Abstract
This disclosure describes SI Impersonation Guard, a method by which AI agents that speak or write to people in live conversations, such as telephone calls, voice assistants, video calls and chat, can never pass as a human or as a particular person or organization without being detected. Every AI agent declares that it is an AI at the start of each conversation, again whenever asked, and again after any break, both in words the listener can understand and in a signed, machine-readable declaration sent alongside the conversation. The declaration carries an accountability token that traces, through a chain of attested agents and delegations, back to a responsible person or organization, so that the listener can see who stands behind the agent without being given anyone's personal data. The listener's device, network or application checks the token against a public registry and shows one of a small set of plain labels, such as verified AI agent of a named organization, unverified AI, or no declaration. A protected-identity registry lets people and organizations enrol their voice, name, image and other likeness as protected; certified agents refuse to generate or present a protected likeness unless the owner has given signed, revocable, scoped consent, and the declaration then names the consent. Signals suggesting impersonation, such as a human-like voice with no declaration, a declaration contradicted by behaviour, or a match to a protected likeness without consent, are answered in grades: a warning to the listener, then interruption or blocking, then a signed report to a quarantine network, with every step recorded and open to appeal. Passing these tests is a condition of certification for any AI agent that converses with the public.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Impersonation Guard™: Declared, Verifiable AI Identity in Live Conversations, in which Every AI Agent Declares Itself and Carries an Accountability Token Traced to a Responsible Party, Listeners' Devices Check It against a Registry, Protected Voices, Names and Likenesses Are Refused without Signed Consent, and Suspected Impersonation Is Answered in Grades with Appeal", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12662