German Court Holds Google Liable for AI Overview Errors

A German court just told Google that its AI summaries are not a carrier's neutral relay but a publisher's editorial product, making the company liable for every hallucination in those overviews.

That ruling, from earlier this month, rejected Google's defenses that users can check for themselves or should know AI isn't always trustworthy. The court held that the summaries are "above all an expression of Google's business activities." No more straddling the carrier-publisher line.

Why the Old Carrier vs. Publisher Distinction Collapses Against AI

Internet companies have spent decades playing both sides of the carrier-publisher divide. Section 230 of the 1996 Communications Decency Act let them off the hook for user-generated content by saying they aren't "publishers" of others' speech. Traditional search engines got away with it because they archived and linked to third-party pages.

AI overviews don't quote and republish. They rewrite. That's editorial discretion, exactly like a newspaper article or an original essay. The German court saw through the pretense. If Google hires human writers to write summaries, Google is liable for errors. AI doesn't change that duty of care.

16,000 Errors Every Second Changes the Liability Calculus

Tests from earlier this year found Google's AI overviews produce mistakes roughly 10% of the time. Multiply that by Google's 5 trillion-plus searches per year, and you get 16,000 erroneous summaries every second. Most are benign, but some cause real harm - like the AI summary that falsely labeled Canadian fiddler Ashley MacIsaac a sex offender. His Ontario lawsuit is ongoing.

Two years ago, Air Canada tried the "separate legal entity" defense when its chatbot promised a discount the company later rescinded. The court ruled the airline responsible for its chatbot's statements, just like any employee's. The German ruling extends that logic to search engine output.

Liability Is the Only Lever That Forces Trustworthy AI

Visa and OpenAI just announced a partnership to build personal AI agents that make purchases on our behalf. If Visa can disavow those purchases, nobody will trust the system. Liability allocation isn't a legal abstraction - it's the mechanism that determines whether AI agents become useful tools or corporate liability shields.

Companies that won't stand by their AI's statements don't deserve users' time or money. The German ruling, if it holds, will force Google to invest seriously in reducing those 10% error rates. For the rest of us, that's the only outcome that matters.

Source: AI and Liability
Domain: schneier.com