15 U.S.C. § 278h–1 — Standards for artificial intelligence

(a) Mission: The Institute must advance shared frameworks, standards, guidelines, and methods for artificial intelligence (AI); help build a framework for managing the risks of deploying AI systems; help develop technical standards and guidelines that make AI systems trustworthy; and help develop standards for testing whether AI training data and applications are biased. (b) Supporting activities: The Director may support measurement research and best practices for trustworthy AI systems, covering: privacy and security of AI training and testing data and of AI hardware and software; advanced computer chips built for AI; managing data and making it more usable, including cleaning, labeling, and standardizing data, and using common open licenses; keeping AI systems safe and robust, including their ability to handle unexpected inputs and attacks; ways to audit and benchmark AI systems for accuracy, transparency, and safety; using machine learning and AI to help other science and engineering fields; documenting AI models, including their performance, limits, fairness, and training and testing methods; documenting how AI systems connect to and depend on each other, and problems those connections can cause; and any other area the Director sees as important to trustworthy AI. The Director may also produce curated, standardized, high-value, secure, and privacy-protected datasets for AI research and use; support one or more institutes under section 9431(b) to advance AI measurement science and standards; support building voluntary consensus standards, including international ones, through open processes; and enter into contracts, cooperative research agreements, grants, and similar arrangements needed for this work. (c) Risk management framework: Within 2 years of January 1, 2021, the Director had to work with other public and private groups — including the National Science Foundation and the Department of Energy — to build, and periodically update, a voluntary risk-management framework for trustworthy AI. This framework must identify standards, best practices, and procedures for building trustworthy AI systems, assessing their trustworthiness, and reducing their risks; set common definitions for trustworthiness ideas like explainability, transparency, safety, privacy, security, robustness, fairness, bias, ethics, and validation; give case studies showing the framework in use; line up with international standards where appropriate; use voluntary consensus standards and industry best practices; and never require using specific technology products or services. (d) Participation in standard setting organizations: The Institute must take part in developing AI standards and specifications, to make sure those standards promote trustworthy AI and reflect current technology through open, consensus-based processes available to all stakeholders. (e) Data sharing best practices: Within 1 year of January 1, 2021, the Director had to work with public and private groups to create guidance helping industry, federally funded research centers, and federal agencies set up voluntary data-sharing deals to advance AI research, including partnership models that give each side a reason to share the data it collects. (f) Best practices for documentation of data sets: Within 1 year of January 1, 2021, the Director had to work with public and private groups to create best practices for AI training datasets, including metadata standards covering the data's origin, the purpose it was created for, allowed uses, what populations are included or left out, and any other properties the Director decides matter; plus standards for keeping datasets with personal information private and secure. (g) Testbeds: Working with other federal agencies, the private sector, and colleges, the Director may build testbeds — including virtual ones — to support robust, trustworthy AI and machine learning, including testbeds that look for weaknesses and conditions that could cause these systems to fail, malfunction, or be attacked. (h) Authorization of appropriations: Congress authorized this much money for the Institute to carry out this section: $64,000,000 for 2021; $70,400,000 for 2022; $77,440,000 for 2023; $85,180,000 for 2024; and $93,700,000 for 2025.