Flarepoint

Munich Court v Google: AI Overviews Ruled Google's Own Words

2026-10-06

On 28 May 2026, the Munich Regional Court I granted two Munich publishers a preliminary injunction barring Google from repeating claims in its AI Overviews that tied them to fraud schemes and dubious business practices, holding that the generated summaries are Google's own content rather than a display of search results. The decision is not final and Google has said it will appeal. It matters because a European court has treated an AI search summary as the operator's own words, and because the publishers built their case on a dated record of what the system said before and after Google was told.

What did the AI Overview say?

According to PPC Land's account of the judgment, a search on 20 January 2026 for the first publisher's name with the German word "Betrugsmasche" (fraud scheme) returned an AI Overview that opened by affirming the publisher was known for disreputable business practices. Legal Tribune Online reported that the summary went on to list supposed features of the scheme, mention subscription traps and debt-collection demands, and advise readers to consult a lawyer. A second search on 26 January 2026 returned a similar answer.

The court found the system had mixed up the publishers with other, genuinely questionable companies and drawn connections that appeared in none of the sources it linked. The publishers submitted sworn affidavits that they had no connection to those companies.

The court's press release does not name the two publishers, and neither does this article. Printing their names next to the word "scam" would repeat the pairing the case was brought to stop.

What did the court decide?

The 26th Civil Chamber heard the case under reference 26 O 869/26. According to the court, Google argued that the claim failed because the publishers were not natural persons, and that it was not liable because it did not adopt third-party information as its own. The chamber disagreed on both points: German speech law protects the personality rights of companies, and an AI Overview that summarises and evaluates results in its own words creates independent statements for which the search operator is responsible.

The court described these as "independent, new, and substantive statements". It rejected Google's argument that users could check the linked sources themselves, drawing a parallel with press law, where a publisher answers for a headline that is understandable on its own even if readers never open the article. It also refused Google hosting-provider protection under the EU Digital Services Act, because Google's own system generated the content.

LTO notes that the court treated the "fraud scheme" sentence as an opinion, but held that an opinion built on demonstrably false facts carries little weight against a company's reputation. Breaches of the injunction can draw fines of up to €250,000 each, and Google was ordered to pay 80 percent of the costs.

Why did it matter that Google had been told?

Per PPC Land's reading of the judgment, the publishers' lawyers sent a formal warning letter on 2 February 2026. Google's complaints address replied the same day that it could not process the complaint and pointed them to an online form, where it was resubmitted. A search on 10 February 2026 still returned a substantially similar answer. The court found Google had had the chance to act and had not.

Nor did the case end when the text stopped appearing. The Decoder reports that the court found a continuing risk of repetition: Google had not signed a cease-and-desist declaration with a penalty clause, and nothing stopped the system generating the same statements again.

In the US, on 14 September 2026, Judge Thomas Durkin let most of author James Keene's defamation claims over AI Overviews proceed, reasoning that Keene's legal removal report either reached someone at Google, or "no human at Google reviewed" it. We covered that ruling in the ruling that turned correction emails into evidence. Different law, same building blocks: a dated output, a dated notice, and a dated check afterwards.

Is this settled law?

No. Google told reporters: "This case focuses on specific and narrow errors, not the foundational way AI Overviews displays web content. We disagree with the ruling and plan to appeal." The Decoder reports that a Berlin court reached the opposite conclusion in early June, in a different context, treating AI Overviews as a form of search result with only limited liability.

Writing in Lawfare on 22 September 2026, Oxford law professor Ignacio Cofone stresses that this is a first-instance preliminary injunction with no binding force over other courts. He also argues that treating generated output as the provider's own statement could carry over into data protection and consumer protection claims.

What this means for lawyers and comms teams

The Munich publishers had a record made at the time. Many companies learn of a false AI answer only from its consequences: in Wolf River Electric's case against Google, the complaint cites a customer who terminated a $150,000 contract after raising Google's claim.

  • Record the query, not just the answer. The Munich searches paired a company name with "fraud scheme", the exact search a cautious customer or counterparty runs. The wording shaped the output.
  • Compare the answer with its citations while both exist. The court relied on the finding that the linked sources did not say what the summary said. That comparison cannot be rebuilt once the answer changes.
  • Use the channel the provider designates, and keep proof. Google redirected a lawyer's letter to a web form on the same day. Both submissions became part of the timeline.
  • Check again after notice. The 10 February search is what showed the claim survived Google being told.

None of this depends on how the appeal comes out. Courts in Munich and Chicago have now both shown that a claim about an AI answer rests on dated evidence: what was said, on which query, when, and what changed after the operator was told.

Sources