35.2 and contempt require a requisite intention", and "the Court does not accept that the conduct was inadvertent. It amounted to intentional conduct". It treated his apology and the published judgment as a sufficient penalty, and the judgment does not name him (paragraph 25).
In paragraph 26, the court said that any unverified citation "will be considered a breach of the conduct required of all advocates before the Court and sanctioned accordingly. The sanctions will include the full identification of the lawyer or law firm and the consequent public disgrace."
The DIFC Gheorghe order: a warning, with the AI question left open
Some websites on AI risk describe a DIFC case more strongly than the order does. In Stelian Gheorghe v (1) BSA Ahmad Bin Hezeem & Associates LLP (2) Jimmy Haoula, CFI 045/2025, Justice Sapna Jhangiani gave an order with reasons on 28 October 2025 (DIFC Courts). Paragraph 10 records that the witness statement of a person "said to be the Claimant's lawyer" "contains a number of serious errors, including non-existent authorities and wrongly cited authorities. The Defendants contend that that evidence, along with the Claimant's claim form, may have been at least partly generated by artificial intelligence." The claimant was on record "as a litigant in person".
The Court did not decide the AI point. It observed "that errors of law have no place in witness evidence filed by lawyers before this Court." The order is a warning with no sanction for AI use, and the person said to be the claimant's lawyer was not on the DIFC Courts register.
What the DIFC and QFC guidance says about checking citations
The DIFC Courts' Practical Guidance Note No. 2 of 2023, in effect since 21 December 2023, is guidance only. It says: "Parties should not rely on AI-generated content without first verifying its accuracy. This should be done through using independent sources such as case law, statutes and credible legal commentary" (section 4.1). It points to contempt proceedings for false statements verified by a statement of truth (RDC 29.137), and notes that "The Courts have the power to reject any content generated through AI-systems under Rule 29.10 of the Rules of the DIFC Courts". Parties "should declare at the earliest possible opportunity if they have used or intend to use AI-generated content during any part of proceedings" (section 4.2). The mandatory Code of Conduct adds that practitioners "shall never knowingly or recklessly make any incorrect or misleading statement of fact or law to the Courts" (Part B(4)(i)).
The QFC Court's Practice Direction No. 1 of 2026, signed on 6 January 2026, makes it "a breach of duty to mislead the Court or submit materials containing fabricated or unverifiable legal citations" (paragraph 8). Paragraph 16 says: "All citations must be checked against primary sources or recognised legal databases." The direction also bars the use of AI tools to generate evidence (paragraph 18).
On disclosure, "Pre-emptive declaration of the use of AI is not required unless the Court specifically asks for it", but court users "must be prepared to identify portions of documents that used AI-generated content and explain how they verified the output. The Court may require an explanation or an affidavit" (paragraphs 24 and 26). The sanctions in paragraph 27 run from disregarding pleadings or evidence to "Adverse costs orders", "Referral to professional regulators" and "Contempt of court proceedings where appropriate". The Practice Guidance issued with it asks court users to "maintain clear records of how they have verified AI-generated content, and be prepared to explain their verification process if challenged" (paragraph 5(iv)).
In ADGM, Practice Direction 2 deals with AI only for document searches: a party that intends to use AI to search for documents "must inform all other parties of that intention" (paragraph 2.85).
In the UAE and Saudi Arabia, the professional rules already cover a false citation
The other Gulf courts I checked have issued no AI guidance for lawyers: the federal, Dubai and Abu Dhabi courts in the UAE, the Saudi courts, Qatar's onshore courts, and the courts of Bahrain, Oman and Kuwait. In the UAE and Saudi Arabia, the professional rules reach a false citation anyway. The UAE Code of Ethics says: "The Lawyer shall refrain from presenting false or misleading information to the court" (Article 13(4)(d)). Saudi Arabia's Rules of Professional Conduct for Lawyers forbid a lawyer to "rely on repealed statutory provisions, overturned judgments, or false data or data prohibited by law" (Rule 32, my translation from the Arabic; gazette notice).
When the UAE Minister of Justice launched the Guiding Principles for the Use of Artificial Intelligence Systems by Law Firms on 29 September 2026, he said: "professional responsibility cannot be delegated. The outputs generated by artificial intelligence systems must always be subject to careful human review, verification and oversight. Lawyers remain responsible for ensuring the accuracy and suitability of these outputs" (press release). I could not find the text of the ten principles when I checked on 8 October 2026. In Saudi Arabia, SDAIA's non-binding Generative AI Guidelines For Public say: "It is users' responsibility to verify the content generated by GenAI."
The lawyer answers for the citation, whatever tool produced it
Neither court blamed the software. In ADGM, the fault sat with the person who ran the search, and the QFC Court expected every lawyer to know by the second half of 2025 that a case from an AI system has to be checked. The ADGM judgment also addresses the people who lead the firm, and expects them to put measures in place. For a firm, the check belongs in its AI policy, under a named partner, and buying a different tool does not move the duty.
What a citation check in code looks like
I built a legal research assistant over one jurisdiction's legislation, and the case study describes how it handles citations. The model is a Gemini agent on Google's ADK, and it acts through a small set of tools: search the legislation, open a specific article, cite a law. Search runs on Vertex AI Search, and every quoted legal text is read from a Postgres database.
Code then checks each citation twice. The law the model cites must be one of the laws the search returned for that question; any citation to another law is dropped before display, so the model cannot cite a law the search did not surface. The article number must exist for that law in the database; if it does not, the answer keeps the law reference and drops the number.
When a search filtered to one jurisdiction finds nothing, it widens to the whole library and returns a "broadened" flag, and the model has to tell the user that the results are not a match for the jurisdiction asked about. The texts come only from official government legislation sources, each stored with a SHA-256 fingerprint, so when a source republishes, a new version is written and the old one is kept. Prompt rules separate what "the law states", which must be sourced, from what the model knows "generally", which is labelled as such.
Each answer carries clickable footnotes, and each footnote resolves to a real law with a source URL from an official portal. When the assistant cannot verify something, such as an article number or a jurisdiction outside its coverage, it says so.
Code covers the first half of the ADGM test: it can prove that a law or an article exists and show its text. Whether the authority stands for the proposition it is cited for is the lawyer's judgment, made by reading the source, and a footnote that opens the article makes that a one-click check.
That system works on legislation. The Gulf cases concern judgments, and the same checks apply once the firm holds the judgments it cites in a database, which fits the QFC direction's reference to "primary sources or recognised legal databases". For the "clear records" that the QFC guidance asks for, I would store the result of each check with the answer: which citations passed, and against which version of the text.
The citation check and its record belong in the firm's AI policy, which the AI policy generator on this site helps draft.