Google “AI Overviews” case (LG München I, 26 O 869/26) and the clinic chatbot case (OLG Hamm, 4 UKl 3/25)
Recently, two German decisions concerning artificial intelligence have successively dominated headlines in Chinese and international media: one is the Munich court holding Google liable for disseminating false information through its “AI Overviews,” and the other is the Hamm court holding a medical aesthetics clinic liable for the physician title fabricated by its chatbot. Reports have poured in, yet some exaggerate the “first-of-its-kind” or “landmark” nature, others overgeneralize the notion that “any false statement by AI triggers liability,” and still others misapply concepts from U.S. law to the German cases. Amid the clamor, the essential points are obscured.
This article does not join the fray. It proceeds solely from the original texts of the two judgments, presenting the facts of each case, analyzing their reasoning, and clarifying their core holdings, before examining their background and significance, and rectifying several widely circulated misunderstandings so as to set the record straight. At the end, download links to the original judgments are provided for those who wish to trace the sources and verify for themselves.
I. Case One: The Munich Google “AI Overview” Case (26 O 869/26)
Facts of the Case
The plaintiffs are two publishing companies based in Munich. Their troubles began with a new Google Search feature—"AI Overviews" (German: Übersicht mit KI, i.e., AI Overview). When a user searched for the keywords "publisher's name + scam (Betrugsmasche)", the AI Overview displayed at the top of the search results page generated a composite answer linking the plaintiffs to scams, improper business practices, "subscription traps (Abo-Fallen)", and furthermore claimed that they were connected to certain other enterprises, cited fictitious telephone calls, and even asserted that they continued to demand payment after customers had already paid.
The crux of the matter is this: these associations largely resulted from the AI confusing the plaintiffs with other genuinely dubious businesses—attributing the misdeeds of one party to another. What is worse, the AI's asserted connections even went beyond the source webpages listed by Google itself. This is a classic "entity confusion hallucination."
The Court’s Reasoning
The court grounded the claim in sections 1004 and 823(1) of the German Civil Code (Bürgerliches Gesetzbuch, BGB), read together with the corporate personality right (goodwill) as established under Article 2(1) in conjunction with Article 19(3) of the Basic Law (Grundgesetz). The pivot of its reasoning lies entirely in the four words “sich zu eigen machen” (adopt as one’s own).
The court found that the AI Overview is not a mere listing or referral to third-party webpages, but constitutes Google’s own content, for which it is accountable. This is because it rewrites in its own language, restructures in its own format, and makes its own assessments of the content (the affirmative opening “Yes, … is known for unfair business practices,” the self-designed columns, and the action suggestion “What you can do”). In particular, the association it asserts is not even supported by the very sources that Google itself cites. All of this goes beyond the proper role of a “signpost” and amounts to an “independent, new, substantive statement.” Since it is Google’s own statement, it must answer for it as a direct interferer (unmittelbare Störerin).
On this basis, the court made a crucial departure from precedent. Earlier case law of the Federal Court of Justice (Bundesgerichtshof, BGH)—the traditional search engine case (VI ZR 489/16) and the autocomplete case (VI ZR 269/12)—had consistently treated search engine operators leniently: they were subject only to the limited liability of an indirect interferer, meaning they are obliged to remove content only after receiving notice and only if the infringement is “manifest.” The rationale was that search is indispensable for use of the internet, and imposing a general ex ante review obligation would be both unworkable and would stifle the search engine’s core function. The Munich court, however, expressly stated that this logic of leniency cannot be transplanted to the AI Overview—because a mere list of links is already sufficient to tame the flood of information, and the AI Overview is not indispensable for internet use. Since the premise differs, the basis for exemption cannot simply be carried over.
As for the characterization of the content, the court followed the orthodox German speech-law distinction: “statements of fact” that can be verified as true or false, and “expressions of opinion” that convey approval or disapproval. A crucial rule applies: even an opinion loses its protection if the factual basis on which it rests is itself untrue (unwahre Anknüpfungstatsachen). The plaintiff demonstrated by affidavit (eidesstattliche Versicherung) that it has no actual connection to the alleged subscription-trap enterprises, and Google did not mount a substantive rebuttal—if the foundation is gone, what can remain? Both of Google’s defenses were likewise defeated. The first—that “users can click on the links to verify for themselves”—was rejected by the court on the ground that the possibility of subsequent debunking does not negate present liability, and that if an AI Overview requires verification of each individual item before it can be trusted, the very purpose of “quickly providing reliable answers” would be self-defeating. The second—that “the content is no longer displayed”—was met with the court’s finding that the risk of repetition (Wiederholungsgefahr) had not been eliminated, because Google had not issued a cease-and-desist declaration with a penalty clause, and the algorithm could still generate the content again.
Key Holding
Google must answer for the false statements generated by its AI Overviews as a direct interferer in its own right and may not invoke the liability exemption for search engines and hosting service providers (commonly referred to as the “safe harbour”). The present case was brought in interim injunction proceedings (einstweilige Verfügung); the standard is prima facie substantiation to a high degree of probability. The judgment is a final decision (Endurteil) rendered after oral argument; however, “final” here means only that it concludes the interim injunction proceedings at this level of instance – it is not a “final determination on the merits” – and the judgment is not yet final and binding; Google may appeal. The court did not decide entirely in favour of one party but granted the application in part and dismissed it in part (the end result being roughly an 80/20 split): for example, the phrase “working with a debt collection agency” did not reach the threshold for reversing the burden of proof, and the phrase “selling subscriptions and dunning in an improper manner” was not even contained in the Overview. The injunction is not limited in effect to the territory of Germany. In addition, the court clarified several points of law: jurisdiction is based on Article 7(2) of the Brussels I Regulation (recast); the applicable law is German law pursuant to Article 40 of the Introductory Act to the German Civil Code (EGBGB); the General Data Protection Regulation (GDPR) does not apply because the claimant is a legal person; the EU Artificial Intelligence Act (KI-VO) cannot serve as a basis for a claim; and the Digital Services Act (DSA) does not preclude the application of national law.
II. Case Two: The Hamm Clinic Chatbot Case (4 UKl 3/25)
Facts of the Case
The plaintiff is the North Rhine-Westphalia Consumer Protection Association (a plaintiff in a representative action). The defendant is a clinic providing minimally invasive aesthetic medicine, whose two operators are physicians publicly known as “Dr. Rick & Dr. Nick” but who in fact lack the qualification of “specialist in plastic and aesthetic surgery.” The defendant’s website featured a customer service chatbot capable of answering questions and booking appointments in real time. In April 2025, when responding to three user queries, the chatbot directly described the two operators as “specialist in plastic and aesthetic surgery,” “specialist in aesthetic medicine,” and “specialist in aesthetic treatment” — the latter two professional titles do not exist at all.
Of particular note: the chatbot was developed by a third party under commission, and its training data was drawn solely from the defendant’s own correct website content and FAQ, which originally contained no misrepresentation of qualifications. In other words, the errors were purely hallucinations fabricated by the machine out of thin air. The defendant raised the defence accordingly: this was not its own act; the chatbot system operated autonomously according to statistical probabilities, as an uncontrollable black box; the training data was originally correct; and the public is well aware that AI can err and will independently verify the information.
The Court’s Reasoning
The court derived the cause of action from section 2 of the Gesetz über Unterlassungsklagen bei Verbraucherrechts- und anderen Verstößen (Unterlassungsklagengesetz – UKlaG, Act on Injunctive Relief for Consumer Rights and other Violations) in conjunction with sections 3, 5(1) and 5(2)(3) of the Gesetz gegen den unlauteren Wettbewerb (UWG, Act Against Unfair Competition) concerning misleading commercial practices, including deceptive statements about a trader’s qualifications or identity.
First step: does the robot’s answer constitute a commercial practice “of its own” by the defendant? The court held: yes. The concept of “commercial practice/practice” is broad and extends to the use of technical means and to automatically generated output. The court cited the BGH decision on the contract document generator (I ZR 113/20) – a document generated purely by software is still a commercial practice of the person behind it. The robot is merely a “technical means” over which the defendant has sufficient control (the fact that it could easily be corrected afterwards and be fitted with a neutral-answer instruction and a keyword filter is compelling proof). The “black box problem” does not alter this: the contract document generator likewise operates without anyone intervening in each individual case, yet liability is still attributed to the operator. Whoever introduces the AI, sets its “operating framework” and makes it run is responsible for its output.
Second step: the defendant’s invocation of the wettbewerbsrechtliche Verkehrspflicht (duty of care under competition law) to seek exoneration was also rejected. The court distinguished: that duty concerns situations where “the operator creates a risk of infringement by a third party” (such as operating an e-commerce platform on which third parties sell counterfeit goods); in the present case, however, “the robot is not a third party” but forms part of the defendant’s own organisation, so what is at issue is an infringement committed by the defendant itself, not liability for a third party.
Third step, displaying even greater doctrinal precision: the defendant invoked the BGH’s autocomplete decision (VI ZR 269/12) in its defence, but the court rejected this – the case concerned “Störerhaftung (interferer liability) of a search engine for third-party content”; and according to the settled case-law of the BGH, in the field of “Verhaltensunrecht” (conduct-based wrongfulness, i.e., competition law), Störerhaftung no longer applies at all, so the autocomplete decision has nothing whatsoever to do with the “own infringement” in the present case. Here the contrast with the Munich decision is striking: Munich reclassified within the interferer framework (finding Google to be a direct interferer), whereas Hamm simply abandoned the interferer framework and proceeded directly on the basis of Täterhaftung (actor liability).
The court finally drew the threads together with a “belt and braces” conclusion: even if, for the sake of argument, the duty of care or the doctrine of interferer liability were applied, the defendant would still be liable – because its activity was not purely passive; a prudent trader should have foreseen that users would ask about professional titles and that the AI might hallucinate in this respect; the medical field called for stricter standards; and since it was so easy to correct afterwards, ex ante preventive measures were to be expected and were reasonable (zumutbar). As to whether it was misleading, the court determined: an average consumer has certain expectations of a “specialist doctor” (a recognised programme of further training and an examination must have been passed), and the robot’s answer, which did not correspond to the facts, was capable of deception. The defendant’s argument that “the public knows AI makes mistakes and will therefore check for itself, so there is no commercial relevance” was expressly rejected by the court: there is no such empirical rule; on the contrary, the public places particular trust in answers given by a machine precisely because it is perceived as less prone to error than a human being – otherwise why would the defendant have deployed this robot in the first place?
Key Holding
The robot’s false response constituted the defendant’s own misleading commercial practice, for which the defendant must be liable as a actor (Täter). Liability attaches regardless of whether the error was caused by a technical malfunction or an AI hallucination. All claims were granted: prohibition of the use of three professional titles, with a fine of up to €250,000 (Ordnungsgeld) for each violation, and reimbursement of warning costs in the amount of €260. The risk of repetition has not been eliminated, because although the defendant took down and corrected the robot, it failed to provide a cease-and-desist declaration with a penalty clause. It must be specifically noted that this case is not a preliminary injunction, but an action on the merits for a permanent injunction (Hauptsache-Unterlassungsklage); the standard of proof is full proof. Moreover, it was heard by the Higher Regional Court (OLG) at first instance — pursuant to Section 6 of the Injunctive Relief Act (UKlaG), the Higher Regional Court of Hamm has had exclusive jurisdiction over such representative actions for injunctive relief in North Rhine-Westphalia since July 2025. Given the fundamental importance of the case, the court expressly granted leave to appeal to the Federal Court of Justice (BGH). The judgment is not yet final and binding.
III. Viewing the Two Cases Together: Convergence and Divergence
There are three points of convergence. First, the locus of attribution is identical: both courts treated the AI output as the operator’s “own” content/conduct, not as a third party’s act—the arguments of “autonomy, black-box, statistical, uncontrollable” were all rejected. Second, both courts demolished the “user self-checking / disclaimer” defence, and the reasoning is virtually identical: both said there is no empirical rule that “the public knows AI makes mistakes and will therefore verify”; on the contrary, the public trusts machine answers more. A general disclaimer (“AI may make mistakes”) was judged insufficient to escape liability in both cases. Third, “if it can be corrected, it should have been corrected early” served as the same hammer to defeat the defence: in both cases the defendants easily corrected the output afterwards, which itself proves that ex ante preventive measures were feasible and a duty of care could be borne; and because neither had submitted a cease-and-desist declaration with a penalty clause, the risk of repetition (Wiederholungsgefahr) was not eliminated.
The divergences are also pronounced. The first lies in the tier and weight: Munich was a Regional Court decision in interim injunction proceedings, applying the prima facie substantiation standard; Hamm was a Higher Regional Court decision in an action on the merits, applying the full-proof standard, and expressly permitted an appeal to the BGH—as regards the weight of “approaching a settled rule,” Hamm is in fact more substantial, while Munich has gained wider publicity because of Google’s name. The second is the legal field: Munich followed the path of expression law / corporate personality rights, Hamm followed the path of unfair competition law / consumer protection. The third, and the most intriguing, is the different treatment of Störerhaftung (interferer liability): Munich preserved the framework and reclassified (direct interferer, therefore the search engine exemption does not apply), while Hamm abandoned the framework and directly attributed to the actor (in the field of conduct-based unlawfulness, Störerhaftung no longer applies; the robot is not a third party). One preserves the framework, the other discards it, yet both arrive at “own liability” by different routes—this is not coincidental, but results from doctrinal differences in the two fields: personality rights injunctive relief still uses the interferer doctrine, whereas competition law has long since shifted to the actor (Täter) plus duty of care (Verkehrspflicht).
A consistent insight thus emerges: liability attaches not to a particular answer, but to the system architecture deployed and controlled by the operator. German law has already produced a cross-field (expression law + competition law), cross-tier (regional court + Higher Regional Court) convergence on attributing AI-generated content to the operator; and both judgments touch on the EU Artificial Intelligence Act as supporting authority (Munich invokes the concept of “provider,” Hamm touches on transparency obligations), yet both use existing civil law and competition law as the backbone of the decision, without using the AI Act as a basis for a claim.
IV. Background, Significance, and Setting the Record Straight
The deep background of the two cases lies in the fact that generative AI is quietly shifting the platform’s role: from “pointing the way” to “giving answers,” from “listing links” to “synthesizing and generating.” The safe harbour regime (whether notice-and-takedown or the red flag rule) derives its justification from the platform’s status as a “passive intermediary”—merely accessing, transmitting, storing, and indexing third-party content, without producing or evaluating it. When this premise collapses, applying the old immunity is tantamount to allowing a “speaker” to enjoy the exemption of a “postman.” This is the structural time lag that safe harbour rules reveal in the face of generative AI. The significance of the two judgments lies precisely in their keen recognition of this shift and in providing a decisional path: when a platform transforms from a referrer into a speaker, immunity dissolves.
Amid the excitement, however, several misunderstandings must be clarified to prevent misinformation from spreading:
First, not a “global/German first” case. As early as September 2025, the Hamburg Regional Court had already issued a temporary injunction (324 O 461/25) against the false statements of the X platform’s chatbot Grok. Looking abroad, the US case Walters v. OpenAI (filed in 2023, dismissed in 2025 on domestic procedural grounds) and the Australian matter Hood v. OpenAI (a “concern notice” sent in 2023) predate these cases by several years. The label “first” is thus inaccurate.
Second, not an absolute proposition that “AI is liable for any false statement.” This is the most dangerous misunderstanding. Liability arises not from the term “hallucination” itself, but from the cumulative effect of several factors: the content is a false factual assertion (or an opinion based on a false fact); the operator has adopted the content as its own (or it arises from the operator’s own commercial conduct); and the operator could have verified or foreseen the falsity. If any of these three elements is missing, the conclusion may not be the same. Readers must not extrapolate this into a blanket rule that “AI platforms are henceforth unconditionally liable for all outputs.”
Third, not a final and binding judgment. Both judgments are not yet final and binding: the Munich case is subject to appeal to the Munich Higher Regional Court, and the Hamm case has been granted leave to appeal to the Federal Court of Justice. The final outcome thus remains uncertain.
Fourth, not a direct source of authority for Chinese law. Given the differences in legal systems, German judgments have only persuasive value for Chinese law and are not binding. Some commentaries have cited Section 230 of the US Communications Decency Act to interpret the German cases, which is a classic case of misapplication—Section 230 is unique to US law; it neither binds German courts nor served as the basis for these two judgments.
As a mirror for Chinese law, the two cases reflect two distinct legal grounds: the Munich case resembles a direct infringement of the right to reputation under Article 1024 of the Civil Code; the Hamm case resembles false advertising under Article 8 of the Anti-Unfair Competition Law (false medical qualifications, which may also engage specialized regulations on advertising and medical advertising). The logic that “the robot is not a third party; it belongs to the operator’s organization and the operator has control and therefore must bear liability” can be directly translated into Chinese law as “the operator’s own conduct / network content provider,” sufficient to dismantle any excuse of “technological neutrality / safe harbour.” Moreover, the Hamm case is a consumer association collective action, which shares the same mechanism as public interest litigation by consumer associations in China (Article 47 of the Consumer Rights Protection Law)—a point worth noting in tandem.
Conclusion
Viewed together, the two cases can be summed up in a single phrase: Though machines may speak, accountability ultimately rests with humans. AI can write, analyze, and answer, yet it cannot be held responsible for what it says; that responsibility falls upon the person who introduces it, controls it, and sets it in motion. The answers provided by the German courts are not necessarily final—after all, both judgments have yet to take effect—but the way they frame the questions strikes at the core of the matter, and deserves serious attention from anyone concerned with platform liability. The spectacle is fleeting; the principle endures. Let us observe it together.
Annex: Download of Original Judgments in Both Cases
Case 1 – Munich I Regional Court, 26 O 869/26 (28 May 2026, final judgment on a preliminary injunction · redacted certified copy)
- Full-text PDF (official redacted certified copy hosted by the-decoder):
https://the-decoder.de/wp-content/uploads/2026/06/26_O_869_26_begl_Abschrift_Urteil_v_28_05_2026_Geschwaerzt_Geschwaerzt_Geschwaerzt.pdf - Case-law database entry (dejure.org): https://dejure.org/2026,16717
Case 2 – Hamm Higher Regional Court, 4 UKl 3/25 (12 May 2026, final judgment · official full text)
- NRW official case-law database HTML:
https://nrwe.justiz.nrw.de/olgs/hamm/j2026/4_UKl_3_25_Urteil_20260512.html - NRW official case-law database PDF download:
https://nrwe.justiz.nrw.de/pdfdownload/downloadEntscheidung.php?entscheidung=/nrwe/olgs/hamm/j2026/4_UKl_3_25_Urteil_20260512.html - ECLI:
ECLI:DE:OLGHAM:2026:0512.4UKL3.25.00
[Factual and Temporal Note] The factual basis of this article consists of the two judgments above, supplemented by publicly available reports from both Germany and China, and verified as of June 2026. The judgments in both cases are not yet final (the Munich judgment is appealable; the Hamm court has already granted leave to appeal on points of law to the BGH), and the ultimate conclusions are subject to change. References to German and EU law are subject to the contents of the judgments. Comparisons involving Chinese law constitute a framework analysis; application to specific cases must be separately verified against the latest legislation and judicial decisions. Earlier overseas precedents (US Walters, Australian Hood) are cited solely to examine the “first case” claim. The names of the parties involved have been anonymized in the German court judgments and in most media reports; this article follows the same practice.