Google's own words

The Regional Court of Munich issued a temporary injunction against Google last week, and buried in case number 26 O 869/26 is the most consequential sentence yet written about AI search: the AI Overview is Google's own content. Not a search result. Not third-party material made findable. Google, speaking.

For twenty years the web's liability architecture rested on one move: the platform points, someone else speaks. Safe harbors in the EU, Section 230 in the US, a line of German Federal Court of Justice rulings shielding search and autocomplete — all of it assumes the operator is an intermediary between a reader and a source. The Munich court ruled that the move doesn't survive generative AI. That's the whole story, and it's bigger than the case.

The case itself was ugly in an instructive way. Google's Overviews falsely tied two Munich publishers to scams and subscription traps — connections that appeared in none of the linked sources. The AI mixed the plaintiffs up with genuinely sketchy companies and opened its answers with lines like "Yes, [company] is known for dubious business practices." The court examined the old search-engine shield — operators are only indirect infringers because they merely make third-party content findable — and found it doesn't transfer. A search engine points. An Overview asserts, generating what the court called "independent, new, and substantive statements."

Google's defense was that users can check the linked sources themselves. The court rejected it, and the data says why: about 1% of users ever click a source link from an AI Overview. The court reached for press law instead — a publisher is liable for a teaser that's understandable on its own, even if nobody reads the full article. And the scale math is unforgiving. An analysis by Oumi for the New York Times found the current model answers correctly 91 percent of the time, which sounds fine until you multiply the remainder by Google's query volume: millions of wrong answers every hour. The same analysis found 56 percent of the *correct* answers weren't supported by the sources Google linked. The product asserts things nobody said, with citations that don't contain them.

The honest caveat: this is one regional court, a temporary injunction, and Google will appeal. The BGH could restore the shield. But the court's protection-gap logic travels across borders, because every jurisdiction has the same hole. The source sites never made the false claims, so victims can't sue them. If the operator isn't liable either, then nobody is — a defamation with no defendant. Courts abhor that vacuum, and they don't need German statutes to find it.

The engineering consequence is the part worth sitting with. Every AI search product — Google's, OpenAI's, Anthropic's, Perplexity's — has treated hallucination rate as a quality metric, a number to improve between releases. In at least one major market it's now legal exposure, denominated in injunctions and 80 percent of the other side's legal costs. The web's oldest alibi was "we just point to it." The machine stopped pointing. The alibi went with it.