Google didn't say it. Its AI did. A German court says that's the same thing.

Picture the pitch meeting. Google builds an AI that reads the web, decides what it means, rewrites it in confident prose and staples that prose to the top of the most visited page on the internet. Then, when said prose defames somebody, the same company explains that it is merely a humble librarian pointing at other people's books. A court in Munich has just looked at that argument and declined to buy it.

On 28th May 2026, the Regional Court of Munich I granted an injunction against Google (case 26 O 869/26). The AI Overview had told searchers that two German publishers, Verlagshaus24 and its GeraMond imprint, were mixed up in scams, subscription traps and unscrupulous debt collection. None of it was true. The reason Google are in trouble: none of this appeared in the sources the overview itself cited. Google's AI had blended the publishers with genuinely shady firms and invented the connection from scratch. That’s defamation.

The case was first reported by The Decoder, the full translated ruling is up at the Transparency Coalition, and Wired and Ars Technica are among the many notable sources to have covered it.

The sentence that did the damage

The old rule was comfortable for everyone: a search engine is a noticeboard, not the author of what's pinned to it. The court agreed that this still holds for the ten blue links. It then said an AI Overview is not a noticeboard. Because the AI "independently compiles" the sources and summarises them "in its own words," the output is, in the court's phrasing, Google's "own statement made by its own AI."

That one move is the whole story. It strips away the intermediary safe harbour the entire industry has been resting on. The "AI can make mistakes, please verify" disclaimer? The court ruled that a warning does not cancel liability, partly because the bot failed to back itself up with sources, so there would be nobody else to sue. And the AI's output does not count as protected speech, because it is the product of an algorithm, not a person with an opinion. The penalty runs to €250,000 per breach, or detention for the directors, plus 80% of costs. Google calls the decision "not yet final" and is "carefully reviewing" it, which is corporate speak for "see you on appeal."

Right ruling, messy precedent

Ultimately, the ruling is correct, and the precedent is a mess.

Correct, because you cannot have it both ways. Nick LeRoy made this point well this week, and his #SEOForLunch newsletter is one of the few genuinely worth the inbox space: you don't get to claim the productivity gains when the AI is right and blame the algorithm when it's wrong. That is the entire principle in one line.

Messy, because "summarises in its own words" is not some narrow defect. It describes every AI Overview ever served, on roughly half of all searches depending on whose tracking you believe and what they happen to be selling. A test that broad is more like a tap left running than a mere leaky pipe. And Google is not going to re-engineer the model to turn it off.

You don't get to sell the answer as authoritative and disown it the second it's wrong. Pick a lane.

There is a grim symmetry to it, too. Those same overviews have spent a year happily draining the open web: where one appears, organic click-through on the top result falls by around a third, and close to seven in ten Google searches now end with no click to anyone at all. Google took the traffic first. Munich has just handed it the liability to match.

Watch the big G weaponise this

If you’ll excuse me for being cynical, and don’t expect anyone in Mountain View to back me up here, but I think this will put Google on the defensive. The cheapest answer to "your AI keeps saying questionable things" is not better AI. It is less AI, applied selectively, and presented as if turning it off is a favour we should be grateful for.

Google would sooner switch the answers off than stand behind them, and it will sell you that retreat as a favour.

What I’m waiting to see next. First, geo-fencing: AI Overviews get thinner, more hedged or simply absent in Germany and then the wider EU, marketed as regional refinement rather than legal retreat. Second, the opt-out as alibi: Google already lets some users turn AI Overviews off, and I'd expect that toggle to be pushed harder and framed as respect for your preferences, when its real job is to move the liability out of Google's column and into yours the moment you choose to leave it on. A deliberate choice, for the good of users, that happens to be a liability get out. Third, the copycats. An injunction in Munich is catnip for every business that has watched an AI Overview describe it badly, and this was a corporate-personality-rights case brought by companies, not a privacy claim brought by an individual. However the door is now open to brands and people (or those who think they’re both). Let the floodgates open.

Are the chatbots next?

The obvious follow-on is whether this reaches ChatGPT, Gemini-the-chatbot , Claude and Perplexity, all of whom shelter under the same "may contain errors" disclaimer. The logic transfers cleanly enough: if generating a fresh, unsourced claim is what creates the liability, a chatbot does precisely that, all day.

I wouldn't hold my breath, though. A standalone LLM chatbot is a tool you choose to open, behind a login, with the warning right in front of your face. An AI Overview is pushed at you on a page you visited for something else, presented as the answer, wearing Google's authority. Courts care about that difference: about reliance, prominence and whether a reasonable person treats the result as fact. So the chatbots are exposed in theory and reasonably safe in practice, for now. What closes that gap is not a new statute. It’s the chatbots becoming more confident and less optional, which is exactly the direction in which they’re already marching off.

I tested it on the one brand I'm allowed to defame

If "describes a business badly" sounds like an edge case, I ran it on my own. Ask Google's own Gemini model about Dog on the Table with web access switched on and it is mercifully accurate: solo consultancy, Brighton, one person and a stack of AI tools doing the work of several. Take the web access away, and the same model relocates me to Athens, decides I was founded in 2013, invents a managing partner named George Vareloglou, hands me Samsung as a client, awards me a cabinet of Greek advertising trophies, and staffs me with "between 15 and 30 professionals." I’ve spent this whole piece telling you these systems invent confident nonsense and here the machine could not resist proving me right by building a fictional thirty-person Hellenic version of me. I am, you'll be glad to hear, working on correcting the record.

It is funnier when it’s me. It is less funny when it’s happening to a client. I’ve done some work with JustAir, who broker aircraft parts. Ask the bots what JustAir does and they will tell you, with total confidence, about air-quality monitoring in Detroit, with a glance at an air-conditioning firm in Arizona. The real company, the one with customers and invoices, comes third if it surfaces at all. Nobody defamed them. The AI simply served two other companies in their place and that was that. Being confidently described as the wrong entity is its own kind of damage, and there is no cease-and-desist letter for "you answered with someone else and called it us."

The usual declaration of interest applies, because I sell AI visibility and a court making AI answers a bigger problem is, conveniently, good for my enquiry pipeline. Apply the suspicion you'd apply to a locksmith warning you about burglars. But the burglars are real, and most business owners have never once checked the locks. You can find out what Google's AI, ChatGPT and the rest actually say about your business in an afternoon. If yours is being described by a machine that, on a bad day, thinks it's Greek, that's worth knowing before a customer reads it first.