German court makes Google liable for AI Overview answers

The Hamburg regional court has done what regulators in Brussels and Washington have only threatened: it has held Google directly liable for what its AI Overviews say. The Decoder reports that the court rejected the safe-harbour defences that have shielded search engines for two decades, ruling that when Google's AI invented a link between two named publishers and a fraud case, those were Google's words, not a neutral pointer to someone else's.

The distinction matters. Under the E-Commerce Directive and its German implementation, search engines have long been treated as intermediaries: they index, they rank, they are not the author. Overviews break that frame. The model summarises, paraphrases, sometimes hallucinates. The court's view is that synthesis is speech. A generated paragraph that names a company and attaches it to a crime is a statement by the publisher of the paragraph, which is Google.

The factual core of the case is uncomfortable for Mountain View. The Overview asserted facts that appeared in none of the cited sources. This is the standard failure mode of retrieval-augmented generation: the model stitches together plausible-sounding claims that the underlying documents do not support. Engineers call it hallucination. A German court has now called it defamation.

What changes when the summary becomes the speaker

For brands, the ruling reorders the risk map in two directions at once.

The first is defensive. Any company that has been misdescribed in an AI Overview, an unflattering merger rumour, a fabricated regulatory action, a confused attribution of a competitor's scandal, now has a clearer legal route in Germany, and a template for action elsewhere. Financial services firms, where a single false sentence about solvency or sanctions can move counterparties, should be paying particular attention. So should multilaterals and policy institutions, whose authority depends on never being quoted saying something they did not say. Until now, the practical remedy for an LLM hallucination about your organisation has been to file a feedback form and hope. A liability regime changes the conversation.

The second is structural, and more interesting. If Google is the speaker, Google will behave like a speaker. Expect tighter grounding, more conservative phrasing, and a higher bar for which sources the model is willing to paraphrase from. That bar will not be set by SEO heuristics. It will be set by what Google's lawyers consider defensible. Sources that are verifiable, attributable, and institutionally credible will be preferred over sources that are merely popular. Wikipedia, Reuters, government registries, standards bodies, and named expert publications become safer raw material for a system that can now be sued for what it says.

This is the part senior marketers should sit with. The brands that have spent the last eighteen months optimising for LLM citation through volume, Reddit seeding, comparison-page farming, listicle placement, will find that strategy decays as legal risk rises. The brands that win citations in a post-liability Overview will be the ones the model's operators trust to repeat. For industrial groups publishing technical documentation, for UN-system bodies publishing data, for ISO and IEEE publishing standards, that is a tailwind. For everyone else, it raises the cost of being quoted.

The precedent travels

German regional court rulings do not bind Google globally, and an appeal is near-certain. But the reasoning is portable. The same argument, that generative synthesis is authorship and therefore attracts the liabilities of authorship, will be made under the UK's defamation regime, under Article 82 of the GDPR for inaccurate personal data, and under the various state-level consumer protection statutes that US plaintiffs' lawyers are already testing against OpenAI. The EU AI Act, in force from August, gives this argument an additional hook for high-risk applications.

Google's response will not be to abandon Overviews. They are too central to the defence against ChatGPT and Perplexity, both of which face the same liability logic but with thinner balance sheets. The likelier path is a quiet narrowing: fewer Overviews on queries that name companies or individuals, more hedged language, more visible source attribution to push the speech act back onto the cited publisher. Perplexity has been edging in this direction for months. Google will follow, because the alternative is to litigate every hallucination in every jurisdiction that adopts the Hamburg reasoning.

For communications leaders, the operational implication is concrete. Monitoring what Overviews say about your organisation is no longer a curiosity exercise. It is the early-warning system for a category of reputational and legal exposure that did not exist two years ago, and the evidentiary record you will need if you ever have to demand a correction. The brands that build that monitoring discipline now will spend the next decade shaping what the model is willing to say about them. The ones that do not will find out the way the two German publishers did.