A German court ruled Google liable for what its AI Overviews say, and drew the liability line for the RAG era

Summary

The Regional Court of Munich (case no. 26 O 869/26) issued a temporary injunction barring Google from spreading false claims about two Munich-based publishers through its AI Overviews. The decisive part is not the injunction. It is how the court classified the feature: an AI Overview is not a list of search results, it is Google’s own content, which makes Google a direct infringer and directly answerable for what it says.

The weight here is not that Google got hit. It is that a court has, for the first time, written down a boundary that had been left vague. The moment a system stops merely pointing you to content elsewhere and instead rewrites, judges, and recombines multiple sources into a new conclusion, that output is no longer a protected intermediary result. It becomes a statement the operator made. If that reasoning holds, it reaches far past Google. Every team that has shipped retrieval-augmented generation (RAG), AI summaries, or an AI support agent is standing on the same line.

What happened

For certain search queries, Google’s AI Overviews linked two publishers to scams, subscription traps, and shady business practices. The court found that the AI had mixed up information about other, genuinely sketchy companies with the plaintiffs and drew connections that appeared in none of the linked sources. These accusations were not in the sources and merely relayed. The AI read a handful of pages, synthesized, and produced them on its own. The publishers sent a cease-and-desist letter, Google did not respond appropriately, and the matter went to court.

The court classified Google as a direct infringer, and the reasoning is plain: AI Overviews work nothing like traditional search results. The AI rewrites and judges results “in its own words and according to its own structure,” the ruling says. In the case at hand, an overview opened with a confident assertion like “Yes, [company] is known for dubious business practices,” then built its own structure: a summary, a list of red flags for the alleged scam, and tips for users. The court called these “the defendant’s own statements.” Google built the AI and offered it to users, so Google owns what it produces, “because it alone has influence over the AI’s offering and the algorithms with which the AI operates.”

The court then dismantled Google’s two main defenses, and this is the part worth reading closely if you build with LLMs.

The first defense leaned on prior rulings from Germany’s Federal Court of Justice (BGH), which gave traditional search engines and autocomplete limited liability. The BGH’s logic was that search operators are only indirect infringers because they merely make third-party content findable, and a proactive duty to check every result would threaten how search engines work. The Munich court said that logic does not transfer to AI Overviews. A regular search engine points you to outside websites, while an AI Overview generates “independent, new, and substantive statements” by evaluating and combining content from various third-party sites. And only Google can check those statements, the court said, “at least by comparing the underlying third-party websites with its own statements based on them.” The court added that an AI Overview is “by no means absolutely necessary” for using the internet, since regular search results already help users sort information. The overview is an extra feature, so providing it is a choice, and the choice carries the liability.

The second defense was Google’s courtroom claim that users can check the linked sources themselves, and that they generally know “information generated with AI should not be blindly trusted.” The court rejected this with a line that carries real weight: the mere possibility of disproving a statement through further research does not “regularly exempt from liability for this statement.” The overview was “understandable on its own” and contained “a self-contained statement with independently understandable content and no reference to other possible interpretations or even unreliable content.” The court drew a parallel to press law, where publishers are liable for teasers that stand on their own even when readers never open the full article. It also noted that studies show users almost never click the sources in AI Overviews, and that Google’s own argument would “significantly diminish” the feature’s value if the overview had to be “generally recognized as unreliable” to escape liability.

The court closed by pointing to a protection gap. If Google were liable only for obvious violations, victims of fabricated AI claims would have no real recourse: the third-party sites used as sources never made the statements, so the victim cannot sue them, and under the old rules cannot effectively sue Google either. Someone has to be answerable, and the court placed that on the operator.

Why it matters

This is the clearest judicial answer yet to the question of who is liable for AI-generated content, and the answer is that whoever generates it owns it.

For two decades the search and platform industry has rested on intermediary immunity. You are a conduit that transmits, indexes, and points users to third-party content, so problems in that content are not on you. That single idea underwrites the business models of search engines, social platforms, and user-generated-content sites. What the Munich court did was point out that AI Overviews quietly crossed the line: the feature no longer relays other people’s words, it produces its own. Once the output is a “new statement reached by evaluating and combining,” the conduit identity collapses, and the immunity it rested on goes with it.

The limits of this ruling deserve honest labeling, so no one reads it as settled global law. It is a temporary injunction from a regional court, applying German law to one specific case. Google has not commented, and whether the reasoning survives appeal is unknown. It is not yet a binding precedent. But the chain of reasoning, that rewriting is speaking, that only the operator can verify, and that “check it yourself” fails because users almost never click through, is the kind of logic a judge in any jurisdiction could adopt. As the source notes, if it gains traction internationally the fallout could hit every AI provider whose systems paraphrase content from the web.

One thing to hedge: the ruling is not saying AI Overviews are “too inaccurate to ship.” The court was not aiming at overall accuracy. It was aiming at a specific kind of failure: the AI made its own claims that appear in no linked source. The ruling stresses repeatedly that none of the linked sources drew any connection between the plaintiffs and the “shady companies” the overview named, that the feature mixed up and invented the link itself. Accuracy is a product question. Where liability lands is a legal one, and this ruling pulled the second out on its own: even if you are right most of the time, the operator answers for the claim it generated rather than sourced.

Builder impact

If any feature in your product pulls a few sources, has a model read them, and hands the user a conclusion in its own words, this ruling is aimed at you, not just at Google. RAG, AI summaries, AI support agents, document Q&A, smart retrieval: what these share legally is exactly the trait the court pinned on AI Overviews, which is generating new statements that are not in the sources.

The first thing to reset is the default assignment of responsibility. You probably assumed that because you cite sources, the correctness of the content is the source’s problem and you are just relaying. This ruling says that once your system rewrites and judges rather than pointing through verbatim, you are the author and the correctness is yours. That is not a copy fix, it is a reclassification of the product. You have to first accept that what you produce is your own speech, then talk about how to stand behind it.

Second, disclaimers are largely useless here. The reflex for many teams is to add a line like “content is AI-generated, please verify.” Munich foreclosed that path directly: the fact that a user could verify does not exempt you, and studies show users almost never click the sources anyway. A disclaimer does not buy immunity, and that is the single most practical takeaway to remember.

Third, the workable direction: be traceable, and do not rewrite the judgment. The two faults the court kept returning to, that the AI drew connections absent from the sources and that its output offered “no reference to other possible interpretations,” point straight at a defense. Making generated content traceable line by line back to the original is step one. The more important part is not letting the model issue characterizations that the sources never made. When a user asks “is this company a scam,” presenting what the sources actually say, with disagreement and uncertainty preserved, is a legally different kind of output from declaring “yes, it is known for dubious practices.” Selling “synthesized judgment” as a product feature is, in this court’s eyes, selling legal exposure built into the product.

Fourth, the EU jurisdiction in particular deserves attention. This is a German court applying German law, and the EU’s posture on platform liability and AI governance is already far more aggressive than the United States’. If your product serves European users or you have a European entity, the real-world probability of this kind of suit is not zero. The pragmatic move is to carve out high-risk categories, such as negative characterizations of people or companies, health, legal, and financial, and either keep the AI from drawing conclusions there at all and only aggregate sources, or use a more conservative, citation-forced generation strategy. Waiting for a cease-and-desist letter before changing anything means taking the hit first.

What to ignore

Ignore the comfort of “it’s one German court, a temporary injunction, still appealable, so it doesn’t touch me.” The legal tier and finality should be labeled honestly, but the reasoning travels, and it lands precisely on the shared weak spot of every RAG-style product today. Treat it as an early signal of liability shifting, not a distant regional lawsuit.

Ignore the illusion that high accuracy keeps you safe. What got penalized here was not the overall hit rate but the AI fabricating accusations absent from the sources. A system that is right the vast majority of the time can still be held directly liable over the content in the remaining slice. Watching accuracy is watching the wrong gauge. The gauge is whether the system generated judgments the sources cannot support.

Finally, do not read this as a death sentence for AI summaries. The ruling did not forbid summarization. It said you answer for the content you generate inside the summary. Traceable summaries that do not overstep into unsupported judgments are still fine. It is the “synthesize freely and conclude boldly” style that now carries its own consequences.

Sources

1. Landmark German ruling declares Google's AI Overviews are Google's own words and makes it liable for false answers / news
2. Google haftet - Lausen erstreitet erstes Urteil zu KI-Übersichten / news

No official primary source available; this analysis is based on reliable secondary reporting (named outlets, cross-confirmed).