Google's AI Overviews Are Now Google's Words, and That Changes Who Answers for the Machine

A regional court in Munich has decided that when Google's AI Overviews make a claim, Google is the one making it. In a preliminary injunction issued on 28 May 2026 and reported publicly between 9 and 11 June, the Regional Court of Munich I, sitting as LG Munich I under case number 26 O 869/26, barred Google from repeating false statements its AI-generated summaries had produced about two Munich publishers. The statements had wrongly tied the companies to scams, subscription traps and dubious business practices. The court's reasoning did not turn on whether the information was wrong. It turned on who said it, and the answer it gave reorganises a question that has shadowed generative AI since it reached the mass market.

The decision matters because of the category it chose. German federal courts have for two decades treated search engines as intermediaries, liable only as indirect infringers because they make third-party content findable rather than create it. That settled position is what has shielded Google's core business across Europe. The Munich court found that the shield does not stretch to cover AI Overviews, because the feature does not point users to outside material in the way a search index does. It generates something new.

The plaintiffs, Munich publishing house Verlagshaus24 and a subsidiary, discovered that Google's summaries had linked them to fraudulent companies, fabricated phone calls and claims of unavailability, none of which appeared in any of the sources the AI cited. The court described the output as the company's own statements, reasoning that the AI evaluates, combines and rewrites material from various third-party sites into “independent, new and substantive statements.” Bernhard Buchner, the lawyer who acted for the publishers alongside the firm Lausen Rechtsanwälte, said the court had sent a clear signal that providers cannot hide behind AI when misrepresentations occur. The court was blunt about ownership: Google built the system, offered it to users and controls the algorithms behind it, so Google answers for what it produces. The “created with AI” label, the chamber held, does not change the attribution.

The “users can check for themselves” defence collapsed under its own evidence

Google's central argument in the hearing was that users could verify the summaries against the linked sources and that people generally understand AI-generated information should not be trusted without question. The court rejected this on two grounds, one legal and one empirical. As a matter of law, it held that the possibility of disproving a statement through further research does not ordinarily exempt the speaker from liability for having made it, drawing a parallel to press law, where a publisher is liable for a standalone teaser that reads as a complete claim even if the reader never opens the full article.

The empirical ground is more striking because it uses behaviour against the defence. A Pew Research Centre analysis of 900 US adults across nearly 69,000 searches in March 2025 found that when an AI summary appeared, users clicked a source link inside it only 1% of the time, and clicked any traditional result in just 8% of searches, against 15% without a summary. The court's logic ran directly through that finding: a feature presented as a reliable answer cannot be defended on the basis that users will independently check it, when almost none of them do. The judges noted that Google's own argument would hollow out the value of the product, since the overview would have to be treated as generally unreliable for the defence to hold.

The court went further than intermediary liability and addressed the speech status of AI output directly. It held that an AI's apparent opinion is not the expression of a held conviction but the result of an algorithm, and that offering AI-powered research is primarily an expression of Google's commercial activity rather than a protected exercise of opinion. Weighing the publishers' interest in removing falsehoods against Google's commercial speech interest, the court found the publishers' position outweighed Google's, particularly because the contested statements rested on untrue facts. Google could not fall back on host-provider protections under the Digital Services Act, nor on the notice-and-takedown process available to ordinary search engines.

The practical terms reflect how decisively the court came down. Google was ordered to bear roughly 80% of the legal costs, with the publishers covering 10% each, and faces a penalty of up to €250,000 for non-compliance. The injunction takes effect across the European Union. Google said it was carefully reviewing the decision, which is not yet final, and an appeal is considered probable.

This was not a bolt from the blue, but the point a trend reached ground

The ruling reads as a landmark, yet it sits at the end of a developing line of German reasoning rather than at its start. The Kiel Regional Court held in February 2024 that the operator of a business information portal using AI for data processing is liable as a direct infringer when illegal content is disseminated through its software, regardless of whether the operator was directly involved in the automated process. In September 2025 the Frankfurt Regional Court, in a competition-law matter referenced as 2-06 O 271/25, established that a provider could in principle be liable for false AI summaries, though it dismissed the plastic surgeon's injunction in that case after finding the statement was ultimately not false when read in full context. Munich converted a conditional possibility that Frankfurt had left open into an operative finding of direct liability. IT lawyer Stefan Lutz, in an analysis of the decision, said the ruling strengthens the position of companies whose reputations are harmed by AI-generated search results.

The reclassification carries weight beyond Google because the reasoning attaches to a capability, not a brand. Any system that paraphrases the open web into a coherent answer produces the same kind of synthesised statement the court identified, which is why the decision is read as reaching, at least in principle, the answer engines built by OpenAI, Anthropic and Perplexity. The scale of the exposure is visible in the accuracy data. An analysis by the AI startup Oumi for The New York Times found that Google's AI Overviews running the current Gemini 3 model answered correctly 91% of the time, a rate solid enough for everyday use that nonetheless produces millions of wrong answers given Google's query volume. The same analysis found that 56% of the correct answers could not be substantiated by the sources Google had linked, which is the precise failure the Munich court fixed on: a system generating claims that trace back to no source and that only the operator is positioned to check.

The credibility question is really an ownership question

The reason this ruling speaks to credibility, and not only to liability, is that the two have become the same problem. An AI summary earns its authority by reading as a settled answer rather than a set of pointers, and that authority is exactly what the court treated as the source of responsibility. As long as a provider could characterise its system as a neutral aggregator, the credibility of the output was presented as borrowed from the underlying web, with the provider standing apart from it. Munich removed that separation. Once the output is the provider's own statement, the provider owns both its persuasive force and its errors. Forrester principal analyst Nikhil Lai captured the operational consequence when he observed that AI overviews can no longer be merely helpful summaries but must now be legally defensible outputs.

Tim Pfaelzer, General Manager and Senior Vice President for EMEA at Veeam, reads the decision as a structural turning point rather than a setback. The biggest barrier to enterprise adoption, he argued, has never been capability but accountability, and clarifying that AI outputs are statements attributable to the provider closes the accountability gap that has slowed confidence in AI at scale. He expects a period of caution in the near term, with providers tightening models, reducing risk tolerance in outputs and investing more heavily in validation and provenance, which may slow consumer-facing innovation. The longer arc he describes runs the other way. Enterprises need clarity on who is responsible when AI gets something wrong, Pfaelzer said, because without it AI remains a reputational and operational risk, and with it AI becomes something organisations can govern, insure and deploy with more confidence. Trust, in his account, is what unlocks adoption, and trust follows from accountability.

That framing is contestable, and the contest is geographic.

Europe and the United States are now answering the same question differently

The Munich approach throws into relief how far the United States sits from it. In Walters v. OpenAI, a Georgia state court granted summary judgment to OpenAI in May 2025, ending the first prominent defamation claim over a generative AI output. Radio host Mark Walters had sued after ChatGPT produced text wrongly describing him as a defendant accused of embezzlement in a case he had no part in. The court's difficulty was that defamation in the United States generally requires fault, and applying concepts such as knowledge or reckless disregard for the truth to a non-human system is awkward, since treating a provider as liable merely because its model is capable of generating a falsehood would impose something close to strict liability on anyone operating such a tool. The German court sidestepped that obstacle entirely by routing the question through intermediary liability and defamation principles that ask who made the statement, rather than through a fault standard that asks what the speaker knew.

The divergence is not academic, because both jurisdictions are under parallel commercial pressure that sharpens the stakes. The Independent Publishers Alliance filed an EU antitrust complaint over AI Overviews in July 2025, alleging that Google uses publisher content to build its summaries without consent, payment or an opt-out, and the European Commission opened separate Article 102 proceedings on related grounds in December 2025. In the United States, Penske Media, owner of Rolling Stone, Variety and Billboard, brought a federal antitrust case in September 2025. The litigation around AI output is widening from copyright and competition into direct responsibility for what the systems assert, and the two largest regulatory blocs are arriving at incompatible answers about where that responsibility lands.

For markets that take their regulatory cues from Brussels, the signal is the more consequential half of the ruling. The EU often sets the tone in data protection and digital governance, as it did when the GDPR shaped frameworks well beyond Europe, including across the Middle East. Pfaelzer expects organisations operating globally to align to the strictest applicable standard, which increasingly resembles the European model, and to design for multi-jurisdiction compliance from the outset rather than treating regulation as something to retrofit. He situates the Munich decision within the wider movement of the EU AI Act, whose rules for general-purpose AI took effect in August 2025 ahead of full applicability in August 2026, and reads it as evidence that AI is moving from experimental technology to a regulated infrastructure layer. That reading is an interpretation of where the law is heading, not a description of what the Munich court decided. The chamber's reasoning rested on German intermediary-liability and defamation principles and on the limits of the Digital Services Act, not on the AI Act, a distinction worth holding onto as the decision is cited in support of broader claims about output-based jurisdiction.

What the court did settle is narrower and, for that reason, more durable. It decided that the credibility an AI answer projects and the responsibility its provider carries are two sides of one object. Whether that holds on appeal remains open, and a single regional injunction is not a final precedent. The reasoning, though, is portable in a way that a fact pattern is not, and it gives every company building answer engines a question it can no longer route around: if the output is yours, so is the error.