The White House names five expectations for automated systems, and each one is an engineering mechanism

2022-10-04US Policy

White House Office of Science and Technology Policy, "Blueprint for an AI Bill of Rights," whitehouse.gov (October 2022), released Oct. 4, 2022, with technical companion

The White House Office of Science and Technology Policy released a framework of five principles for automated systems together with a technical companion describing how each principle is to be met in practice. The five are safe and effective systems, algorithmic discrimination protections, data privacy, notice and explanation, and human alternatives, consideration and fallback. The framework applies to automated systems that have the potential to meaningfully impact the American public's rights, opportunities, or access to critical resources or services. The document states in its legal disclaimer that "The Blueprint for an AI Bill of Rights is non-binding and does not constitute U.S. government policy." The technical companion carries the substance. Under safe and effective systems it expects that "Systems should undergo extensive testing before deployment," that "Automated systems should have ongoing monitoring procedures, including recalibration procedures," that "Automated systems should be designed to allow for independent evaluation," and that responsible entities "should provide regularly-updated reports" covering performance testing and monitoring results. Under notice and explanation it expects systems to "provide explanations that are technically valid, meaningful and useful" to the people affected and to operators. Under human alternatives it expects an opt-out where appropriate and access to a person who can promptly consider and remedy a problem. Nothing in the document creates an obligation, a right of action or a standard an agency must apply.

What it changesRead the technical companion as a list of mechanisms an artificial intelligence product team will be asked to build whether or not this document ever binds anyone: pre-deployment test harnesses, drift monitors with recalibration triggers, evaluation interfaces for outside auditors, explanation generators, and routing to a human fallback. Each is a system with inputs, state and outputs, and each is patentable when claimed as the mechanism rather than as the goal, which the Federal Circuit's treatment of result-oriented claims makes the only way to claim it. The in-house action is to put the compliance roadmap in front of the patent committee: every feature the governance team specifies to meet one of these expectations is an invention disclosure candidate, and the design document written to satisfy an auditor is usually most of a specification. The same roadmap marks what to keep as a trade secret, since monitoring thresholds and recalibration rules are easier to hide than to claim.