Accountable Inference Delivery Protocol (AIDP)

Every conversation a person has with an AI model passes through software on their own screen. In current consumer deployments that software is built and operated by the company that trained the model, so the only component positioned to hold a duty to the user answers instead to the party the user may need protection from. Other technologies of comparable potency resolved this by interposing a party whose loyalty runs the other way: the pharmacist, the broker, the browser.

This paper specifies one for artificial intelligence. An inference advocate is an independently operated client bound by an enacted duty of loyalty to the user alone, operating under the Accountable Inference Delivery Protocol (AIDP). The protocol establishes which model produced a given response, confirms the endpoint was authorized to serve it, evaluates the response independently, gates delivery according to the provider's accumulated conduct and the user's own jurisdiction, and reports incident rates without reporting content. The user's interaction history stays on their own device, and the few facts enforcement needs about them are proven without revealing who they are.

Almost none of this requires invention. The machinery is borrowed from internet protocols we already know how to build and run at scale, so the engineering is tractable and the cost is modest. Two elements have no developed precedent: enforcing provider conduct by aggregate incident rate rather than case-by-case adjudication, and applying the user's own jurisdiction at the point of delivery.

The need is already documented in law. Four jurisdictions have prohibited sycophancy, engineered emotional dependence, and engagement optimization that overrides a system's own safety behavior, while other instruments mandate machine-readable provenance and independent auditing. Every one of these obligations describes something that has to be observed where a response reaches a person, and every one leaves the observing to the party being observed.

Nobody is watching on the user's behalf. This paper defines that role, places it at the boundary where a response reaches a person, and sets out how to keep it from being captured. The aim is plain enough: the software on your screen should work for you.