When the AI Speaks, Is It Using Your Voice?
A pattern is emerging across courtrooms in three countries. Most organisations deploying AI haven’t noticed it yet.
I was researching AI model pricing earlier this year, looking at how it had changed and what the direction of travel looked like. I had read enough to have a working view, and when I asked an LLM to summarise what I had read, it told me something that flatly contradicted it.
I pushed back.
It told me I was wrong.
Eventually, I went back to the sources and found the specific URLs, copied and pasted them in. The AI looked at this new evidence and replied: “You’re right to push back on that, this is opposite to what I said.”
That sentence has stayed with me. Not because the AI hallucinated, that part barely registers any more. What stayed was what the sentence doesn’t contain. No account of how it happened. No acknowledgement that it had just told me I was wrong about something it was wrong about. One moment it was certain and the next it wasn’t.
Most people don’t push back. Many people don’t have enough time or prior knowledge of a subject to know when they should, and when they don’t, there is no correction. There is just the original statement, sounding confident and remaining uncorrected.
Courts in three countries are now working out whose problem that is. I am not a lawyer, and nothing here should be read as legal advice. But you don’t need a law degree to see something taking shape.
In late May, the Regional Court of Munich issued a preliminary injunction against Google after its AI Overviews feature made false and damaging statements about two local publishing companies. The AI had mixed up information about genuinely dubious businesses with the plaintiffs, producing confident assertions linking them to scams and shady practices. None of those assertions appeared in any of the sources the system cited.
Google’s defence, we are just surfacing what is out there. We are a conduit, not an author.
The court rejected it, ruling that AI Overviews produce independent, new, and substantive statements. They are Google speaking and not a list of links, so when Google speaks falsely about someone, Google is liable for what it said.
It also closed a secondary escape route. You cannot argue your product is valuable because it removes the need to verify sources, and simultaneously claim no liability because users could have verified the sources.
Google confirmed on 12 June that it will appeal. This is a preliminary injunction from a regional court, not settled law. And the legal picture in Germany is already complicated. Four days after Munich, the Berlin Regional Court dismissed a separate case against Google’s AI Overviews, brought by a perfume manufacturer whose brands were being mentioned alongside cheaper imitation products. The Berlin court found that Google does not present AI Overview content as its own statements and that a normally informed user would understand the text as a summary of third-party sources rather than something Google had authored.
The two rulings are not quite the contradiction the headlines suggest. Munich was a defamation case where the AI fabricated claims that appeared in none of its sources. Berlin was a trademark case where the AI accurately reflected content that did exist on the web. The Berlin court treated AI Overviews as a new search format, not as Google authoring original content. The Munich court treated them as Google’s own statements because the AI had gone beyond its sources and hallucinated. What matters for every organisation deploying AI is which side of that line their systems fall on.
And the Munich ruling did not arrive from nowhere.
In 2024, a Canadian tribunal ordered Air Canada to compensate a customer after its chatbot gave him false information about bereavement fares. The airline argued that the chatbot was a separate legal entity responsible for its own actions. The tribunal did not agree, noting the chatbot was part of Air Canada’s website. It made no difference whether the information came from a static page or from an AI. One detail worth noting: the chatbot’s response included a link to the correct bereavement policy page. It still told the customer the opposite of what that page said.
In March 2025, Wolf River Electric, a solar company in Minnesota, sued Google after AI Overviews told anyone searching for the business that it was facing a lawsuit from the state attorney general for deceptive sales practices. The Minnesota AG had sued four solar companies but Wolf River was not one of them. Google’s AI stitched together sources that mentioned the AG action, none of which mentioned Wolf River, and produced a confident assertion that the company was implicated. Customers started cancelling orders. In early March 2025 alone, the company lost contracts worth hundreds of thousands of dollars. Wolf River is seeking between $110 million and $210 million in damages. That case is still ongoing.
On 12 May 2026, two weeks before Munich, the Higher Regional Court of Hamm in Germany ruled that a cosmetic medicine clinic was liable under unfair competition law for its AI chatbot falsely claiming the doctors held specialist qualifications they did not have. The clinic argued it had not published the falsehoods deliberately and had not intended to mislead. They said they had fed the system accurate data. The court rejected all of it. Intent and the accuracy of the inputs were both irrelevant. The chatbot’s output was the company’s output. The court also found that general disclaimers along the lines of “AI can make mistakes” do not reliably protect against liability.
Accurate inputs, false output, full liability.
In each of these cases, the deployer made the same argument: the AI is a separate thing, not our words, not our responsibility. And in each case where the AI had generated content beyond what its sources actually said, courts rejected that argument. Where courts have found the AI faithfully summarised existing content, as in Berlin, the deployer has been treated as a search engine. Where the AI fabricated or invented, the deployer has been held responsible for the output. These cases span different legal systems and different theories of liability, and I would not claim they represent settled law anywhere. But where courts have considered the question, the direction is clear enough to pay attention to.
This is not just about search engines or defamation claims.
Think about the HR platform that summarises a performance record, or the customer service system that explains what a policy covers, or the internal knowledge tool that answers a staff question by pulling together content from across the organisation. Every one of these is generating output that carries the deploying organisation’s authority.
None of these are returning documents. They are generating conclusions. In most cases, the people reading those conclusions have no reliable way to tell whether they are accurate. An analysis conducted for the New York Times found that even when Google’s AI Overviews gave correct answers, more than half could not be verified through the sources cited. If systems are routinely arriving at conclusions their own evidence does not support, that is not a Google-specific problem. It is how these systems work.
These systems are useful precisely because they do more than retrieve. Courts are not saying synthesis is wrong. They are saying synthesis is authorship, and authorship has consequences.
The OLG Hamm case is the clearest example. The clinic gave the chatbot correct data. The chatbot hallucinated qualifications the doctors did not hold, but the company was still liable. That is not a ruling about how carefully you built the system. It is a ruling about who owns what it produces.
Most organisations deploying AI have not seriously asked who is responsible when the system gets something wrong. The absence of legal consequence made it easy not to, but that is starting to change.
The answer my AI gave me, “you’re right to push back on that”, assumed the pushback was mine to do. That I knew enough to notice and had time to go back and check. Most users of most deployed AI systems have none of those things. In every deployment where nobody pushes back, the original statement stands.
It will take time for the law to catch up with what AI systems are actually doing. These are early cases in different courts and under different legal theories, and we are likely years away from clear precedent on how courts will assign responsibility for harm caused by AI. But every case so far has been asking the same question, and it is one you can answer now without waiting for a court to answer it for you. What is your AI system asserting? Who checked it before it reached the person reading it? And if it is wrong, whose name is on it?
You already know the answer to the last one.
I write about AI, cybersecurity, and technology every Friday. Subscribe to get it in your inbox.
Sources
Regional Court of Munich I, preliminary injunction against Google re AI Overviews (Az. 26 O 869/26, 28 May 2026). Reported by The Decoder, Reuters, and Der Spiegel. Full translated decision published by Transparency Coalition.
Regional Court of Berlin II, perfume manufacturer trademark claim dismissed (Az. 52 O 62/26 eV, 1 June 2026). Reported by Kanzlei Plutte and blogspan.net.
Moffatt v. Air Canada, 2024 BCCRT 149, Civil Resolution Tribunal, British Columbia (14 February 2024).
Wolf River Electric (LTL LED, LLC) v. Google LLC, filed Ramsey County District Court, Minnesota (March 2025). Removal to federal court June 2025; remanded to state court January 2026 (Judge Jeffrey Bryan). Reported by Minnesota Star Tribune, Politico, and Reason (Volokh).
Higher Regional Court of Hamm (OLG Hamm), Case No. I-4 UKl 3/25, cosmetic medicine clinic AI chatbot liability (12 May 2026). Appeal to Federal Court of Justice (BGH) permitted. Reported by Library of Congress Global Legal Monitor, DLA Piper, and IR Global.
Oumi / New York Times analysis of Google AI Overviews accuracy using SimpleQA benchmark (April 2026). Gemini 2: 85% accurate; Gemini 3: 91% accurate. Ungrounded rate for correct answers: 56% (Gemini 3), up from 37% (Gemini 2).
Google search volume: Google internal figures, reported as “over 5 trillion annual searches.”
AI Overviews trigger rate: approximately 48% of searches as of March 2026. BrightEdge research.