AI Hallucination Cases in Europe: UK, Germany, Austria, Switzerland and the Netherlands

The solicitor who filed Horleston v SSHD [2007] EWCA Civ 654 did not get it from a chatbot. He typed a question into Google, and Google’s “AI Mode” answered with a Court of Appeal authority, complete with a panel of judges. The case does not exist; the citation belongs to an unrelated equal-pay decision. When the Upper Tribunal rephrased the question itself, the AI produced a different bench for the same non-existent case. That is [2026] UKUT 81 (IAC), promulgated on 17 November 2025, and of all the AI hallucination cases in Europe it is the one to read first: the risk has left the chatbot window.

A wasted-costs order in London, an Ermahnung in Berlin and a nullity appeal thrown out unread in Vienna are different consequences for the same duty, which an English court form now spells out in three words: exists, locatable, supports.

Why Europe looks different: 69 UK, 11 German, 2 Austrian decisions

Damien Charlotin’s database stood at 2,039 decisions on 12 September 2026: USA 1,396, Canada 217, Australia 110, UK 69, Italy 15, France 13, Germany 11, Austria 2. Charlotin’s own explanation is access, not virtue: PACER and CourtListener make American decisions searchable, while European anonymisation hides cases. It lists only decisions where a court found or implied reliance on hallucinated material, so every count is a floor.

The second difference is enforcement style. Markus Hartung of the DAV’s ethics committee, writing in Anwaltsblatt in September 2025, contrasts England’s Hamid jurisdiction with Germany, where “wir setzen auf Eigenverantwortung”: “Spezielle Berufspflichten gibt es jedenfalls noch nicht … Verletzungen dieser Pflichten sind bislang folgenlos geblieben.” His diagnosis: AI practises “Love Scamming”, “sprachliche Verführung als literarische Heiratsschwindelei”. Europe records fewer cases and fines less than the US sanctions timeline shows; it does not verify better, and the Upper Tribunal reported “a considerable increase in the latter half of 2025 in the citation of fictitious authorities”.

England and Wales: Ayinde and Al-Haroun in the Divisional Court

Ayinde v Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin) (6 June 2025) is the reference judgement.

In Ayinde, a barrister’s grounds cited five cases that do not exist. Haringey Law Centre’s email called the fakes “cosmetic errors”. Ritchie J ([2025] EWHC 1040 (Admin)) ordered the barrister and the law centre to pay £2,000 each in wasted costs: “They should have been shocked when they were told that the citations did not exist.” The Divisional Court found the threshold for contempt met, declined to initiate proceedings, and referred the barrister to the Bar Standards Board and the solicitor to the SRA. In Al-Haroun, an £89.4m claim, 18 of 45 cited authorities did not exist; the lay client had generated them with public AI tools, and the court found it “extraordinary that the lawyer was relying on the client for the accuracy of their legal research, rather than the other way around” (para 79).

Freely available generative artificial intelligence tools, trained on a large language model such as ChatGPT are not capable of conducting reliable legal research. — Dame Victoria Sharp P, Ayinde, para 6

Paragraph 82 is the line for managing partners: “promulgating such guidance on its own is insufficient.”

The Upper Tribunal: Google AI Mode, “Horleston” and the statement of truth

UK and R (Munir) v SSHD [2026] UKUT 81 (IAC) joined two immigration matters. In the first, the solicitor had “inadvertently used the ‘AI Mode’ of a Google search” and filed the non-existent Horleston. In the second, a solicitor blamed a part-time trainee. Both are under SRA investigation; headnote 2 made supervision the point: supervisors “must ensure that fee-earners under their supervision are aware of the dangers of using non-specialist Artificial Intelligence (AI) for legal research and drafting”.

The first solicitor had also pasted client emails and Home Office decision letters into ChatGPT to “improve” them. Paragraph 21, later quoted verbatim by the SRA:

To put client letters and decision letters from the Home Office into an open source AI tool, such as ChatGPT, is to place this information on the internet in the public domain, and thus to breach client confidentiality and waive legal privilege. — Upper Tribunal, [2026] UKUT 81 (IAC), para 21

He was also referred to the Information Commissioner (which tools count as “open” is covered in is ChatGPT confidential for lawyers).

The judicial review claim form now requires the legal representative’s statement of truth to confirm that every authority “(a) exists; (b) may be located using the citation provided; and © supports the proposition of law for which it is cited”. A model can build the checklist if it verifies nothing itself.

From the draft below, list every authority cited in a table: Authority as cited | Paragraph | Proposition it is cited for (quote my sentence) | (a) Exists | (b) Locatable by this citation | (c) Supports the proposition | Checked by / date.
Leave (a), (b) and (c) blank; I will complete them from legislation.gov.uk, the National Archives or BAILII. Do not assess, correct or add any authority. Flag any citation without a neutral citation or pinpoint as "[INCOMPLETE]".

Draft:
[paste]

Germany: AG Köln, LG Frankfurt and the KG Berlin’s “fantasierende” KI

Charlotin lists eleven German entries and Hartung names further 2025 decisions at OLG Celle, LG Hildesheim and VGH Mannheim, but three German decisions from 2025 are the ones to read in full, two of them in family matters (see AI for family lawyers). None fined anyone; all three wrote language worth reciting.

AG Köln, 2 July 2025, 312 F 130/25. From page eight of the brief, every source was invented, down to a monograph (“Brons, Kindeswohl und Elternverantwortung, 2013”) nobody could find. The court told counsel to desist because such submissions “das Ansehen des Rechtsstaates und insbesondere der Anwaltschaft empfindlich schädigen”, and flagged § 43a Abs. 3 BRAO (beck-aktuell).

LG Frankfurt a.M., 25 September 2025, 2-13 S 56/24. Three verbatim “quotations” from BGH decisions were “komplette Fälschungen”. The forensic point: the file numbers carried the registration sign “ZB”, which should have raised suspicion, since the BGH does not hear Streitwertbeschwerden (Otto Schmidt).

Es gehört zu den Grundpflichten anwaltlicher Tätigkeit, dass ein vortragender Anwalt weder Fundstellen im Volltext erfindet noch von einem Chatbot vorgeschlagene Textquellen ungeprüft in einen Schriftsatz übernimmt. — LG Frankfurt a.M., 2-13 S 56/24

KG Berlin, 20 November 2025, 17 WF 144/25. In a legal-aid family matter the brief cited “BGH, Beschl. v. 14.11.2007 – XII ZB 183/07, FamRZ 2008, 137” and “OLG Brandenburg, Beschl. v. 28.07.2016 – 13 UF 103/16”. Neither exists. The Senat issued an Ermahnung (BRAK), and its headnote is now the German standard: lawyers “sind gehalten, mithilfe von KI verfasste Schriftsätze zu prüfen, insbesondere darauf, ob darin enthaltene Rechtsprechungszitate Ergebnis einer ‘fantasierenden’ KI sind”.

This screen applies the Frankfurt and Berlin lessons before the database check; triage, not verification.

Rolle: erfahrener Rechtsanwalt, der einen Schriftsatzentwurf gegenprüft. Erstelle aus dem Entwurf unten eine Tabelle aller Rechtsprechungs- und Literaturzitate: Zitat wie im Entwurf | Behauptete Aussage | Formale Auffälligkeiten | Prüfstatus.
Prüfe nur die formale Plausibilität: Passt das Registerzeichen zum Gericht (der BGH hört keine Streitwertbeschwerden, also kein "ZB")? Passen Senat, Datum und Fundstelle zueinander? Ist das Werk mit Autor und Jahr bekannt? Trage bei jedem Zitat "IN JURIS/BECK-ONLINE PRÜFEN" ein. Entscheide nicht, ob ein Zitat existiert; ergänze und korrigiere nichts.

Entwurf:
[einfügen]

Austria: the OGH refuses to read an AI-drafted nullity appeal

On 7 October 2025 the Oberster Gerichtshof dismissed a nullity appeal in a Graz drugs case, 14 Os 95/25i, under § 285d Abs 1 Z 1 StPO. The appeal cited supreme court decisions that “zum Großteil gar nicht oder jedenfalls nicht zum angegebenen Thema” existed: a submission “offenbar ohne fachliche Kontrolle (vgl aber § 9 Abs 1 RAO …) durch sogenannte ‘künstliche Intelligenz’ erstellte”, which “entzieht sich daher einer inhaltlichen Erwiderung” (Anwaltsblatt).

The OGH imposed nothing on the lawyer itself; the client’s appeal was simply never answered on the merits. The commentary in AnwBl 2026/101 draws the line under § 9 Abs 1 RAO, “bereits fahrlässige, unrichtige Formulierungen verletzen diese berufliche Sorgfaltspflicht”, notes that disciplinary consequences remain possible, and ÖRAK’s guidance requires “eine ausnahmslose Überprüfung” of all AI output (see the BRAK, DAV and ÖRAK rules). One Austrian practice guide answers with four-eyes review, every citation opened in the RIS, and a file note recording the check.

Draft a short internal file note in formal German, under 200 words, titled "Kurzvermerk zur Prüfung und zum RIS-Gegencheck der Zitate" for the brief described below. Fields: Aktenzeichen; Schriftsatz und Datum; verwendetes KI-Tool und Zweck; each case citation with "im RIS geöffnet am [Datum] durch [Kürzel], Aussage bestätigt/nicht bestätigt"; Korrekturen; Vier-Augen-Prüfung durch [Name] am [Datum]. Leave every date, initial and result as a placeholder; do not state that anything was verified.

Brief details:
[paste]

Switzerland, Italy, France and the Netherlands

Switzerland. On 2 June 2026 the Bundesgericht declined to enter into a self-represented litigant’s appeal, 9C_235/2026, whose first version still carried the subtitle “Entwurf – Sachlich-starke Version”. Martin Steiger: it failed “nicht an der Verwendung von künstlicher Intelligenz (KI), sondern an der nicht zielführenden Verwendung von KI”. The Swiss Bar’s Wegleitung of June 2024 adds the rule every European guidance implies: you cannot simply ask an AI whether its output is true, “da KI-System dazu nicht in der Lage sind”.

Italy. The Tribunale di Firenze dealt with invented Cassazione numbers on 14 March 2025. Counsel said a “collaboratrice di studio” had used ChatGPT without his knowledge; the court declined aggravated liability under art. 96 c.p.c. as no damage was pleaded.

France. As reported by L’Agence Sauvage, courts in Grenoble, Orléans and Périgueux issued warnings in December 2025, Périgueux being the first to use the word “hallucinations”; as of July 2026 no avocat had been sanctioned by his ordre.

Netherlands. On 24 August 2026 NL Times reported a court ordering a firm to reimburse EUR 2,135 of the other side’s costs, inside a EUR 26,000 costs award, for citing a non-existent Supreme Court ruling. The lawyer denied using AI; the court held that AI use was unprovable and irrelevant, and called the citation “reprehensible”. The European rule in one sentence: the source of the fake does not matter; the signature does.

The guidance that followed the cases

The SRA’s warning notice “Misuse of AI” of 17 August 2026 followed 42 reports of potential misuse between July 2025 and July 2026, cites Ayinde, the UKUT decision, BCP v A Mother and Cork v Smith, and states that “Reliance on an output of AI would not be a suitable defence”. Details are on its own page.

On the continent, the CCBE’s guide of 2 October 2025 warns that a lawyer may be in breach “already at the point of inputting the data into a system”, and BRAK’s Hinweise (December 2024) say that, where possible, only “abstrakte” prompts that allow no inference about a specific mandate should go to a language model. Tools built for these rules are compared in legal AI tools for Germany, Austria and Switzerland.

What European courts now expect

Six expectations follow. Every authority must exist, be locatable and support the proposition. The check is against a primary database. It is never delegated to the client (Ayinde, para 81) or a trainee (UKUT headnote 2). It is never performed by another AI. It is recorded, in a Kurzvermerk or a statement of truth. And when it fails, the lawyer says so at once: “cosmetic errors” and a blamed trainee cost more than the fakes did.

I practise in Vienna, so the verification session in AI Lab for Lawyers works through DACH and UK examples like these, not only the American ones.

Where to go next: the six-step verification protocol turns the six expectations into a routine, the US sanctions timeline shows where the same duty leads when courts fine, and the verification hub holds the rest. The prompts are in the prompt library.