How Google Lost AI Immunity in German Court

Google just lost immunity for AI mistakes. On June 13, 2026, the Munich Regional Court issued a preliminary ruling that Google is liable for false statements generated by its AI Overviews feature — a decision that could fundamentally reshape how every AI company on the planet operates. Two publishers discovered that Google’s AI-generated summaries linked them to scams, dubious business practices, and subscription fraud with zero factual basis, and the court agreed: when your AI invents claims that don’t exist in any source, you own them.

How the Case Began: AI That Invents Defamation

The case started when two publishers noticed something alarming in Google search results. When users searched for their companies, Google’s AI Overviews confidently stated that they were “known for dubious business practices” and “often perceived as a scam.” The problem? No source on the internet ever made those claims about these specific publishers.

Google’s AI had combined information about other companies flagged for illicit practices with data from the plaintiffs, creating false associations through what the court called “a misinterpretation of information available on the internet.” The AI essentially hallucinated a connection between innocent businesses and fraudulent activity, then presented that hallucination as fact at the top of search results.

Earlier this year, the affected companies sent Google a cease-and-desist letter. Google denied liability, arguing that its AI Overview feature warns users that information may contain errors and should be independently verified. The court was not persuaded.

Step 1: The Court Asks — Is AI Overview Google’s Own Content?

The first question the Munich court tackled was fundamental: does an AI-generated summary count as Google’s own content, or is it just a neutral repackaging of third-party information?

The court examined whether Google’s LLM actively generates new text rather than passively linking to sources. Traditional search engines display lists of blue links with brief snippets — exact excerpts from third-party pages. AI Overviews, by contrast, use Gemini’s retrieval-augmented generation (RAG) system to synthesize information from multiple sources into novel sentences that never appeared in any single document.

The court found that Google’s tool produced “independent, new, and substantial statements” based on its own interpretation of web content. That distinction — generating new text versus reproducing existing text — was the foundation of the entire ruling.

Step 2: The “Zu Eigen Machen” Test

German law has a well-established concept called zu eigen machen — literally “to make something one’s own.” When a platform adopts third-party information as its own by presenting it without clear source attribution, it assumes legal responsibility for that content.

The court applied this test to AI Overviews and found that Google was doing exactly that. The synthesized answers appeared at the top of search results in a format that users naturally interpret as Google’s authoritative response. There were no individual source links for each claim in the summary — just a general attribution to “the web.” The court concluded that Google had made the AI’s output its own.

Step 3: Peering Inside Gemini’s RAG Architecture

The judges didn’t stop at surface-level analysis. They evaluated the technical architecture of Gemini’s retrieval-augmented generation system to determine whether grounding in search results qualifies as mere indexing or content creation.

The distinction matters. In traditional search, Google’s crawlers index pages and rank them, but the content displayed belongs to the original publishers. With RAG, the system retrieves documents, passes them through a language model, and generates entirely new text that reflects a weighted interpretation of those sources. The model doesn’t copy — it creates. And creation, the court found, carries creator liability.

Step 4: Snippet Extraction vs. Generative Summarization

The court drew a sharp line between two technical approaches:

Traditional snippet extraction copies exact text from a source page and displays it verbatim. If that text is defamatory, the original publisher is liable — not the search engine displaying it.

Generative summarization transforms multiple sources into novel sentences that no single source ever contained. When those novel sentences are false or defamatory, the original publishers can’t be held responsible because they never wrote them.

This distinction was devastating for Google’s defense. The court noted that the challenged AI Overview “contains statements that do not appear at all in the search results.” You can’t sue a third-party publisher for content they never published. Only Google controls the model that produced those statements.

Step 5: The Telemedia Act Removes the Shield

Under German Telemedia Act Section 7, paragraph 1, liability exemptions for intermediaries are removed when platforms create or select information for presentation. The court applied this provision directly to AI Overviews.

For 23 years, Google had operated under the E-Commerce Directive’s intermediary protections in Europe, which shield platforms that merely transmit or store third-party content. The court’s ruling effectively strips that protection when a platform’s AI generates new statements rather than displaying existing ones.

This has enormous implications beyond Germany. The EU’s Digital Services Act Article 16 requires platforms with over 45 million monthly users to implement risk mitigation systems, but the German ruling establishes stricter liability standards than the DSA’s safe harbor provisions. If AI-generated content triggers content provider liability under national law, the DSA’s platform protections may not save you.

“Nobody needs AI to search the Internet, so AI firms can’t just let their tools attribute false claims to fake sources without assuming any liability.” — Munich Regional Court

Step 6: Editorial Control Means Editorial Responsibility

The court examined Google’s editorial control mechanisms, including reinforcement learning from human feedback (RLHF) and constitutional AI filters that shape output content. Google designs the training data, tunes the model, sets the safety filters, and controls the presentation format. The AI’s output is, in the court’s view, “primarily an expression of the defendant’s commercial activity.”

This framing is significant. It rejects the notion that AI outputs are some emergent, uncontrollable phenomenon. Google chose to deploy this technology. Google chose to place AI Overviews at the top of search results. Google chose the model architecture and the training data. With those choices comes accountability.

Step 7: User Perception Matters More Than Disclaimers

Google argued that its AI Overviews include warnings about potential inaccuracies, and that users should verify information independently. The court dismantled this defense on multiple fronts.

First, the court noted that the AI tool’s utility “would be significantly diminished if the ‘AI overview’ were generally regarded as unreliable and if every single displayed link required independent verification.” Google can’t simultaneously claim the feature is useful and that users shouldn’t trust it.

Second, a Pew Research survey from July 2025 showed that most people don’t click on AI Overview source links. A May 2026 analysis published by The New York Times found that AI Overviews with the current Gemini 3 model are inaccurate about 9 percent of the time and include inaccurate source links about 56 percent of the time. These findings suggest Google’s AI tool may be cranking out millions of wrong answers daily, with few users verifying the information.

Third, the court held that AI-generated results cannot be protected under principles of free speech, as they are “the product of an algorithm designed, trained, and managed by a company, and not the expression of an individual opinion.” When an AI hallucinates a defamatory statement, that’s not speech — it’s a product defect.

Step 8: A Precedent That Touches Every AI Company

The court’s final analysis establishes a precedent with global reach: generative AI outputs trigger content provider liability when the platform controls the model, training data, and presentation format.

This doesn’t just affect Google. Companies like OpenAI, Anthropic, and Perplexity AI all warn users that responses may contain errors and recommend verifying information. Under the Munich court’s reasoning, those disclaimers are insufficient. When your AI generates a new statement that doesn’t appear in any source, you are the publisher of that statement — and you bear legal responsibility for its accuracy.

The ruling required Google to remove the defamatory statements, prevent their recurrence, and cover 80 percent of the legal costs. Google’s spokesperson said the company is “carefully reviewing this decision, which is not yet final,” suggesting a likely appeal.

Why This Ruling Could Change Everything

The implications extend far beyond one German court. If other jurisdictions adopt similar reasoning:

AI search engines would need to treat every generated summary as a published statement subject to defamation law, potentially requiring pre-publication review of AI outputs.

Chatbot providers could face liability for any false or defamatory claim their models produce, even when users explicitly ask for the information.

Content platforms that use AI to generate recommendations, summaries, or responses would need to reassess whether their intermediary protections still apply.

Insurance markets would need to develop new liability products for AI-generated content, similar to media liability insurance that traditional publishers carry.

The Munich ruling represents the first time a court has held an AI company liable for the speech its model generates. It won’t be the last.

But here’s what the ruling didn’t address: who’s liable when the training data itself contains the false information.