Abstract
This disclosure describes how certified AI agents that answer questions and make recommendations for people are held to the facts published by the businesses and institutions they describe, and how the fairness of their recommendations is made checkable. An organization publishes facts about itself, such as prices, availability, terms, opening hours and policies, as signed, dated, versioned records whose identity and current version are anchored in a public registry. A certified agent, before stating a fact about an organization, must check for such a record; it then states the fact with the record's date, labels any fact it could not verify at source as unverified, and never presents an unverified claim as the organization's own. Each answer that states facts or recommends options carries a signed answer receipt held by the person who received it, listing the sources relied on and their versions. When an organization corrects a fact, the registry carries the correction along its references to every certified agent that relied on the earlier version, and each agent notifies the people whose receipts show they received the superseded fact, within a declared time. For recommendations, each answer's receipt also discloses the basis of ranking, any paid placement, any affiliation between the agent's operator and the options shown, and the number of eligible options considered and left out; accredited bodies audit samples of recommendations against an independent baseline ranking, including the presence of smaller sellers that meet the person's stated criteria, and publish the results as part of the agent's live certification status. When an agent deals with sellers' agents on a person's behalf, it owes the person a duty of loyalty: it refuses inducements, discloses commissions before the person commits, and records each negotiation for the person's review. Finally, certified agents identify themselves cryptographically when they read websites, read each site's published terms for retrieval and for training, follow them, and record each reading session in a form the site can audit, so that sites can admit honest agents instead of shutting all of them out.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Verified Sources™: Certified AI Agents that Check Signed Source Facts Before Stating Them, Give Every Answer a Signed Receipt of Its Sources, Carry Corrections Back to the People Who Received Wrong Answers, Disclose the Basis of Every Recommendation with an Accredited Audit Against an Independent Baseline, Act with a Recorded Duty of Loyalty When Dealing with Sellers' Agents, and Read Websites Under Declared, Recorded Terms", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12918