A German Court Just Ruled That Google
Owns The Lies Its AI Overviews Tell.

Most of the AI failures we document here are about a product breaking in public. A chatbot invents a court case, a database sits exposed, a health bot fumbles an emergency. This story is about something rarer and arguably more important. It is about a court answering the question the whole field has spent two years dodging. When a generative system fabricates a damaging falsehood, who answers for it? On May 28, 2026, the Regional Court of Munich I gave its answer, and the answer was Google.

The case, docketed as 26 O 869/26, came from two Munich-based publishing companies. When certain searches ran through Google's AI Overviews, the feature that summarizes results in a confident paragraph above the blue links, the summary told users the publishers were tied to scams, subscription traps, and dubious business practices. According to the plaintiffs, the AI stitched their names to shady operations they had nothing to do with. The connection did not exist in the underlying sources. The model simply asserted it, in fluent and authoritative prose, to anyone who happened to search.

Why This Ruling Is Different

Why does this matter beyond one defamatory paragraph? Search engines have spent a quarter century protected by a simple principle: they point at other people's content, they do not author it, so they are not responsible for what that content says. That shield is the legal bedrock the entire search business sits on. Google's defense leaned directly on it, arguing that an AI Overview is just another way of surfacing results and that users can always click through and verify for themselves.

The court did not buy it. It found that the AI Overview was not a neutral pointer to third-party material but an independent statement that Google itself created. In the court's framing, because the system independently compiles and condenses information into a new summary text, the result is the defendant's own presentation, something Google must be held accountable for. The summary made claims that did not even appear in the search results it supposedly drew from. That, the court reasoned, is not indexing. That is authorship.

Judges went a step further and dismissed the idea that the burden of verification belongs to the reader. Google had argued that because the linked sources sit right there, a careful user could check the claim and catch the error. The judges rejected that, finding the overview was understandable on its own as a self-contained statement with independently comprehensible content. In plain terms: if your product hands a person a finished, confident paragraph, you do not get to blame the person for believing it.

The Host-Provider Defense Failed

Google also reached for the protection that platforms across Europe rely on under the Digital Services Act, the host-provider safe harbor that shields a company from liability for user-generated content until it is notified and fails to act. The court closed that door too. A host provider stores and transmits material that someone else created. An AI Overview is created by Google's own system. You cannot claim to be a passive host of content your own model wrote. With that, the notice-and-takedown logic that lets platforms say "we will remove it once you flag it" simply did not apply.

What came next was an injunction. Google is barred from repeating the false statements about the two publishers, and under German civil procedure that kind of order carries real teeth, with escalating penalties available for non-compliance that can reach the company's management. The court also assigned the costs lopsidedly, ordering Google to carry the large majority of the legal bill while each plaintiff covered a small fraction. This was not a split decision dressed up as a win. It was a clear finding that the company running the model is the party responsible when the model defames someone.

The Defense The Whole Industry Was Counting On

To understand why a single regional court in Germany is making headlines on three continents, you have to understand the bet the AI industry placed. Every major player has shipped systems that confidently produce false information, and every one of them has quietly relied on the same fallback: the output is not a statement of fact, it is a generated guess, and nobody owns a guess. It is the same posture that runs through our wider record of documented AI failures, where the harm is always real but the responsibility is always somebody else's. The Munich court looked at that posture and ruled it does not survive contact with defamation law. If your machine writes a sentence and publishes it to the world, the sentence belongs to you.

That logic does not stop at search overviews. The same reasoning that makes Google answerable for a hallucinated summary makes any company answerable for a hallucinated answer from any generative product. It is the doctrinal cousin of the wave of cases we have tracked where courts have stopped treating AI errors as harmless accidents, from the Mississippi judge who sanctioned lawyers on both sides for citing AI-invented cases to the federal appeals court that drew a hard line on AI fabrications in legal filings. Those rulings punished the humans who trusted the machine. This one goes after the company that built it.

What Happens Next

Google is not accepting the result. The company said it is carefully reviewing the decision, which is not yet final, and has signaled that it intends to appeal. A higher court could narrow the reasoning, and a ruling in one German regional court does not bind judges in the United States or anywhere else. Nobody should pretend this single decision rewrites global law overnight.

But precedent does not need to be binding to be powerful. It needs to be persuasive, and this one is. It hands every defamed business, every wrongly described person, and every regulator a clean and quotable argument: the AI summary is the company's own publication, and the safe-harbor defenses built for the link-listing era do not stretch to cover it. We have spent months cataloging the human cost of these systems in our running archive of AI lawsuits, and the recurring theme has always been a liability gap, a space where harm lands on real people while no institution has to answer for it. Munich just started filling that gap.

Underneath all of this sits a question about incentives. As long as the cost of a confident falsehood is zero, the rational move is to ship faster and apologize never. The moment that falsehood carries legal weight, the math changes. Suddenly the guardrails that get cut to hit a launch date have a price tag attached. That is the real significance of a small courtroom in Bavaria. It is not the size of the case. It is the principle that a company cannot build a machine to speak in its name and then disown every word it says.

Has an AI tool published something false about you or your business? Tell us what happened.