For years, the unspoken contract between a search engine and its user was based on the concept of the pointer. You typed a query, and the engine provided a list of directions to other people's houses. If the house you visited contained a lie, the mapmaker was rarely held responsible for the contents of the home. But the interface is shifting. Today, millions of users no longer click through to the source; they read the summary at the top of the page and accept it as the final word. This shift from navigation to narration has finally collided with the legal system in Germany.
The Munich Ruling and the Death of the Intermediary
The Munich Regional Court recently issued a landmark interim injunction against Google, ruling that the company must be held directly responsible for the inaccuracies produced by its AI Overviews. The case centered on two publishing houses based in Munich that found themselves the target of AI-generated falsehoods. According to the court documents, Google's AI Overviews began disseminating claims that these publishers were involved in fraudulent activities and improper business practices. These were not mere errors in indexing; they were confident, structured assertions of misconduct.
The court's analysis focused on the mechanical nature of how AI Overviews function. Unlike a traditional search result that presents a snippet of text from a third-party website, the AI rewrites the information using its own linguistic structure and logic. In this specific case, the AI did not simply link to a suspicious site; it actively constructed a narrative. It began its response with a definitive claim that the companies in question had suspicious business practices, then proceeded to build a structured layout including summaries, red flags, and user tips. By organizing the information into a cohesive, authoritative argument, the AI ceased to be a directory and became an author.
Because the AI creates a new, independent statement rather than merely reflecting an existing one, the court defined these summaries as Google's own content. This distinction is critical. By classifying the output as original content, the court determined that Google is not just a host of information but a producer of it, thereby assuming full legal liability for the truthfulness of those statements.
The Algorithmic Gap in Legal Protection
The core of Google's defense rested on the traditional legal shield afforded to search engines: the principle of limited liability for intermediaries. Under this framework, a platform is generally not responsible for the content it hosts as long as it acts as a neutral conduit and removes illegal content once notified. Google argued that it should be protected under the Digital Services Act (DSA) as a hosting provider, which would allow it to utilize the notice-and-take-down procedure to resolve errors.
However, the Munich court rejected this argument, creating a sharp legal pivot. The judges ruled that AI-generated content does not qualify for the protections typically granted to human expression or neutral hosting. Because the output is the result of an algorithmic process rather than a personal belief or a curated collection of third-party links, it cannot be protected under the umbrella of freedom of expression. The court essentially decided that an algorithm cannot have a belief, and therefore, its output is a product, not a protected opinion.
Furthermore, the court dismantled the argument that users can simply verify the information by clicking the provided source links. The ruling noted that AI Overviews often create connections that do not exist in the original source material. In the case of the Munich publishers, the AI had conflated the plaintiffs with other companies that actually had suspicious records, inventing a link that no human reader could find by clicking the citations. The court also cited research indicating that a vast majority of users do not click through to the sources, treating the AI summary as the primary source of truth. This behavior transforms the AI summary from a helpful guide into a definitive statement of fact, removing the possibility that the user's own verification process mitigates the platform's liability.
This ruling effectively separates the legal status of a link provider from that of a generative AI provider. When a system evaluates, combines, and rewrites data to produce a new sentence, it is no longer an intermediary. It is a publisher. This means the notice-and-take-down system is no longer a sufficient shield for generative AI; the responsibility for accuracy now shifts to the front end of the process.
The implications of this shift are underscored by the technical volatility of the models themselves. While internal benchmarks might show high accuracy, the gap between a correct answer and a legally actionable lie is narrow. Analysis from Oumi indicates that while the Gemini 3 model may achieve a 91% accuracy rate, there is a staggering 56% rate of outputs where the source cannot be accurately traced. When over half of the citations are untraceable or hallucinated, the risk of producing defamatory content is not a technical edge case but a systemic feature. As courts begin to treat AI as a content producer, these percentages move from being engineering metrics to becoming legal liabilities.
