A legal brief for people who run hotels, written by someone who is not a lawyer, about a machine that is not a person, which said something that was never true, citing sources that never said it.
In re: The Machine That Said It
Regional Court of Munich I, Case No. 26 O 869/26, decided May 28, 2026. Preliminary injunction granted. Appeal expected. Adult supervision pending.
Question Presented
When an AI answer engine tells millions of people that your business is a scam, and that claim appears in none of the sources it cites, and no human being anywhere ever wrote it, who said it?
For thirty years the internet’s answer has been: nobody. The platform just pointed. The publisher just published. The user just clicked. Everyone was a conduit and no one was a mouth. It was the greatest liability disappearing act in commercial history, and the whole industry agreed not to notice, because noticing was bad for the stock price.
A court in Munich just noticed. Out loud. In writing. With costs.
The court’s reasoning, compressed to bar-napkin length: Google built the thing, Google runs the thing, Google alone controls the algorithms the thing operates with. So when the thing speaks, Google speaks. The machine is not a witness. The machine is a mouth, and the mouth has an owner.
Now, the lawyers will remind you, correctly, that this is a preliminary injunction, which is legal Latin for “we’ll do this again later, longer, at greater expense.” Fine. Noted. The modifiers will get two years of billable attention. The sentence stands anyway, and Munich is the loudest any court has yet said it: the mouth has an owner. Everything else is footnotes. But the footnotes are where hotels live, so let’s read them.
Statement of Facts
Exhibit A: Munich
Two German publishers, Verlagshaus24 and a subsidiary, discovered that Google’s AI Overviews had been telling searchers they were known for scams, subscription traps, and dubious business practices. The AI had confused them with genuinely shady companies and then, in the confident complete-sentence voice these systems use for everything from pancake recipes to felony accusations, presented the confusion as fact.
Now here is the detail that should keep every marketing director staring at the ceiling: per the court’s findings, the claims appeared in none of the linked sources. Not one. The court found the AI made statements that were not made anywhere in the search results it cited. No blogger wrote it. No reviewer wrote it. No competitor, no ex-employee, no guy in a basement wrote it. The machine composed it. Original work. Never before published. World premiere, top of the world’s largest search engine, admission free, corrections unavailable.
The publishers sent a cease-and-desist in early February 2026, which is the legal system’s way of saying “knock it off.” The machine did not knock it off. On May 28, the Regional Court of Munich granted a preliminary injunction, which reporting indicates reaches across the European Union, and ordered Google to pay 80 percent of the costs. That is a lot of ground for one hallucination to cover, and somewhere in Mountain View, a very expensive lawyer updated a very long slide deck.
Exhibit B: Minnesota
Same movie, American subtitles. Wolf River Electric, a solar installer in Isanti, Minnesota, discovered that Google’s AI Overview was telling anyone who searched the company that it was being sued by the Minnesota Attorney General for deceptive sales practices.
The Attorney General had in fact sued four solar-lending companies. Wolf River was not one of them. Wolf River’s crime was appearing near the end of a newspaper article about the lawsuit, in a paragraph that did not accuse it of anything. Per the complaint, the machine read that article, connected dots that do not exist, and published the connection as fact, citing four sources for the claim. Zero of the four said it. That is not a summary. That is a rumor with a bibliography.
Then the invoices arrived, because in America the damages always arrive itemized. The complaint lays them out like a tab: a canceled $39,680 contract on March 3, 2025. A lost prospect on March 4. A dead $150,000 deal on March 5, when the customer attached a screenshot of the AI Overview as the reason for walking. Wolf River sued in Minnesota state court in March 2025, claiming $24.7 million in 2024 losses alone; its later initial disclosures estimated total damages between $110 million and $210 million.
Then came the procedural comedy. Google removed the case to federal court, where Section 230 defenses tend to get a warmer reception. In January 2026, a federal judge sent it right back to state court, because Google had missed the removal deadline. Wolf River’s own settlement letter, claiming $24.7 million, had put Google on notice months earlier, and Google’s lawyers filed too late. A company whose entire product is knowing things on time did not know a thing on time. The case is now back in Ramsey County District Court, pending, unresolved, and generating the one thing litigation reliably generates, which is more litigation.
Two continents. Two businesses. Two claims of falsehood that, on the record so far, existed nowhere on Earth until a synthesis engine assembled them from parts. One court has preliminarily agreed. One court hasn’t ruled. Put that fact pattern in your pocket. It is the whole case, and it is coming up again.
The Holding
German law, like American law, has generally treated search engines as pointers, not speakers. Germany’s Federal Court of Justice gave search and autocomplete limited liability on the theory that they merely make third-party content findable. Reasonable enough. You cannot ask the card catalog to vouch for every book, because the card catalog did not write the books. It just knows where they are.
The Munich court looked at an AI Overview and made the obvious observation the industry has spent two years paying people not to make: this card catalog talks. It does not point at statements. It makes independent, new, substantive statements by evaluating and combining content from many sources. The output is a self-contained claim, understandable on its own, delivered in the operator’s product, generated by the operator’s algorithms. That is not a librarian. That is an author. And courts have a very old habit of making authors answer for what they publish. The habit predates electricity. It will outlast the pitch decks.
Google’s defense was the disclaimer defense, the industry’s favorite piece of fine print: users can click the links and check, and everyone knows not to trust AI blindly. Think about the architecture of that argument. We built a product whose entire value is that you do not have to click anything, we put it above everything you might click, we wrote it in the calm voice of a fact, and our legal position is that you should not have believed it. The product is the answer. The defense is that it isn’t. Both of these are Google’s positions, held simultaneously, presumably in different buildings. It is Schrödinger’s defense: the Overview exists as both the answer and a disclaimer at once, and the state only collapses when a court opens the box. Munich opened the box.
The court declined to play along. In its reasoning, the possibility that a statement can be disproven through further research does not exempt the author from liability for making it. The judges drew a parallel to press law, where a publisher is liable for a misleading teaser that stands on its own, even if no reader ever opens the full article. Same principle here, except the teaser is the whole product.
And the behavioral data sides with the judges. A Pew Research Center analysis published in July 2025 found that users clicked a source link inside an AI Overview on roughly one percent of visits. One percent. Nearly everyone else takes the answer and moves on, exactly as designed. The court even noted the trap Google built for itself: if the Overview were generally treated as unreliable and everything had to be independently verified, the feature’s asserted utility would collapse. You cannot market a product as the answer and defend it in court as a suggestion. Pick one. Munich picked for them. Preliminarily. Which is how everything gets picked, the first time.
Necessary caveats, stated plainly, because precision is the whole point of publishing this: this is a preliminary injunction, not a final judgment. Its formal reach is a matter for enforcement lawyers, not headline writers. Google says it is reviewing the decision, which is not yet final, and an appeal is expected. A Frankfurt court, in a competition-law dispute in September 2025, accepted a related theory but denied the injunction sought there, so even the German courts are not yet singing in unison. This is a marker, not settled doctrine. Markers matter anyway. Every doctrine you have ever heard of started as one judge somewhere refusing to pretend.
The American Problem
In the United States, the question runs into Section 230 of the Communications Decency Act: twenty-six operative words written in 1996 that immunize platforms from liability for information provided by another information content provider. Twenty-six words. The Gettysburg Address needed 272 words to bind a nation’s wounds. Congress needed 26 to make sure nobody would ever be responsible for anything online again. Yes, there are exceptions. The exceptions have exceptions. The twenty-six words are still undefeated.
The statute made sense for the world it was written in. Every harmful statement had a human author somewhere. Sue the author, not the host. The host is just a wall; go find who held the spray can. That was the deal, and for thirty years it mostly worked, because there was always a hand on the can.
Generative AI walks in and, the argument goes, breaks the deal’s central assumption. When the claim was synthesized by the machine and exists in no underlying source, there is no author to sue. Follow the logic to its destination: if Section 230 covers the synthesis, then nobody is liable. Not the platform, because immunity. Not a third party, because there isn’t one. The statement floats free, doing eight figures of alleged damage, owned by no one. Call it the accountability gap, which is polite language for “the money escaped.” A machine can now say your hotel has bedbugs, cite four articles that never mention bedbugs, cost you a season of bookings, and the answer to “who pays for that” may genuinely be a shrug in a robe.
Honesty requires reporting the candid expert read, and it cuts against the plaintiff. William McGeveran, dean of the University of Minnesota Law School, has described AI Overviews as a souped-up version of Google’s search algorithm and predicted a judge would likely find Section 230 applies under current law. He also acknowledged the discomfort in the same breath: the Overview reads like something Google wrote, because functionally it is. And the defense travels; Section 230 is a federal statute Google can raise in Ramsey County just as easily as in federal court, deadline permitting. There is little to no directly on-point precedent on generative AI and Section 230. The question is genuinely open, which in law means the question is expensive.
So the current scoreboard reads: one German court says the machine’s owner is the author, preliminarily, with an appeal loading. America has not yet decided who, if anyone, answers for the machine’s output. And the machines keep talking either way, because the machines do not read case law. They summarize it, occasionally incorrectly, with citations.
Analysis: The Author With No Name
Here is where this stops being a story about two lawsuits and starts being a story about your hotel.
An analysis by the AI startup Oumi, conducted for The New York Times across 4,326 searches and published in April 2026, found that Google’s AI Overviews answered correctly about 91 percent of the time under the current model. Sounds great. Ninety-one percent is an A-minus. Now do the multiplication at Google’s volume, and the 9 percent becomes millions of wrong answers on a rolling basis, each one delivered in the same confident voice as the right ones, because the machine does not have a voice for “I’m guessing.” Google called the study flawed, objecting that the benchmark was built from deliberately difficult questions, which is a fair methodological point and also the traditional response to arithmetic you did not commission.
But the second Oumi finding is the one to sit with: among the answers that were correct, 56 percent could not be supported by the sources the AI cited. And here is the part nobody puts in the keynote. Under the previous model, that figure was 37 percent. The upgrade made the machine more accurate and less checkable at the same time. It got better at being right and worse at proving it, which is the exact career trajectory of every witness a lawyer has ever had to prep.
Read that again, slowly. More than half the time, even when the machine is right, it cannot show its work. The citations are decoration. Parsley on the plate. Whatever the engine is doing, “summarizing the web” is a generous description of it. It is forming its answers by processes even its operator describes only in brochure language, and then speaking in declarative sentences to an audience that clicks through one percent of the time. We have built a global oracle that is usually right, rarely traceable, occasionally accusatory, and constitutionally incapable of doubt. Then we put it in front of every purchasing decision on Earth and called it search.
The travel industry has spent two years asking how to show up in AI answers. Munich just reframed the question. The AI answer is a published statement about your property. Authored, in the preliminary view of at least one court, by the platform. Assembled from whatever the machine could find, plus whatever it inferred from what it found, minus any obligation to distinguish the two. It is a statement you did not write, did not approve, cannot preview, and, per the data, cannot reliably trace even when it flatters you.
Your hotel almost certainly already has these statements in circulation. Statements about your rates, your renovation, your service, your ownership, your alleged bedbugs. Some are accurate. Some are Wolf River with a lobby. And the difference between the two is not entirely in your control. Nothing about these systems is entirely in your control; that is rather the point. But the controllable part is the quality, structure, and verifiability of the knowledge the engines formed those statements from.
This is the part the AI-visibility vendor class keeps missing while it sells rate feeds and booking widgets. A live-rate pipe into a chatbot controls your price card after the machine has already decided what you are. The deciding happens upstream, in the corpus, where the engine learns what kind of property you are, what you are known for, and whether the claims about you trace to anything at all. Munich just attached legal consequence to that layer, and here is a prediction, clearly labeled as one: it will not be the last court to try, because the authorless statement that harms with impunity is exactly the kind of loophole legal systems eventually get around to closing. Slowly. Expensively. After the damage. But they get around to it, and the businesses that assumed they wouldn’t are the ones in the case captions.
What This Means for Operators
Not legal advice. Operating advice. Three items, none of which require a lawyer, all of which are cheaper than needing one.
First, know what the machines say about you. Not once. Continuously. Wolf River found out from a customer’s screenshot stapled to a dead $150,000 contract. That is a monitoring program. It is the most expensive one on the market, and the alerts arrive as cancellations.
Second, demand traceability, from everyone. From the engines, from your agencies, and especially from anyone selling you an AI visibility report. If a claim about your property cannot be traced to a live, verifiable source, it is not intelligence. It is a hallucination with a logo on it. Nikhil Lai, a principal analyst at Forrester, reacting to the Munich ruling, predicted the rise of what he called defensible AI, where verifiability and traceability become more valuable than polish. He was describing a legal trend. He accidentally described a procurement standard. Adopt it, and watch how fast the vendor meetings get shorter.
Third, understand where the fight actually is. The statement the engine publishes is downstream of the knowledge the engine forms. You can sue, obviously; Wolf River did, and it even won the procedural skirmish when Google’s lawyers missed a deadline. Whether anyone can win the merits is a twenty-six-word question no American court has answered, and litigation is a fine way to spend three years and a renovation budget learning the answer. What you can do today is govern the inputs: publish structured, dated, verifiable, self-consistent information about your property across the surfaces these systems learn from, so that when the machine composes its confident little paragraph about you, the composition has somewhere true to come from. The machine is going to write about your hotel either way. The only question you control is what it read first.
Prayer for Relief
The plaintiffs in Munich asked the court to make Google stop saying things about them that were never true. The court agreed. Preliminarily. In one jurisdiction. Pending appeal.
So that is what victory currently looks like in the answer economy: a temporary order, in one country, requiring one machine to stop repeating one falsehood it invented about one company. Everything else the machines are saying, about every other business, in every other jurisdiction, remains in circulation. Uncited. Unclicked. And for the moment, answered for by no one.
The hearing is ongoing. You are in it whether you filed or not. The machine has already read your file.
Case references: LG München I, 26 O 869/26 (May 28, 2026), preliminary injunction, appeal anticipated. LTL LED, LLC d/b/a Wolf River Electric v. Google LLC, filed March 2025 in Ramsey County District Court, Minnesota; removed to the U.S. District Court for the District of Minnesota in June 2025; remanded to Ramsey County District Court in January 2026 after the court found Google’s removal untimely; pending, merits unresolved. All Wolf River damages figures are allegations from the plaintiffs’ filings, not adjudicated findings. Sources referenced: Pew Research Center analysis, July 2025; Oumi analysis for The New York Times, April 2026 (4,326 searches, SimpleQA benchmark; methodology disputed by Google); William McGeveran, University of Minnesota Law School; Nikhil Lai, Forrester; remand coverage, The Volokh Conspiracy, January 2026; translated Munich decision, Transparency Coalition. The author is not an attorney, and nothing here is legal advice. The machines discussed throughout were not available for comment, although several of them will confidently summarize this article anyway, citing sources that do not say what you are about to read.
The machine is going to say something about your hotel today. KFO is how you get a vote. The machine shows you citations. We show you results.
