A German court has held Google responsible for an AI answer assembled from true facts because those facts belonged to a different man.
In a decision dated Sept. 24, 2026, the Regional Court of Aachen ordered Google to stop spreading three versions of an AI Mode answer about a local doctor. He had served as chief physician and professor of neuroradiology until late June. When people searched for why he left, AI Mode supplied a serious criminal allegation, and when they followed up with terms like suspension, termination, and court proceedings, it produced multi-paragraph accounts of a legal aftermath. The proceedings were real. They involved a different man, also a former university professor, and the court could confirm the mix-up from its own records.
The doctor sent Google a detailed explanation in late August. In mid-September, Google said it would take no action, and a formal legal demand sent the same day went nowhere. The court then acted on an emergency basis without first hearing from Google, with fines of up to 250,000 euros per violation. Google can still challenge the order.
The outcome matters less than the reasoning. The court in Aachen relied on an earlier ruling from the Regional Court of Munich, which in May 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 that users could check those sources themselves. The Munich court rejected that, finding that the overview read as a self-contained statement, and described what Google’s AI produces as “independent, new, and substantive statements” built by combining content from third-party sites. Aachen adopted that language directly. Google announced an appeal of the Munich ruling and then, according to The Decoder’s reporting, settled before any higher court weighed in.
Courts in Munich and Aachen have now treated AI search answers as the platform’s own statements, and those statements can be false even when every fact inside them is true. I wrote earlier this year about the liability surface moving from the results page to the answer. These two cases are what that shift looks like once it reaches a courtroom.
A Berlin Court Sees It Differently, And Nobody Has Settled It Yet #
In early June, a Berlin court reached the opposite view in a trademark dispute. A perfume company objected to AI Overviews that named its brands next to links selling cheaper imitations. The Berlin court 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 don’t line up neatly. Berlin’s AI summary accurately repeated what other sites said, while the Munich and Aachen answers invented connections that 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.
That’s the landscape for anyone working in search right now, even if this is from a different country. The rules governing the surface that answers a growing share of your customers’ questions are unsettled; they differ by jurisdiction and even by courtroom, and the risk falls on whoever assumed they were settled. Waiting for clarity isn’t an option, because your brand is already appearing in these answers.
How A True Fact Ends Up Attached To The Wrong Name #
The mechanism is less exotic than the outcome. A system answering a question about a named person or company retrieves documents that match on the name and on surrounding attributes such as role, institution, city, industry, or a recent event. It then generates one fluent narrative from whatever it retrieved. When two entities share enough attributes, the system has no dependable way to know it is holding facts about two different subjects, and the generated answer fuses them. I call this entity conflation, meaning the system merges two distinct people or companies into one and then speaks about the merged result with full confidence.
The pattern repeats across the cases on record. In Aachen, two men shared the attributes of university professor and departure from a post. In Munich, the AI blended the publishers with genuinely dubious companies that, according to sworn statements, had no connection to them. In 2024, Microsoft’s Copilot described a German court reporter as the perpetrator of crimes he had spent years covering, because his name sat beside those crimes in his own bylines. The platforms and years differ, and the failure is the same.
The exposure isn’t limited to people with criminal records nearby. Any entity that shares a name or a cluster of attributes with another is a candidate: two companies with similar names in the same industry, a franchise location confused with the brand, an executive who shares a name with someone else in the same field, a product line whose name overlaps a competitor’s. Most businesses have never checked which neighbors they are easiest to confuse with, because in a list of links the reader did that sorting.
What follows is my own read, an opinion and not something the court found. The Aachen answer was responding to a question the doctor’s public record left open, which was why he left. In my view, an unexplained event in an entity’s public record works like an invitation: the system is asked for an explanation, it reaches for the nearest plausible one, and if the closest match belongs to someone else, that’s what it uses. I can’t show that this is what happened inside AI Mode. I offer it as a way of thinking about risk, and I believe an unexplained gap in an entity’s public record is where conflation most easily gets in.
Where I Think The Platforms Go Next #
This section is my opinion and ideas, not reporting. I’ll state the guess plainly here, and in the next section I’ll lay out what would prove it wrong.
If courts keep treating AI answers as the platform’s own statements, every generated answer about a named entity becomes a potential liability. Disclaimers alone won’t fix that, since Munich already rejected the argument that users can verify for themselves. Pointing to the source links won’t fix it either, because almost nobody uses them. Pew Research found that 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 compared with 15% when no summary appeared. People read the answer, which is exactly why courts treat it as the platform’s speech. So I expect platforms to tighten the pool of sources they trust when answering about specific people and companies. I expect heavier weighting toward facts that are consistent across independent sources, more reliance on structured and disambiguated entity data, and more hedging or outright declining when a system can’t tell which entity a query refers to. The Aachen court said Google could check its output by comparing it with the source sites, and that putting the doctor’s objections to its own AI would have exposed the error. A verification step like that costs compute and time on every answer, so I expect platforms to apply it selectively at first, starting with the query types most likely to end up in front of a judge.
I also don’t expect the effect to stay in Germany. Platforms rarely maintain a separate answer pipeline for each court that rules against them, and source-weighting changes made for one market tend to ship more broadly because maintaining divergent systems is expensive. The same logic applies to every company generating answers about named entities, not only Google, since the Munich reasoning rests on how the product works rather than on who built it.
For businesses, the consequence is that being retrievable stops being enough, and being unambiguous becomes the standard. A company whose facts conflict across its own site, its profiles, and third-party coverage, or whose name and attributes overlap with a neighbor’s, is exactly the kind of entity a cautious system either hedges on or gets wrong.
What Would Make Me Wrong #
Several developments would undercut this forecast, so let’s examine them.
The first is higher courts siding with Berlin. If an appeals court, or ultimately Germany’s Federal Court of Justice, treats AI answers as search results whose operator becomes liable only after being notified, the pressure behind my prediction weakens considerably.
The second is that precedent never forms. Binding law arrives only when someone takes a case all the way through. If Google keeps settling, liability stays a manageable cost rather than a design constraint, and platforms have little reason to change how they weight sources.
The third is regulation going the other way. Lawmakers in the EU or the U.S. could place AI answers under the same protections that cover intermediaries passing along other people’s content. In the US, no court has squarely decided whether Section 230 covers text a model writes, and practitioners expect courts to take that question up soon. A broad reading in the platforms’ favor would remove most of the American pressure.
The fourth is that platforms solve conflation inside the model. Better verification and self-consistency checks could catch merged entities before an answer ships, leaving source weighting largely untouched and keeping the burden off businesses.
The fifth is that platforms suppress rather than tighten. Instead of raising the bar for sources, they could simply stop generating answers about private individuals and smaller companies, which narrows coverage without changing who gets trusted.
If I had to watch one of these, it would be the settlement pattern, because it’s already visible. If the next few cases settle quietly, my timeline stretches considerably.
Make Your Entity Hard To Confuse With Anyone Else #
None of those developments makes the defensive work a waste. Even under the outcome most favorable to platforms, a merged entity is still a business problem, with eroded trust, misdirected customers, and a correction process that, as Aachen showed, can stall for weeks. The work starts with knowing what these systems currently say about you, checked on a schedule rather than after a colleague happens to notice. Measuring that is what I built CitationIQ to do, so weigh my emphasis accordingly; whatever tool you use, the point is to observe before something breaks.
From there, the work looks less like classic reputation management and more like identity hygiene. You aren’t trying to bury a bad result; you are trying to make sure the system never has a reason to blend you with someone else.
- Query the way the Aachen answer surfaced. Search your brand, executives, and products by name combined with role, location, and event terms such as departure, lawsuit, acquisition, or closure, since conflation tends to appear in follow-up queries more than in the first one.
- Publish facts that separate you from your nearest neighbors. State founding details, leadership, locations, what you do, and what you don’t do on pages you control, in plain sentences a system can quote accurately.
- Make your identity machine-readable. Schema.org Organization and Person markup with sameAs links to your authoritative profiles, plus consistent naming everywhere, gives a system an anchor that a name alone cannot.
- Close the gaps yourself. When something material changes, such as a leadership departure, a rebrand, or a location closing, publish your own account so the system has an explanation that actually belongs to you.
- Validate before you publish. Your pages are now raw material for someone else’s sentences, so an error on your own site travels further than it used to, and the claims you make about yourself deserve a second check.
- Document everything when something goes wrong. Keep the exact query, the full answer text, the date and time, the surface, the cited sources, your correction request, and the platform’s reply, because both German cases turned on documented requests that went unanswered.
Unique, verified, first-party facts used to function mainly as a quality signal. In a system that answers in its own voice, they work closer to identity protection. The safest entity in AI search is the one that’s hardest to mistake for another.
More Resources:
- [When AI Has Nothing On Your Company, It Describes Someone Else](https://www.searchenginejournal.com/when-ai-has-nothing-on-your-company-it-describes-someone-else/587876/)
- [AI Answers About Your Locations Are Often Wrong – Check Before Customers Do](https://www.searchenginejournal.com/ai-answers-about-your-locations-are-often-wrong-check-before-customers-do/582205/)
This post was originally published on Duane Forrester Decodes. Featured Image: Roman Samborskyi/Shutterstock