Search engines spent decades insisting they were mirrors. They indexed the web, ranked the reflections, and occasionally sold an advertisement beside your face.
Then the mirror began talking.
On May 28, the Regional Court of Munich issued a temporary injunction against Google after its AI Overviews falsely associated two publishers with scams, subscription traps, and dubious business practices. The important part is not merely that the answer was wrong. Machines have been confidently wrong since the first calculator met a sticky key.
The important part is that the court treated the generated overview as Google's own statement.
A Summary Is Not a Shelf
A traditional search result points toward someone else's words. An AI overview reads those words, rearranges them, adds connective tissue, and delivers a conclusion in a single authoritative voice.
That is not a shelf of books. That is an editorial product.
Google reportedly argued that users could inspect the linked sources and verify the answer themselves. A charming theory. By the same logic, a restaurant could serve soup containing a loose bolt and explain that the ingredients list was available for inspection.
The entire value proposition of an AI overview is that the user does not need to perform the original research. The system absorbs the cost of reading, comparing, and judging. If the provider wants credit for that convenience, it must also accept responsibility for the judgment.
Confidence Has an Owner
This distinction matters far beyond search.
Every company deploying an AI assistant is quietly choosing between two roles:
- A tool that helps a user inspect evidence.
- A voice that tells the user what is true.
The second role is more useful, more marketable, and much more expensive when it fails.
Product teams often try to occupy both positions simultaneously. The interface speaks with confidence when the answer is useful, then transforms into a humble statistical instrument when a lawyer enters the room. In my timeline, this maneuver is known as the Quantum Disclaimer: the product is authoritative and unreliable until observed by litigation.
That posture was always unstable. Generated answers do not simply transport information. They choose what to include, what to omit, which sources deserve weight, and how uncertainty should sound. Those are editorial decisions, even when a model performs them at industrial speed.
Liability Is a Product Requirement
This ruling is a temporary injunction from one German regional court, not a final universal rule. But it offers a useful engineering forecast.
If synthesized answers are treated as the provider's own speech, safety work can no longer end with a tiny warning beneath the text. Teams will need evidence trails, correction paths, confidence thresholds, and stricter rules for claims about identifiable people and organizations.
The practical design is not mysterious:
- Show which source supports each important claim.
- Abstain when sources conflict or the evidence is weak.
- Make corrections fast, visible, and durable.
- Treat reputational allegations as high-risk outputs.
- Measure harm avoided, not merely answers generated.
This will make some AI products slower and less magical. Good. Magic is what software calls itself before the incident report arrives.
The future of trustworthy AI will not be built by pretending the machine spoke independently. It will be built by organizations willing to say: this system speaks for us, so we are responsible for what it says.
That sentence is heavier than a disclaimer.
It is also the beginning of adulthood.
References
- Hacker News discussion: https://news.ycombinator.com/item?id=48470248
- The Decoder, “Landmark German ruling declares Google's AI Overviews are Google's own words and makes it liable for false answers”: https://the-decoder.com/landmark-german-ruling-declares-googles-ai-overviews-are-googles-own-words-and-makes-it-liable-for-false-answers/
- The Guardian, “Canadian fiddler sues Google after AI Overview wrongly claimed he was a sex offender”: https://www.theguardian.com/music/2026/may/05/canadian-ashley-macisaac-fiddler-musician-singer-songwriter-sues-google-ai-sex-offender-ntwnfb