A Reputational Wrecking Ball, and Google's Appeal
On May 28, 2026, the 26th Civil Chamber of the Landgericht München I handed down a final judgment (case 26 O 869/26) that Google's AI Overviews had falsely told searchers that Munich publisher Verlagshaus24 GmbH and its subsidiary GeraMond ran fraud schemes and subscription traps — claims that appeared in no linked source at all, effectively hallucinated by the system. The court didn't just order Google to stop repeating the claims. It ruled that AI Overviews are Google's own statements, not a neutral list of links, and that Google therefore forfeits the liability shields search engines have relied on for two decades. Violating the injunction now exposes Google to fines of up to €250,000 per instance, and Google must cover 80% of the litigation costs. On June 12, 2026, Google confirmed it will appeal, arguing through a spokesperson that "this case focuses on specific and narrow errors, not the foundational way AI Overviews displays web content."
The Case for Strict Liability
The publishers' argument deserves to be taken seriously on its own terms. A search engine that merely indexes and links third-party pages has long enjoyed limited liability because it doesn't originate the content — a defense that traces back to the EU's e-Commerce Directive and now the Digital Services Act's Article 6 hosting exemption. But an AI Overview doesn't index; it writes. The Munich court found that Google's system "independently compiles the information... and summarizes it into a summary text," constructing a claim — "Betrugsmasche," fraud scheme — that existed nowhere in the underlying pages. When a system fabricates a defamatory factual assertion and presents it as an authoritative, first-person answer, treating it as passive hosting stretches the DSA's Article 6 and 16 host-provider protections well past what they were designed to cover. The court weighed the publishers' personality rights under Germany's Grundgesetz (Art. 2(1), 19(3) GG) against Google's expression rights (Art. 5 GG) and concluded that constitutional protection for opinion evaporates once it rests on "demonstrably false factual foundations." That's a defensible line, not judicial overreach — a small publisher has no practical way to correct a hallucination that Google itself generated and served to millions of searchers.
Two Courts, One Question, Opposite Answers
What complicates the picture is that German courts don't agree with each other. Four days after the Munich judgment, on June 1, 2026, the Landgericht Berlin II ruled the opposite way in a related but distinct dispute (case 52 O 62/26 eV): a perfume manufacturer's trademark claim over AI Overviews that surfaced cheaper "Duftzwillinge" (scent-alike) alternatives alongside its brand name. Berlin held that the AI output is "recognizably a summary of third-party content," that Google exercises no "decisive influence" over what it says, and that an average user understands the AI Overview is pulling from other sources rather than making Google's own claim. The underlying legal question — is an AI Overview Google's own statement or derivative third-party content? — is the same threshold issue in both cases, even though one sounds in personality-rights tort and the other in trademark law. Two regional courts in the same country, weeks apart, reached opposite conclusions on it. That is not a stable legal environment for any company operating an AI answer engine in Germany, let alone the EU.
Why Proportionality Should Win
Munich's underlying instinct is right: a company that ships a system generating specific, checkable, false factual claims about identifiable businesses should bear responsibility when it gets those facts wrong, especially when the harm is concentrated and the company alone controls the model. But the remedy matters as much as the principle. Munich's actual order was narrow and proportionate — an injunction against repeating specific false statements, not a blanket declaration that every AI Overview is a defamation risk requiring pre-publication verification of every summarized claim. If appellate courts or the DSA's harmonizing pressure instead push toward treating all generative summarization as inherently unprotected speech, the practical effect is to make AI-assisted search commercially unworkable in the EU's largest economy, since no company can guarantee zero hallucination at web scale. The workable middle path — implicit in Munich's own remedy — ties liability to demonstrated, specific factual falsity plus a practical opportunity to correct it, rather than strict liability for any inaccuracy a probabilistic system might produce. That preserves accountability for publishers genuinely harmed by fabricated claims, like Verlagshaus24, without collapsing the distinction between "got a narrow fact wrong" and "is legally a publisher of everything the model outputs."
What Comes Next
Google's appeal goes to the Oberlandesgericht München, and given the split with Berlin, this dispute is a strong candidate to eventually reach the Bundesgerichtshof — Germany's federal court of justice — for a binding answer on how the DSA's hosting exemption interacts with generative AI output. Until then, the Munich judgment is not yet legally binding (Google's own characterization: "not yet final"), and the Berlin ruling stands as a live counter-precedent. Every company running an AI answer engine that summarizes web content — not just Google — has a direct stake in which reasoning wins, since the same threshold question of "whose speech is this" will recur every time a generative system makes a factual claim about a real business.