Restricted analysis, made public daily.
Declassified under standing order Edition No. 020 Thursday, July 23, 2026

A German court ruled an AI’s wrong
answer is Google’s liability.

In May, Munich’s regional court held Google directly liable after its AI Overview wrongly linked real Munich publishers to fraud — ruling that an AI-generated answer is the platform’s own speech, not a neutral summary of someone else’s, and rejecting the argument that a user should have simply checked for themselves.

Someone searched real Munich publishers’ names alongside German terms for fraud and subscription scams, and Google’s AI answered as if it already knew the verdict. The AI Overview stated, in full sentences, that the publishers were known for fraudulent business practices, scams, and subscription traps. Nobody at Google wrote those sentences. The model did — synthesizing them fresh from whatever it had read. In May, the Munich I Regional Court ruled that the distinction doesn’t matter: the words are Google’s, and so is the liability for them.

The case turned on one distinction the court refused to blur: what a search result does with a source, and what an AI Overview does to it.

Google raised the obvious defense. The court didn’t need long to reject it.

An AI Overview isn’t a summary of what other people said about you. It’s a new statement, in the platform’s own voice — and now, in its own name, on its own liability.

What the ruling actually changes

Anatomy of the ruling

Synthesis, not citation, is the trigger. The court’s dividing line: quoting a source is one thing, generating a new sentence none of those sources contain is another. Only the second counts as the platform’s own speech.

Ownership follows control. Because Google built and operates the algorithm producing the statement, the court attributed the words to Google directly — not to the pages the model happened to read.

Fact-checkability is not a defense. The court rejected Google’s argument that users could simply verify the claim themselves. A statement being disprovable doesn’t undo the harm of it having been made.

What this doesn’t fix

The ruling is German, one regional court’s reasoning, not a global standard, and not something a business outside Germany can simply invoke — and Google has said it will appeal, so even this specific holding isn’t final. It also addresses liability after a false claim already existed inside an answer — it does nothing to stop the claim from forming in the first place, or to guarantee it gets corrected quickly once it does. Edition No. 013 covered exactly that mechanic three weeks ago: a wrong fact about a firm doesn’t need a person to repeat it, only a second machine to agree — and the correction has to outrun the citation, not just contradict it. This ruling makes the harm compensable after the fact. It does not make the first draft correct.

That’s the gap Identity Architecture exists inside — not the courtroom remedy for a false statement already published, but the structural work of making sure the systems generating that first sentence read the right sources, in the right order, before they ever draft an opinion of a real business at all.

Twenty editions in, the throughline sharpens: courts are beginning to treat an AI’s sentence as the platform’s own, which means a wrong sentence about you may now be worth suing over. It also means a correct one, encoded first, is worth more than any lawsuit ever recovers.

Sources

The Munich I Regional Court (LG München I) ruled May 28, 2026 that Google is directly liable for false statements generated by its AI Overview feature — reported by the U.S. Library of Congress Global Legal Monitor.

Case background and the court’s reasoning on synthesized vs. quoted content — Malwarebytes.

Analysis of the ruling’s significance for AI-platform liability generally — The Decoder.

Google confirmed it will appeal the ruling — reported by Reuters.

Munich just ruled an AI’s sentence is the platform’s own.
SIA makes sure yours is correct
before any court has to.

A German court just held Google liable for what its AI said about a real business, unprompted. That’s a remedy for after the damage — a claim filed once the wrong sentence has already shipped. Identity Architecture works earlier: structuring what a system can verify about you so the first sentence it drafts is the accurate one — not a claim your counsel has to correct in civil court, assuming you ever find out it was made at all. Most businesses never do. They just quietly stop getting recommended.

Every edition, in order, from No. 001 · A new edition releases daily, 05:30 CT.

Open the Record
‹ Edition No. 019 Edition No. 021 ›