A German court ruled earlier this month that Google is on the hook for its AI search summaries. The court rejected Google’s defenses — that users can verify results themselves and should know AI output isn’t gospel. The ruling stated that AI summaries are an expression of Google’s business activities, making the company responsible for what they say.
This fits into a much older debate about internet publishing. Traditionally, there were carriers and publishers. Phone companies are carriers — they transmit whatever you say and aren’t liable for it. Newspapers are publishers — they choose what goes to print and can be sued for what they publish.
Internet companies have spent decades trying to be both. Section 230 of the 1996 Communications Decency Act enshrined that duality, shielding platforms from liability for user-generated content. That worked when platforms just displayed posts in chronological order. But once Facebook started algorithmically curating feeds, the carrier argument got a lot harder to make.
Google’s AI overviews take this even further. Traditional search archives and links to third-party content. AI overviews rewrite that content, exercising editorial judgment — more like a newspaper article than a neutral index. The same logic applies to restaurant review summaries, legal procedure summaries, or any AI-generated digest of someone else’s work.
Two years ago, Air Canada learned this the hard way. Its AI chatbot promised a discount the airline tried to walk back, claiming the bot was a “separate legal entity.” The court disagreed, ruling the airline was just as responsible for its chatbot’s promises as anything on its website.
If companies could dodge liability simply by delegating to AI, the incentives would be disastrous. Why hire human writers, lawyers, or doctors when you can use cheaper AI and walk away when it screws up?
Visa and OpenAI recently announced a partnership to build AI agents that make purchases on your behalf. If Visa’s AI buys something you didn’t want, who pays? Without clear liability, trust in these systems evaporates.
The numbers matter too. Tests found Google’s AI overviews make mistakes about 10% of the time. With over 5 trillion searches per year, that’s 16,000 bad summaries every second. Most are harmless, but some cause real harm — like when it falsely labeled Canadian fiddler Ashley MacIsaac as a sex offender. His lawsuit is still ongoing.
Bottom line: if a company won’t take responsibility for what its AI says, don’t use that AI.
