Google cannot ignore its AI’s false accusations

On 24 September, a regional court in Aachen ordered Google to stop telling users that a German chief physician had been convicted of sexual coercion (KPW, Beck aktuell). The conviction belonged to someone else. Google had been told about the error and refused to remove it. Whether Google carries the same duty to act on such errors elsewhere in Europe is still unsettled.

Two men, one answer

On 28 August, a colleague searched the doctor's name together with the German word for "suspension." Google's AI Mode said he had been convicted, dismissed and paid off with a six-figure settlement (Beck aktuell). That history belonged to another former senior figure at the same clinic.

The AI had merged two people. The articles it drew on were about the other man and never named the doctor. The doctor says no complaint, proceeding or conviction ever existed against him. When he asked Google to take the answers down, the company replied on 15 September that it would take no action (LHR). A formal warning from his lawyer also failed.

What Google must now stop

The order is narrow. It bars three specific versions of the false answer. It is aimed at Google Ireland Limited, the company's European subsidiary (KPW).

The court treated Google as a Störerin. Under German personality-rights law, that is a party obliged to end a violation it has the power to stop, even if it did not originate it. A false criminal accusation is serious and urgent, so the court ruled without hearing Google first. A breach could bring a fine of up to €250,000, though no fine has been imposed (LHR).

An earlier German ruling ran in the same direction. On 28 May, the Munich I Regional Court examined an AI Overview that linked two local publishers to subscription fraud (Munich judgment). It found that Google's system had not simply quoted its sources. The system had combined them into a new statement, and the court held Google responsible for that statement.

Google appealed the Munich ruling, but the appeal ended in a settlement, so no higher court tested the reasoning (The Decoder). Both decisions come from first-instance courts. Neither binds judges elsewhere.

The gap in EU law

The published accounts record that Google refused, not why. The strongest argument open to it is scale. A system answering vast numbers of questions cannot check every sentence in advance, and users can follow the cited links.

EU law addresses that argument without settling it. The Digital Services Act is the EU's 2022 rulebook for online platforms. It protects services that store or pass on other people's content. Those services have no general duty to monitor that content, and they are not liable if they act promptly once told something is unlawful (Digital Services Act).

The open question is whether a sentence written by Google's own AI counts as other people's content at all. If it does, Google is a host that must act once notified. If it does not, Google looks more like the author of the claim. No ruling from the EU Court of Justice has answered this (Lawfare).

Aachen did not have to pick a side. Under either reading, Google had been shown a documented falsehood and left it online. That is our reading of the reported facts, not the court's stated reasoning.

Italy offers one point of comparison. On 29 April, its communications regulator, AGCOM, asked the European Commission to examine AI Overviews and AI Mode under the act's rules on systemic risks (AGCOM). Those rules require the largest platforms to assess and limit harms their services cause. The Italian move was driven mainly by publishers' worries about losing readers, not by false accusations. A Commission inquiry could push Google to manage such risks better. It would not, by itself, decide who is liable when an AI answer defames someone.

The doctor got what this case could deliver: a court order against three false answers about him, issued after Google ignored his complaint. Europe still has no settled rule for AI search answers in general. People elsewhere are not without remedies, because national defamation and privacy laws still apply. But each court must decide for itself whether Google's AI speaks for Google. A Europe-wide answer would take a higher national court, a Court of Justice ruling or new legislation, and none has yet arrived.