Industry

A German court held Google responsible for what its AI said about a business

A Munich court ruled Google's AI Overviews are Google's own words. Then the case settled. What it changes, and what you can actually do about it.

Multi Chats Team
September 27, 2026 · 9 min read

Type a company name into Google in Germany earlier this year, add the word the search box suggested next, and the "Übersicht mit KI" panel might have told you the business was a known scam. Two Munich publishing companies saw exactly that. On 28 May 2026 the Munich I Regional Court (Landgericht München I) ordered Google to stop, holding that Google owns the sentences its AI writes.

That judgment never became binding. Google appealed to the Munich Higher Regional Court, the parties then settled before the appeal was decided, and the Bavarian state case law database that publishes the full text records the judgment as having no further effect. That makes it a useful piece of reasoning to read. It does not make it law anyone can rely on.

Below are the questions people actually ask about it, each answered in the same shape.

What did the court actually decide?

The publishers applied for a preliminary injunction, and the court granted most of what they asked in a final judgment within those interim proceedings, case 26 O 869/26. Google was barred from repeating a specific list of claims about them, backed by the standard German enforcement clause of up to 250,000 euros per breach.

Two headnotes carry the reasoning. First, a search engine operator is liable as a direct interferer for AI-generated overviews that contain untrue factual claims or inadmissible opinion damaging a company's business personality right. Second, AI-generated overviews count as the operator's own attributable content, so the operator cannot invoke the liability shields for hosting providers or search engines under the Digital Services Act or the E-Commerce Directive. The court rejected Google's argument that readers could go and check the linked sources themselves: the panel was self-contained, and the chance to falsify a claim by further research does not relieve whoever made it of responsibility. Two smaller parts of the claim were refused.

Short answer: A first instance court, in urgent proceedings, ordered Google to stop repeating specific false claims and said the search engine liability shield does not cover text the AI writes itself.

Who was suing, and what did the AI say about them?

A Munich publishing house and its subsidiary. The trigger was searches combining the company name with "Betrugsmasche", the German word for a scam, which Google's own autocomplete offered alongside the name.

Legal Tribune Online quotes the opening line of the overview: "Ja, das Verlagshaus ist bekannt für unseriöse Geschäftspraktiken und wird oft als Betrugsmasche wahrgenommen." In English: the publishing house is known for dubious business practices and is often perceived as a scam. The panel went on to list supposed hallmarks: subscription traps, invoices for phone calls that never happened, continued payment demands after a bill was already paid, frequent changes of name and web address. It closed by advising the reader to get a lawyer.

The court's problem was sourcing. None of the pages linked under the panel connected the publishers to the third party companies the overview named. The claimants filed a sworn declaration that no such connection existed, Google offered no substantiated case to the contrary, and the court treated the link as untrue for the purposes of the proceedings.

Short answer: The overview asserted connections that appeared in none of the sources it cited, and the publishers proved they did not exist.

What does "never became binding" mean here?

German civil cases start at a regional court, and the losing side can appeal to a higher regional court. Until that appeal is decided, the first judgment is not final. Google lodged an appeal, docketed 18 U 1744/26 at the Munich Higher Regional Court. Before that court ruled on anything, the parties ended the dispute by settlement, and the Bavarian case law database notes that the 26 O 869/26 judgment is consequently without object.

Two further things limit its weight. It was interim relief, decided on the lower standard of credible substantiation, below the full proof a main action demands. And German courts are not bound by each other's rulings the way common law courts are, so even a surviving judgment would have been persuasive at best. It has been published in NJW and K&R and lawyers are writing practice notes on it, so the reasoning is being taken seriously. Publication is not authority.

Short answer: The reasoning is on the record and widely discussed. The outcome binds nobody, and no appeal court ever reviewed it.

Does any of this reach beyond Germany?

The claim itself ran on German constitutional and civil law, specifically the business personality right built from Articles 2 and 19 of the Basic Law with sections 823 and 1004 of the Civil Code applied by analogy. The court did decline to confine the injunction to German territory, on the reasoning that a decision given inside the Brussels I recast regime is recognised across the Union.

One strand of the reasoning is wider still. Google could not fall back on the liability privileges in the Digital Services Act, Regulation (EU) 2022/2065, including the hosting privilege in Article 6, or on the equivalent in the old E-Commerce Directive. Those privileges cover other people's content, and the overview text was the provider's own. The DSA applies across the whole Union, so a court in Belgium or Ireland facing the same defence would read the same regulation, from scratch, without a Munich judgment to point at.

Short answer: The substantive claim was built out of German law, and the DSA argument is available anywhere in the EU. Both would have to be won again elsewhere.

What can you do when an AI answer says something false about you?

Every route below existed before this case and survives its settlement. Nothing here is a right the Munich judgment created.

Capture evidence before it changes. The Munich file ran on the overview text as it appeared on a dated search, the exact query including the autocomplete term, emails carrying the panel as an attachment, and a sworn declaration. AI answers are regenerated constantly, so a timestamped capture is worth far more than a description of what you saw.

If you are an individual in the EU or EEA, your data protection rights already covered this before Munich. Article 16 of the GDPR gives you the right to obtain from the controller, without undue delay, "the rectification of inaccurate personal data concerning him or her". Article 17 covers erasure in defined circumstances. Both are free to exercise and the controller has to answer.

If you are a company, that route is closed to you. Google argued exactly that in Munich, that an Article 17 claim failed because the claimants were not natural persons, and the court agreed the regulation did not apply. Businesses fall back on national law. In practice, a lawyer.

Then use the product's own reporting channel and keep the paper trail. The Munich claimants complained on 2 February 2026, by letter and email and then through Google's own online form, and the court found Google had not shown that it or its AI had put the statements through any check at all. If nobody answers, every EU and EEA country has a data protection authority.

These routes all act on an output. Getting one answer blocked or removed is a different problem from changing what a model absorbed, and almost nobody is offering you the second thing.

Short answer: Evidence, then a written request to the controller, then your data protection authority if you are a person or a lawyer if you are a company. None of it depends on this case.

What does it mean for the companies building AI?

The immediate consequence was operational. Google had to remove the contested content and prevent comparable statements about those publishers, on pain of court-ordered penalties, and the judgment put around 80 percent of the costs on Google. Heise read the decision as raising the incentive to run the feature in Germany with more conservative settings, with limits on sensitive subjects like people, companies and health.

Heise traces the reasoning back to a September 2025 decision of the Frankfurt am Main Regional Court, case 2-06 O 271/25, which accepted that liability for AI overviews is not ruled out in principle while rejecting the injunction sought there.

Google's position, given to The Decoder in June 2026, was that AI Overviews are "designed to reflect the information that exists on the web", that they can occasionally miss context in the way ordinary search results can, and that the company was "carefully reviewing this decision, which is not yet final". The durable signal for anyone shipping a generative feature sits underneath the argument: the more a product rewrites its sources into fresh prose, the harder it is to say the sentences belong to someone else. That observation does not need the judgment to survive.

Short answer: One first instance court found that the "we only display other people's content" defence sat badly with a feature that writes its own sentences. The settlement took the ruling away and left the question open.

Where does MultiChats stand on this?

We are YUKARI LABS, a Belgian company. On data, the accurate statement is layered: we do not train on your conversations, what the provider behind a given model does with them varies by model, and the models where we know of a difference carry a notice on their info card in the picker. Models routed through OpenRouter carry the line "Routed via OpenRouter to third-party providers that may briefly retain requests; they do not train on your data", while Grok's card says plainly that "xAI may use your conversations to train Grok models."

On wrong answers about a real person or business, we will not overstate what we have. There is a Report control on any message, on the web app and on Android and iOS, with reasons including "Spam or misleading", and it puts the message in front of our team. For data rights, our privacy policy names YUKARI LABS as the controller and gives you an address for access, correction and deletion requests.

Short answer: Reporting gets a bad answer to a human here. Correcting what a model believes about you is something we cannot promise through a report alone.

What we read