The agent has a principal

A German court has ruled that Google can be held liable for the errors its AI Overviews introduce — the AI-written summary that now sits atop the search page, ahead of the links it's digesting. Not the sites it draws from. Google. The company that chose to put the summary there.

That's the right call, and the principle underneath it is the one this column has spent a month circling without naming. Bruce Schneier states it plainly:

An agent has a principal, and the principal answers for the agent. For two years the industry has run the other way. The chatbot is a "tool." The summary is "generated." The model "hallucinates" — a word chosen, knowingly or not, to make a false statement sound like weather, something that happens rather than something a product did.

"Hallucinate" is a word chosen to make a false statement sound like weather — something that happens, not something a product did.

Schneier's analogy cuts through it: if Google hired human writers to produce those summaries, it would own their mistakes without argument. Swapping the writers for a model doesn't move the liability — it only makes the mistakes cheaper to generate. The deployer set the system prompt. The deployer picked the model, and decided to show its output to a customer without a human reading it first. That is a chain of decisions with a name at the end of it.

It's the inverse of the move the build-out keeps making. The pattern all year has been cost running downhill — onto the maintainer triaging an AI-written bug report, onto the asylum seeker a model misreads at a border, onto whoever happens to be standing where the externality lands. Liability is the pipe that runs the other way. It sends the cost back to the party that chose to create it. That isn't punishment. It's accounting.

The objection is real, and Schneier raises it himself: hold deployers liable and you make them skittish. Some will pull AI features rather than stand behind them, and the ruling will, at the margin, chill deployment. Good. That's the mechanism working, not breaking. A company that won't accept liability for what its AI tells a customer is telling you it doesn't trust the output either — and would rather you carry the risk it won't. The chill is just confidence being priced before the harm instead of after.

All year the question has been whose hand is on the switch, whose name is on the bill. A German court gave the plain answer: the hand that deployed it. Put your AI between yourself and the world, and you don't get to disown what it says in your name.