What Happened
On June 24, 2026, the Hessian Higher Administrative Court (Hessischer Verwaltungsgerichtshof, VGH Hessen) issued a referral to the European Court of Justice — a Vorabentscheidungsersuchen — asking whether Ukrainian nationals lose their claim to EU temporary protection if they spent more than two years legally, but without a residence title, in a third country (in this case, Georgia) before entering the bloc. The case, filed under docket number 3 A 2624/25 by the court's Third Senate, will shape how all 27 EU member states treat a meaningful slice of the Ukrainian displacement population — a serious, continent-wide legal question (VGH Hessen press release).
What drew wider attention wasn't the referral itself but what reporters found in its footnotes seven weeks later: five source links in the court's reasoning carried the suffix "utm_source=chatgpt.com" — the tracking parameter ChatGPT automatically appends to URLs it surfaces in a chat response. The judges had used ChatGPT to research Georgian legal and administrative sources, and the citation trail showed it plainly (it-boltwise.de).
The Errors
At least two of the five citations didn't hold up. One was meant to support a description of a 2026 Georgian government decree but linked to an article published in 2025 — before the decree it was supposedly documenting existed. Another was cited for a 2015 regulation but pointed instead to an unrelated 2025 amendment provision. Both are textbook large-language-model failure modes: plausible-looking source attributions that don't actually match the underlying claim (it-boltwise.de).
Equally notable is how long it took anyone to notice. The court publicized the ruling via press release the same day it was issued, June 24. The dead and mismatched links weren't reported publicly until roughly 47 days later. No clerk, no opposing counsel, no journalist clicked through in the interim (borncity.com).
The court's response has been measured rather than defensive: it says AI use for research is fundamentally permissible — especially useful for locating and translating hard-to-access foreign legal material, like Georgian statutes few German court staff read fluently — but that AI output must be verified before it goes into a ruling, and it maintains the flawed citations weren't outcome-determinative because the decision rested on a broader evidentiary base than ChatGPT alone (heise.de).
The Case for Alarm
The strongest objection isn't really about this one referral's outcome — it's about what citations are for. A court's reasoning is only as trustworthy as its sourcing, because litigants, appellate judges, and — in this instance — 26 other EU legal systems are meant to be able to retrace the court's logic to the underlying law. A hallucinated date or a dead link isn't cosmetic; it's a broken link in the chain of legal accountability, and it's the kind of error traditional judicial review is bad at catching precisely because reviewers trust a citation looks real and move on. That a national court could publish an EU-referral ruling with unverifiable footnotes and have it stand undetected for nearly seven weeks is a legitimate worry about institutional quality control, not just tooling.
Why the Fix Isn't Prohibition
But the lesson here is a citation-discipline failure, not an AI failure that only a ban on judicial AI use could solve. Nothing in this case suggests ChatGPT's assistance changed the outcome, and the underlying legal question — whether prior refuge in Georgia forecloses EU temporary protection — was neither manufactured nor distorted by AI; the court still had to interpret the EU's 2022 Mass Influx framework on its own authority. The actual defect is mundane: nobody cross-checked five links before publication, in a system that, as of this ruling, still has no uniform mandatory training on AI verification for judges, despite Germany's judiciary formally recognizing the gap.
Germany's E-Justice-Rat — the federal-state body coordinating judicial digitization — has been explicit that AI's role is to support court staff while decision-making authority stays with judges, and has been developing AI competency and training measures for the courts precisely because tools like this are already in use (Justizportal / E-Justice-Rat). That's the right instinct, and this case is an argument for accelerating it, not reversing it. A rule requiring judges to cite the primary legal text an AI tool points to — not the AI's paraphrase, and not a link the AI hallucinated — combined with a disclosure norm (flag where AI assisted research) and a pre-publication verification checklist, would have caught both errors before they ever reached print. That's proportionate governance: it preserves the genuine efficiency gain of AI-assisted research into obscure foreign-language sources, while closing the specific gap — unverified citations — that actually caused the problem. An outright ban on judicial AI research, which some German legal commentators have floated in response, would sacrifice a real productivity tool to fix a quality-control process that needed fixing anyway.