Sec. 3. Foundation model data source and training transparency

1,553 words·~7 min read·/bill/118/hr/6881/ih/section-3

A research copy — for the controlling text, always check the official state or federal source. Not legal advice.

Not later than 9 months after the date of the enactment of this Act, the Commission shall— in accordance with section 553 of title 5, United States Code, promulgate regulations that establish standards specifying information to improve the transparency of foundation models by covered entities with respect to training data, model documentation, data collection in inference, and operations of foundation models; and issue guidance to assist covered entities in complying with the standards established under paragraph (1).

In establishing the standards and issuing the guidance required by subsection (a), the Commission shall consult with the Director of the National Institute of Standards and Technology, the Director of the Office of Science and Technology Policy, the Register of Copyrights, and other relevant stakeholders, including standards bodies, covered entities, academia, technology experts, and advocates for civil rights and consumers. The standards established under subsection (a)(1) shall include requirements, with respect to a foundation model, for— what information specified under such subsection shall be submitted to the Commission by the covered entity that provides such model; and what information specified under such subsection shall be made publicly available by the covered entity that provides such model.

The standards established under subsection (a)(1) shall specify the form and manner in which certain information specified under such subsection, selected at the discretion of the Commission, in consultation with the Director of the National Institute of Standards and Technology and the other actors described in subsection (b), shall be made publicly available by covered entities, including— what information shall be made available on the website of a covered entity that relates to any foundation model provided by such covered entity; what information shall be displayed in a central location on a website hosted by the Commission, which shall include, with respect to a foundation model, information that is substantially similar to the information required under paragraph

(1)to be made available on the website of the covered entity that provides such model; that a machine-readable format shall be used with respect to the information specified under paragraphs

(1)and (2); the URL at which the information specified under paragraph

(2)shall be hosted by the Commission; and such additional specifications as the Commission considers appropriate. The standards established under subsection (a)(1) shall specify a process by which the information required under subsection (c)(1) shall be submitted to the Commission. The Commission shall consider specifying in the standards established under subsection (a)(1), with respect to a foundation model, the following information: The sources of training data (including, as applicable, personal data collection and information necessary to assist copyright owners or data license holders with enforcing their copyright or data license protections) and whether and how data is collected and retained during inference. A description of the size and composition of such training data, including broad demographic information, language information, and other attribute information, while accounting for privacy. Information on data governance procedures, including how such training data was edited or filtered. How such training data was labeled, and information regarding how the validity of the labeling process was assessed. A description of the intended purposes and foreseen limitations or risks of the foundation model, an overview of past edits to such model, the version of such model, and the date of release of such model. A description of the efforts of the covered entity to align the foundation model and the transparency of such model with— the AI Risk Management Framework (or any successor framework) of the National Institute of Standards and Technology; or a similar Federal Government-approved consensus technical standard. Performance under evaluation, either self-driven or through audit, on public or industry standard benchmarks, including what precautions the foundation model takes to answer or respond to situations with higher levels of risk of providing inaccurate or harmful information, including, if such model responds to such questions, relating to the following: Medical, health, or healthcare questions. Biological or chemical synthesis. Cybersecurity. Elections. Policing, including predictive policing. Financial loan decisions. Education. Employment or hiring decisions. Public services. Information relating to vulnerable populations, including children and protected classes. Information on the computational power used to train and operate the foundation model. Any other information determined necessary by the Commission, in consultation with the Director of the National Institute of Standards and Technology, to improve transparency of foundation models. In establishing the standards and issuing the guidance required by subsection (a), the Commission shall consider whether to include alternative provisions for— open-source foundation models; or foundation models that are a derivation of or built upon another foundation model, having been retrained or adapted from such other foundation model to any extent. The regulations required by subsection (a)(1) shall apply beginning on the date that is 90 days after the date on which the Commission promulgates such regulations. Not later than 2 years after the date on which the Commission promulgates the regulations required by subsection (a)(1), and not less often than annually thereafter, the Commission, in consultation with the Director of the National Institute of Standards and Technology, shall assess the standards established by such regulations and update such regulations so as to incorporate appropriate updates (if any) to such standards. A violation of a regulation promulgated under subsection (a)(1) shall be treated as a violation of a regulation under section 18(a)(1)(B) of the Federal Trade Commission Act ( 15 U.S.C. 57a(a)(1)(B) ) regarding unfair or deceptive acts or practices. Except as provided in subsection (m)(3)(C)— the Commission shall enforce the regulations promulgated under subsection (a)(1) in the same manner, by the same means, and with the same jurisdiction, powers, and duties as though all applicable terms and provisions of the Federal Trade Commission Act ( 15 U.S.C. 41 et seq. ) were incorporated into and made a part of this section; and any covered entity that violates a regulation promulgated under subsection (a)(1) shall be subject to the penalties and entitled to the privileges and immunities provided in the Federal Trade Commission Act. Not later than 2 years after the date of the enactment of this Act, the Commission shall submit to the Committee on Energy and Commerce and the Committee on Science, Space, and Technology of the House of Representatives and the Committee on Commerce, Science, and Transportation of the Senate a report on the establishment, implementation, and enforcement of the standards required by subsection (a)(1). There are authorized to be appropriated to the Commission to carry out this section— $10,000,000 for fiscal year 2025; and $3,000,000 for each fiscal year thereafter. In this section: The term artificial intelligence has the meaning given such term in section 5002 of the National Artificial Intelligence Initiative Act of 2020 ( 15 U.S.C. 9401 ; enacted as division E of the William M.

(Mac)Thornberry National Defense Authorization Act for Fiscal Year 2021 ( Public Law 116–283 )). The term Commission means the Federal Trade Commission. The term covered entity means any person, partnership, or corporation described in subparagraph

(C)that provides— use of or services from a foundation model which generates, in aggregate, over 100,000 monthly output instances (whether text, images, video, audio, or other modality), including output instances generated from use by users of second party entities that use such model; or use of or services from a foundation model which has, in aggregate, over 30,000 monthly users, including users of second party entities that use such model. The Commission, in consultation with the Director of the National Institute of Standards and Technology and the Director of the Office of Science and Technology Policy, may, by regulation promulgated in accordance with section 553 of title 5, United States Code, update the number of monthly output instances for purposes of subparagraph (A)(i) or the number of monthly users for purposes of subparagraph (A)(ii), as the Commission considers appropriate. The persons, partnerships, and corporations described in this subparagraph are— any person, partnership, or corporation over which the Commission has jurisdiction under section 5(a)(2) of the Federal Trade Commission Act ( 15 U.S.C. 45(a)(2) ); and notwithstanding section 4, 5(a)(2), or 6 of the Federal Trade Commission Act ( 15 U.S.C. 44 ; 45(a)(2); 46) or any jurisdictional limitation of the Commission— any common carrier subject to the Communications Act of 1934 ( 47 U.S.C. 151 et seq. ) and all Acts amendatory thereof and supplementary thereto; and any organization not organized to carry on business for its own profit or that of its members. The term foundation model means an artificial intelligence model that— is trained on broad data; generally uses self-supervision; generally contains at least 1,000,000,000 parameters; is applicable across a wide range of contexts; and exhibits, or could be easily modified to exhibit, high levels of performance at tasks that could pose a serious risk to security, national economic security, national public health or safety, or any combination of those matters. The term foundation model includes an artificial intelligence model otherwise described in subparagraph

(A)even if such model is provided to users with technical safeguards that attempt to prevent users from taking advantage of any relevant unsafe capabilities. The term inference means, with respect to a foundation model, when such foundation model is operated by a user to produce a result. The term training data means, with respect to a foundation model, the data on which such foundation model was trained.

Connectionstraces to 7

Traces to 7 documents

U.S. Code