German Court Orders Google To Stop AI Answer Built From True Facts

A German court has ordered Google to stop publishing three versions of an AI Mode answer about a local doctor, under threat of fines up to 250,000 euros per violation. The decision, dated Sept. 24, 2026, came from the Regional Court of Aachen. Every fact in the answer was true. The facts simply belonged to a different man.

The doctor had served as chief physician and professor of neuroradiology until late June. When searchers asked why he left, AI Mode supplied a serious criminal allegation. Follow-up queries using terms like suspension, termination, and court proceedings produced multi-paragraph accounts of a legal aftermath. The proceedings were real. They concerned another former university professor, and the court confirmed the mix-up from its own records.

How did the dispute reach a courtroom?

The timeline matters. The doctor sent Google a detailed explanation in late August. In mid-September, Google said it would take no action. A formal legal demand sent the same day also went unanswered. The court then acted on an emergency basis, without first hearing from Google. Google can still challenge the order.

The outcome matters less than the reasoning. Aachen relied on a May ruling from the Regional Court of Munich, which held Google directly liable for AI Overviews that tied two publishers to fraud and subscription traps using claims that appeared in none of the linked sources. Google argued users could check the sources themselves. Munich rejected that argument. It found the overview read as a self-contained statement and described Google's AI output as "independent, new, and substantive statements" built by combining content from third-party sites. Aachen adopted that language directly. Google appealed the Munich ruling, then settled before any higher court ruled, according to The Decoder's reporting.

Takeaway: two German courts now treat AI answers as the platform's own speech, even when every underlying fact is accurate.

Why does a Berlin ruling point the other way?

In early June, a Berlin court reached the opposite conclusion in a trademark dispute. A perfume company objected to AI Overviews that placed its brands next to links selling cheaper imitations. Berlin treated AI Overviews as a new format of search result that bundles information from other sites, not as Google speaking for itself, and found no infringement.

The cases do not line up neatly. Berlin's summary accurately repeated what other sites said. Munich and Aachen involved answers that invented connections no source made. Even so, courts in the same country have described the same product in incompatible terms, and no higher court has ruled, because the one case headed there settled. The risk falls on whoever assumed the rules were settled.

What causes a true fact to attach to the wrong name?

The mechanism is not exotic. A system answering a question about a named entity retrieves documents matching the name plus surrounding attributes: role, institution, city, industry, recent events. It then generates one fluent narrative from what it retrieved. When two entities share enough attributes, the system cannot reliably tell them apart, and the answer fuses them. Duane Forrester, the original analyst, calls this entity conflation.

The pattern repeats. In Aachen, two men shared the attributes of professor and departure from a post. In Munich, the AI blended publishers with companies that, per sworn statements, had no connection to them. In 2024, Microsoft's Copilot described a German court reporter as the perpetrator of crimes he had covered, because his name sat beside those crimes in his own bylines. The platforms differ. The failure does not.

Exposure extends beyond people with criminal records nearby. Candidates include two similarly named companies in one industry, a franchise confused with its brand, an executive sharing a name with a peer, or a product line overlapping a competitor's name.

Takeaway: any entity sharing a name or attribute cluster with another is a conflation candidate, and most businesses have never checked their nearest neighbors.

What might platforms do next?

This section is the analyst's opinion, not a court finding. If AI answers count as platform statements, every generated answer about a named entity carries liability. Disclaimers will not fix it; Munich already rejected the self-verification argument. Source links will not fix it either. Pew Research found users clicked a link inside the AI summary on just 1% of searches where one appeared, and clicked any traditional result on 8% of those searches versus 15% when no summary appeared.

The likely response: tighter source pools for queries about people and companies, heavier weighting toward facts consistent across independent sources, more structured entity data, and more hedging when the system cannot identify the entity. Aachen noted that feeding the doctor's objections back into Google's own AI would have exposed the error. Verification costs compute, so expect selective rollout first, starting with query types most likely to reach a judge.

The analyst also flags what would prove him wrong: an appeals court siding with Berlin, continued settlements that prevent precedent, intermediary-style regulation in the EU or US, model-level fixes for conflation, or platforms simply declining to answer about private individuals and smaller companies. In the US, no court has squarely decided whether Section 230 covers model-written text.

Takeaway: the forecast depends on courts and settlements, not on Google's product roadmap.

What should site owners monitor?

The defensive advice holds under every scenario. Search your brand, executives, and products by name combined with role, location, and event terms such as departure, lawsuit, acquisition, or closure, since conflation appears more in follow-up queries. Publish disambiguating facts on pages you control. Add Schema.org Organization and Person markup with sameAs links. Publish your own account when something material changes. And document everything when an answer goes wrong: exact query, full answer text, date, surface, cited sources, correction request, and the platform's reply. Both German cases turned on documented requests that went unanswered.

Watch the settlement pattern above all. If the next cases settle quietly, the timeline for any binding precedent stretches considerably, and the rules stay exactly as unsettled as they are today.