--- title: "The $77,000 Citation" arabicTitle: "استشهاد بـ ٧٧٬٠٠٠ دولار" arabicDescription: "في فبراير ٢٠٢٦، أمرت محكمة في أبوظبي محامين بدفع أكثر من ٢٨٢ ألف درهم بسبب استشهادهم بقضايا اخترعها الذكاء الاصطناعي. هذا ليس حادثاً معزولاً، بل جزء من أزمة متنامية في المحاكم حول العالم." arabicKeywords: - "هلوسة الذكاء الاصطناعي" - "استشهادات قانونية مختلقة" - "محكمة أبوظبي" - "عقوبات المحاكم" - "الذكاء الاصطناعي والقانون" - "التأصيل القانوني" - "الإمارات العربية المتحدة" - "AI Hallucination Sanctions" date: "2026-08-20" description: "An Abu Dhabi court fined a law firm $77,000 for citing AI-fabricated cases. It's one entry in a database of over 1,600 similar rulings worldwide, and it's exactly the failure Dike exists to prevent." author: "Dike Engineering" readTime: "7 min read" heroImage: "/images/artwork/grounded-authority-core.png" tags: ["AI Safety", "MENA", "Grounding"] --- In February 2026, the ADGM Court of First Instance in Abu Dhabi ordered a law firm to pay 282,508 dirhams, about $77,000, in wasted costs after its defense submission in *Arabyads Holding Limited v Gulrez Alam Marghoob Alam* leaned on legal authorities that didn't exist.[1][2] The citations had been generated by an AI tool and filed without anyone checking whether the cases behind them were real. They weren't. The court called the conduct reckless, and made it the first UAE ruling on wasted costs arising specifically from AI misuse. It's not an isolated embarrassment, either. Legal researcher Damien Charlotin has been tracking these incidents since AI hallucinations first started showing up in court filings, and his public database is past 1,600 court proceedings worldwide where fabricated citations reached a judge, growing by roughly eight new documented cases a day as of mid-2026.[3] Most of those are in the United States, but the Abu Dhabi ruling is the one MENA legal teams shouldn't be able to shrug off as somebody else's problem. It's already here. > The fake cases weren't caught by a smarter model. They were caught by a judge doing the job the AI tool should have done in the first place. ## This isn't a fringe problem anymore The scale is what's changed. An Omaha attorney filed a divorce appeal brief in which 57 of 63 citations turned out to be defective, a mix of entirely fabricated cases and citations that simply didn't say what the brief claimed they said. The Nebraska Supreme Court handed down a temporary suspension in April 2026, pending a full disciplinary hearing, calling it a grave error of judgment.[4] A month earlier, a Sixth Circuit panel sanctioned two attorneys in *Whiting v. City of Athens* over briefs containing more than two dozen fabricated citations: $15,000 each to the court registry, the opposing side's full appellate fees reimbursed, double costs, and a disciplinary referral.[5][6] The panel's opinion was blunt about where the responsibility sits: no filing should contain a citation the lawyer hasn't personally read and verified, regardless of whether AI wrote it. Courts have responded the way institutions usually respond to a new failure mode: with process. More than 300 US federal judges have adopted standing orders addressing generative AI in filings.[7] Many jurisdictions now require a signed AI disclosure certificate, a lawyer swearing under their own name that they manually verified every AI-generated citation before filing it. It works, sort of, in that it shifts liability back onto a human who's supposed to catch the fake before it's filed. But it's also an admission that the tools producing these briefs don't check themselves. Verification only became a courtroom compliance step because the AI systems generating the citations have no internal concept of "this one might not be real." ## Why MENA legal AI is even more exposed The mechanism behind a hallucinated citation is the same everywhere: a language model doesn't have a lookup table of case law sitting in its weights, it has a compressed statistical sense of what a citation *looks like*, and when the real one isn't well represented in training data, it generates something with the right shape instead. What's different in the region is how much more often that gap gets triggered. Arabic primary law doesn't live on the kind of clean, indexed pages English-language law enjoys. It's scattered across bare-IP parliament servers, university file dumps, and loose PDFs that were never meant to be a searchable corpus. A general-purpose model trained mostly on English legal text has thin, patchy coverage of Saudi, Emirati, and Egyptian law to begin with, which means more guessing, not less, on exactly the jurisdictions where a Dubai or Cairo lawyer is most likely to be asking. The failure mode doesn't even need a fabricated case name to be dangerous. A model that's thin on real Gulf or Egyptian statutory text but fluent in Arabic can produce an answer that's confidently, grammatically, entirely wrong about which country's law applies, and nothing about the sentence itself gives that away. A lawyer moving quickly, in a language where the model sounds equally sure of itself whether it's right or wrong, has a lot less to catch that on than the Abu Dhabi firm had when it was staring at case citations that turned out not to exist. ## What "verified" actually has to mean The standing orders and disclosure certificates treat verification as something a human bolts on after the model finishes generating. We think that's backwards. If a citation can't be checked against something real before it reaches the user, the check happens too late to matter, because by then it's already sitting in a brief that's about to get filed. Every citation Dike returns resolves against the actual corpus at query time, not against whatever a model recalls. Our reasoning endpoint is instructed to cite only with structured reference tokens, and once it generates an answer, each of those tokens gets parsed and checked against the documents that were actually retrieved for that query. If a reference doesn't resolve, it's stripped out and logged, not softened with a caveat, just removed. The API also reports how many citations came back verified versus how many got stripped, so "the model didn't have a good answer" stays visibly different from a real one. We keep the official Arabic text as the source of truth rather than reasoning over a translation, because a citation's exact wording is what a court actually checks, and a translated paraphrase quietly discards that. English sources sit alongside it, anchored to the same underlying document, so an agent working in either language is grounded in the same source. None of this makes it impossible to file something wrong. You can still ask a bad question, or misread a correct answer. What it closes is the specific gap that cost the Abu Dhabi firm 282,508 dirhams, the gap between a model making up something that looks like a citation and a system that actually went and found one. Whether you catch that gap before you file or after opposing counsel does is, as it turns out, a $77,000 question. --- ## Sources
  1. Morgan Lewis, February 2026. ADGM Court's Judgment Serves as a Warning Regarding Proper Use of AI
  2. Herbert Smith Freehills Kramer. AI Hallucinations: ADGM Court Takes Firm Stance Against Misuse of AI in Drafting Pleadings
  3. Damien Charlotin. AI Hallucination Cases Database, a public tracker of court decisions worldwide involving AI-hallucinated material
  4. WOWT, April 2026. Nebraska Supreme Court Suspends Omaha Attorney Over AI Use
  5. Exterro. Sixth Circuit Imposes Steep Sanctions for "Hallucinated" Citations and Record Misrepresentations, on Whiting v. City of Athens
  6. Sixth Circuit Appellate Blog. Sixth Circuit Sanctions Attorneys for Fake Citations
  7. Ropes & Gray. Standing Orders, Local Rules, and Decisions on the Use of AI, a tracker of federal court AI disclosure orders