German court finds Google responsible for errors in its AI Overviews

Two publishers falsely accused of questionable practices

The case starts from a very concrete problem. For certain search queries, Google’s AI Overviews had linked two Munich-based publishing companies to scams, subscription traps and questionable business practices.

According to the court, the AI ​​had mixed up information about other companiesreally dishonest this time, with those of the complainants, and had established links which did not appear in any of the sources cited.

The two publishers had sent a formal notice to Google, which had not responded satisfactorily. The case therefore ended up before the Munich Regional Court, under file number 26 O 869/26.

Why AI Overviews are not just search results

The heart of the court’s reasoning lies in a technical, but decisive, distinction. A traditional search engine simply points to external websites: it makes third-party content accessible, but does not rewrite it. AI Overviews work differently since the AI ​​reformulates and evaluates the results “ in your own words and according to your own structure ”, according to the terms of the judgment.

In the case examined, the summary generated by the AI ​​opened, for example, with peremptory statements such as “ yes, this company is known for questionable business practices », before developing a structure built by the AI ​​itself: a summary, warning signals about the supposed scam, then advice to users.

The court also found that the AI ​​Overviews made claims that did not even appear in the underlying search results. None of the sources cited established a link between the plaintiffs and the questionable companies mentioned by the IA. For the judges, these are “ own statements » from Google.

Why existing case law does not protect Google

Google relied on case law established by the German Federal Court of Justice (BGH), which grants classic search engines and autocomplete functions limited liability. The BGH’s reasoning is based on the idea that search engine operators are only liable as indirect authors, since they are content to make content produced by third parties findable. Imposing a systematic verification obligation on them would threaten the very functioning of online research.

The Munich court ruled that this logic does not apply to AI Overviews. Unlike a traditional search engine, AI generates “ independent, new and substantive statements » by evaluating and combining content from several third-party sites. And only Google is able to verify these claims, if only by comparing the source websites with the claims it makes from them.

The judges also noted that the AI ​​Overview is “not absolutely not essential » to use the internet. Traditional search results already allow users to sort information; AI Overview is just an additional feature.

User verification argument rejected by judges

During the hearing, Google argued that users could check linked sources themselves to ensure the accuracy of the AI-generated summary. The company also argued that the public generally knows that AI-generated content should not be taken at face value.

The court rejected this argument. The simple possibility of refuting an assertion through additional research does not, according to the judges, exempt from the responsibility linked to this assertion. The AI ​​Overview was “ understandable in itself » and constituted “ a stand-alone statement, the content of which could be understood independently, without any reference to other possible interpretations or unreliable content “.

The judges relied on studies showing that users very rarely click on sources displayed in AI Overviewswhich reinforces their reasoning. They also drew a parallel with press law, where editors are responsible for self-understandable article headings, even if the reader never reads the full article behind it.

Finally, the court pointed out a worrying legal gap. If Google was only responsible for obvious violations, victims would have no real recourse in the face of false AI claims. The third-party sites that served as sources had themselves never made the statements in question: it was therefore impossible to pursue them, and under the old legal regime, it was also impossible to effectively pursue Google. This situation prevented Google from invoking the protections granted to hosts by the Digital Services Act, or from hiding behind the classic notification and removal procedure applicable to search engines.

Reduced free speech protection for AI

The court also addressed the issue of freedom of expression as applied to AI-generated content, with an unfavorable conclusion to Google. According to the judges, an opinion produced by artificial intelligence is “not pas the expression of a conviction acquired by the person expressing himself, but the result of an algorithm “.

Proposing an AI-assisted search tool is “ above all an expression of Google’s commercial activity ”, and at most “ secondarily from an interest in being able to freely express an opinion or convictions “. In the balance between the rights of the plaintiffs and the interests of Google, the company therefore had to bow, especially since the contested assertions were based on inaccurate facts. The IA had associated the plaintiffs with companies that, according to affidavits, had absolutely no connection with them.

Google condemned on almost all points

The court ruled in favor of the plaintiffs on the majority of the claims. It banned Google from spreading claims about scams, links to dubious companies, subscription traps, fictitious phone calls or lack of availability. Only two minor requests were rejected.

The risk of repetition of the offense was considered persistent, even if the specific texts in dispute were no longer displayed at the time of the judgment. Google had not made a cease-and-desist statement with a penalty clause, and there was nothing to stop the algorithms from generating the same claims again in the future. Google must support 80% of legal coststhe plaintiffs sharing the remaining 10% each. According to the court, the scope of this decision could extend beyond German borders.

A technical precision that changes the legal situation

The Munich judgment goes far beyond the scope of this single case. An analysis carried out by the AI ​​startup Oumi for the New York Times showed that Google’s AI Overviews, working with the Gemini 3 model, responded correctly 91% of the time. A rate sufficient for typical daily use, but which still represents, on Google scale, millions of incorrect responses every hour.

Oumi’s analysis also found that 56% of Gemini 3’s correct answers could not be confirmed by the sources Google itself cited. AI therefore provides answers whose exact origin the user cannot trace.

It is precisely this problem that the Munich court resolved: the AI ​​makes its own assertions, which do not appear in any cited source, and the operator must answer for them in court. It remains to be seen whether this reasoning will stand up to appeal. Google confirmed its intention to challenge the decisionconsidering that the case concerns “ specific, limited errors, not the fundamental way AI Overviews displays web content “.

If the court’s reasoning were to take hold beyond Germany, the consequences could affect not only Google, but all AI providers whose systems reformulate content from the web, starting with ChatGPT, Claude or Perplexity.