The Regional Court of Munich (Landgericht München I) ruled on Tuesday that Google is directly responsible for false claims generated by its AI Overviews. The case is the first of its kind in Europe, and the reasoning is the part that should worry every company shipping a large language model (an AI system trained on huge text corpora to generate plausible prose) inside a search box. The court did not split hairs about training data, model architecture, or the question of whether the AI knew better. It said the answer was wrong, the answer was published as a search result, and Google is on the hook for the result.
I have been writing about AI search products for a couple of years, and I have seen plenty of hand-wringing about hallucinations (cases where a model produces confident, fluent, and entirely made-up answers). Most of the discussion has been about model behaviour, training data, or evaluation benchmarks. The Munich case is a different kind of problem. It is a product liability question, and the answer is going to be very uncomfortable for Google, Microsoft, Perplexity, and every other company that has wrapped a chatbot around a search engine.
The case in plain terms
A user searched for information that returned an AI Overview containing false statements about a person. The user sued. Google argued that the AI Overview was generated by a third-party model and that the company could not be held liable for the model’s output the same way a publisher would not be liable for what a printing press produces. The court disagreed. The ruling treats the AI Overview as a Google product. The search results page is Google’s product, the answer is Google’s answer, and the user has no realistic way to know that one of the snippets is human-curated and the other is generated.
The reason this matters is the framing. The court is not asking whether the model could have produced the right answer. The court is asking whether Google should have shipped the product in a way that conflated model output with factual search results. Under German product liability law, the answer to that question is no. The court is treating the AI Overview the way it would treat a misleading advertisement or a defective product manual. The user has no way to evaluate the model’s confidence, and Google has no business making them try.
Why Google built it anyway
Google did not have to ship AI Overviews. The company built them because OpenAI shipped ChatGPT and the search incumbents were scared. The internal calculus was simple: if users could get direct answers from a chatbot, they would stop clicking search results, and Google’s ad business would slowly leak. AI Overviews were the response. They keep the user on the search page, they keep the user looking at Google ads, and they keep the user from typing chat.openai.com into the address bar.
The cost is that Google is now publishing AI-generated text inside the product that has, for twenty-five years, been treated as a factual reference work. That is the product liability problem. The Munich court is saying, in effect, that you cannot move the goalposts without taking on the new liability. If your product tells the user something, you are responsible for whether it is true. The fact that a model produced the text is not a defence.
The same argument applies to Perplexity, to Bing’s Copilot answers, to DuckDuckGo’s AI assist, and to any other search product that has bolted an LLM (a large language model) onto a search engine. The Munich ruling is the first time a court has been explicit about this, but it is unlikely to be the last. Every other national court that looks at the question will be working from the same playbook. The defendant is a large technology company that publishes AI-generated text inside a factual product. The plaintiff is a user who reasonably expected the answer to be correct.
What changes for users
Short term, not much. Google is going to appeal the Munich decision, and the appeal will take years. AI Overviews are still live, still generating answers, and still occasionally making things up. The product surface is not going to change because of a single district court ruling.
Medium term, expect to see warning labels. The companies that ship AI answers inside search are going to start labeling them as AI-generated, hedging the answers, and adding disclaimers. The disclaimers will be legally motivated, not user-motivated. Once a court has established that AI Overview is Google’s product, the obvious mitigation is to make it look less like Google’s product. I expect to see the following changes become standard:
- “This answer was generated by AI” labels at the top of the snippet
- Confidence indicators that show the model’s internal probability that the answer is correct
- Inline source links for every factual claim, not a list at the bottom
- A “fall back to ten blue links” toggle for users who want the old product
- An opt-out for users who never want to see AI answers in their results
The product experience will get worse for users who liked the AI answers, but the legal exposure for Google will drop. Long term, the answer to the legal question is going to drive the product. If every national court follows Munich, the companies shipping AI search are going to either figure out how to make their answers reliably correct, or stop shipping the answers at all. Search is a high-stakes product. The companies that succeed are going to be the ones that treat the model output the way a publisher treats a submitted article: a draft to be reviewed, not a finished product to be published.
What changes for AI builders
If you are building a product that puts a model in front of users, the Munich ruling should change how you think about the surface you are building. The legal exposure is no longer limited to defamation or copyright. The legal exposure now includes product liability for any factual claim the model makes inside a product that presents itself as a factual reference.
The companies that are going to do well in the next few years are the ones that are explicit about what the model is and what the model is not. If the model is a chatbot, fine. If the model is generating product descriptions, fine. If the model is writing inside a search box that the user treats as a reference work, you have a problem. The Munich ruling is not going to be the last ruling. Expect to see similar cases in France, in the United Kingdom, in the United States, and in any other country where the user has standing (the legal right to bring a case) and the search engine has market share.
Trade-offs
The Munich ruling is good for users, in the sense that it shifts the cost of factual errors from the user to the company that published them. The cost is that AI search products are going to become less useful, less confident, and more hedged. The model that produced the false statement in the Munich case probably had a confidence score (an internal probability that the answer is correct) of around 70%. The product did not show the user the score. The court is going to require it to show the user, or to stop publishing the answer.
For the companies that ship AI search, the trade-off is between model confidence and legal exposure. The most useful AI search product is the one that gives the user the most useful answer. The legally safest AI search product is the one that gives the user the least useful answer, hedged with so many disclaimers that the user gives up and goes back to ten blue links. The companies that figure out how to do both are the ones that survive. The companies that do not figure it out either get sued into a defensive product posture, or get replaced by a competitor that is willing to take the legal risk for the better user experience.
For the user, the trade-off is the same one that has always applied to search. The user has to know when the answer is a fact and when the answer is a guess. Search engines have been telling users this for twenty-five years. AI search engines are going to have to start telling users, more explicitly, when the answer is a guess.
Bottom line
If you use AI Overview, Copilot, Perplexity, or any other search product that puts a model in front of a search box, treat the answer as a starting point, not a fact. The Munich ruling is going to make these products more cautious, more hedged, and more explicit about what they do not know. That is going to make the answers less useful for the cases where the model is right, and more useful for the cases where the model is wrong. The court is right. The product should not be shipped the way it is shipped. The fix is going to be painful for the companies and slightly better for everyone else.