A Court Told Google It Owns Its AI's Mistakes. So Do You.

A German court ruled that Google is legally liable for inaccuracies in its AI Overviews, the summaries the search engine generates above the blue links. Google’s defense was the one every vendor reaches for: users should verify the output themselves. The court rejected it. Per Bruce Schneier’s analysis, the judges called the summaries “an expression of Google’s business activities.” The company that published the answer owns the answer.

That sentence is the whole story for anyone deploying agents in front of customers. The legal system has started treating an AI’s statements as the operator’s statements. Not the model vendor’s. Not the user’s. Yours.

The disclaimer defense just failed in court

The standard mitigation for hallucination has been a sentence of fine print: “AI-generated, may contain errors, verify independently.” It functions as a liability transfer. The company emits a confident answer and assigns the cost of being wrong to the reader.

The German ruling treats that transfer as void. If you operate the system, present the output as an answer, and profit from the engagement it drives, the output is a statement you made. A footnote does not convert a published claim into the customer’s research project.

Schneier frames the deeper point bluntly: “Any company that won’t stand by the statements its agents make, whether human or AI, doesn’t deserve users’ time or money.” The point is commercial before it is legal. A business whose answers come with a built-in denial of responsibility is telling customers the answers are not trustworthy. The disclaimer and the product undermine each other.

The error rate is not a rounding error

Google’s AI Overviews carry an error rate of roughly 10 percent, per Schneier’s piece. Against the platform’s volume, the abstraction collapses into something physical. At around 5 trillion searches a year, a 10 percent error rate produces on the order of 16,000 erroneous summaries every second.

The scale turns that percentage into an operations number. If your agent answers customer questions at any meaningful scale, a 5 to 10 percent error rate becomes a stream of wrong statements you publish continuously, each one now carrying potential liability. The volume that makes AI attractive is the same volume that makes the error rate a balance-sheet item.

Most enterprise teams have never measured their own number. They know the demo worked. They do not know how often the deployed agent tells a customer something false, because nobody instrumented it. You cannot price a risk you refuse to count.

Air Canada already drew the line

This is not a single eccentric ruling in one jurisdiction. The pattern predates it. Air Canada was held liable when its support chatbot invented a bereavement-discount policy and a customer relied on it. Per Schneier’s account, the tribunal rejected the airline’s argument that the chatbot was a separate entity responsible for its own statements. The chatbot’s promise was the airline’s promise, treated the same as any claim on the airline’s own website.

Two facts connect the cases. First, the “the AI did it, not us” defense loses. Courts collapse the agent into the operator. Second, the standard being applied is not novel. Companies have always been liable for what their representatives tell customers. A salesperson who promises a discount binds the firm. The ruling simply confirms that an AI agent is a representative, not a magic exception to agency law.

If you have written about where AI and the courts collide, this is a distinct front from the ones already mapped. We covered who owns AI-generated code and the governance that follows, and the precedent set when a frontier lab was ordered to preserve user data. Those concern inputs and access. This concerns outputs, the claims your deployed agent makes in the ordinary course of serving a customer.

Liability is the enforcement mechanism accuracy never had

For two years the argument for AI output quality has been mostly aspirational. Be accurate because it is good practice. Reduce hallucination because users prefer it. None of that carries a price tag, so none of it forced a budget.

Liability changes the incentive structure. A wrong answer that costs nothing gets tolerated. A wrong answer that creates legal exposure gets measured, bounded, and governed, because now it shows up where the company actually pays attention. The German court and the Air Canada tribunal are doing for AI accuracy what regulators did for data privacy: converting a soft preference into an enforceable obligation with a cost attached.

That reframes the work. Output accuracy is no longer a model-quality concern owned by the data science team. It is a governance concern owned by whoever signs off on customer-facing systems. The question shifts from “is the model good enough” to “can we stand behind every statement this agent makes to a customer, and prove we tried.”

Do this now

Run one exercise this week. Take your highest-volume customer-facing agent and ask three questions, in order.

What is its actual error rate in production? Sample real outputs and have a human grade them for factual accuracy. The benchmark and the demo do not count. If you have never done this, the number will be higher than leadership assumes. That number is your liability surface.

Where do its answers create binding commitments? Pricing, policy, eligibility, refunds, legal or medical guidance. Map every place the agent can state something a customer might act on and hold you to. Those are the statements a court will treat as yours.

What happens when it is wrong? Trace one false output end to end. Is it logged? Can you reconstruct what the agent said, to whom, and why? If a customer relied on it, can you detect that before they escalate? If the honest answer is no, you are running an unmeasured liability at production volume, and a disclaimer will not cover you.

The companies that treat this as a governance problem will instrument their agents, bound the high-risk answer categories, and stand behind what ships. The ones that treat it as a disclaimer problem will keep emitting confident wrong answers until a customer, or a court, hands them the bill.

This analysis synthesizes AI and Liability (Bruce Schneier, Harvard Kennedy School, June 2026).

Victorino Group helps companies instrument, bound, and govern customer-facing AI agents so they can stand behind every statement that ships. Let’s talk.