The endorsements are the market refusing the risk. The statute is the law assigning it anyway. On the ninth of December 2026 the European Union's revised Product Liability Directive — 2024/2853 — takes its place in the national law of every member state, and for the first time it names software and AI as products in the strict sense. Strict liability does not ask whether anyone was careless. It asks one question: did the product fail to provide the safety a person was entitled to expect. Negligence litigates conduct — what the deployer should have done. Strict liability litigates the artifact — what the thing did. The first needs evidence of fault. The second needs evidence of defect. And evidence of defect about its own behaviour is the one thing software has never been able to produce.
This is the pincer the deployer is standing in. From the carrier's side, the commercial general liability form now excludes the exposure by default. From the statute's side, strict liability attaches the exposure whether or not the deployer was careful. The exclusion removes the transfer; the directive keeps the liability. Neither document contains the one thing that would relax the pincer, and it is a physical object, not a principle: a record written at the moment the agent acted, saying where that action landed and how far that is from what the agent was authorised to do — a record a stranger can recompute later and get the same number.
Before the argument goes any further, one question to take to the broker: does the directors-and-officers tower affirmatively cover an oversight failure in a system nobody can monitor? Most of the people who approved the deployment have never asked, and the answer is on the same page as the exclusion.
A defect in a physical product is visible — a cracked weld, a contaminated batch, a recall serial number. A defect in an agent's behaviour is a semantic event, and the industry's answer to that — evaluation harnesses, guardrail models, drift detectors, explainability dashboards, every one of them real and worth running — is in each case software watching software. They run on the same machine, through the same scheduler, over the same cache the thing they are watching has just conditioned, and they report in the same currency the thing under watch reports in: a claim about meaning, offered by a system with no way to step outside the claim and check it. That is Rice's theorem in one sentence: a monitor that shares the failure domain of what it watches cannot certify it, and its certificate is a story about a story. This is not a claim that those tools are bad; it is a claim about what a court can do with their output. Strict liability has always run on measurable defect. AI governance has had dashboards. That is not a gap in the law. It is a gap in the instrument the law assumes already exists.
The receipt is that instrument, and it speaks the one dialect this market already trusts: the parametric one. A parametric policy does not pay on a loss adjuster's opinion; it pays on a deterministic trigger — a wind speed, an earthquake magnitude, an index that either crossed the line or did not. The drift receipt is a parametric trigger for role continuity. It does not argue that the agent stayed in its lane; it measures whether the action's position matched its authorised intent, at the substrate, where the answer is a physical mismatch and not a contestable narrative. Out-of-lane stops being a debate and becomes a number that crossed a line. The defect strict liability requires becomes a trigger the parametric desk already knows how to underwrite — and the same signed artifact the deployer hands the carrier to transfer the risk is the one a court reads to apportion it.
What made the roads legal was not a policy written over the factory. It was a policy written over each car. Compulsory motor liability is worth remembering because most people have it backwards: it did not arrive to slow the automobile down, it is what put the automobile on the road. Before it, a driver was an unbounded stranger-risk no city could rationally permit; after it, that risk was a per-vehicle premium somebody could pay. The carmaker carried its own cover, of course — but no policy written over the plant could ever have made one car's harm, on one road, on one Tuesday, into a priced thing. That took a per-vehicle contract and a defect an adjuster could point at: a skid mark, a failed brake line, a number on a dial. An agent is the car, not the factory. The vendor cannot carry where you drove it; the deployer can, one action and one lane at a time, the moment there is something to point at.
Which is why the ninth of December is a pedal, and not the pedal anybody expects. A liability rule with a measurable defect underneath it is the accelerator: it turns an exposure nobody will quote into a premium somebody will, and deployment that was frozen by unpriceable risk gets to move. The same rule with nothing measurable underneath it is the brake — harder than any moratorium, because the only rational answer to an unbounded exposure is to not deploy, or to deploy quietly and hope nobody reconstructs it afterwards. The directive does not decide which of the two it becomes; the instrument underneath it does. Which is worth saying without the hedge: responsible deployment is the only kind that has anything good actually happening in it, and measurement is not the tax charged on that — it is what makes it possible. The deployment that can show what it did is the one still running after the first incident. The one that cannot is pulled, quietly, along with everything good it was doing and the team that was doing it.
So the two legal vectors converge on a single missing object. The American exclusion — the carrier refusing to cover a defect it cannot measure — and the European directive — the statute attaching the loss whether or not anyone was careless — are not competing readings of AI risk; they are the demand side and the supply side of the same instrument. The market will not insure a defect it cannot measure, and the statute will not excuse a defect it can attach. Both clear the moment the defect becomes a signed, recomputable, parametric fact — which is the moment the substrate produces the receipt.