S. 3308: Artificial Intelligence Civil Rights Act of 2025

The Artificial Intelligence Civil Rights Act of 2025 would create a broad federal civil-rights framework for AI systems used in consequential decisions. It bars developers and deployers from using covered algorithms in ways that cause disparate impact or discrimination, and it defines protected characteristics expansively: race, color, ethnicity, national origin or immigration status, religion, sex including sexual orientation, gender identity and pregnancy, disability, limited English proficiency, biometric information, familial or marital status, source of income, income level, age, veteran status, genetic information or medical conditions, plus any other classification protected by federal law. The bill also defines the "Commission" as the Federal Trade Commission and applies across all 50 states, D.C., Puerto Rico, the U.S. Virgin Islands, Guam, American Samoa, and the Commonwealth of the Northern Mariana Islands.

A core feature is front-end testing. Before deployment, any covered algorithm used for a consequential action must undergo an evaluation, and if harm is considered plausible, an independent auditor must be brought in. Those evaluations, assessments, reviews, and even developer contracts with deployers must be kept for at least 10 years. Summaries of the reviews must be posted on the developer or deployer website and sent to the FTC within 30 days of completion. The FTC would also have to write rules within 2 years of enactment covering preliminary evaluations, public summaries, and regulations for opt-out rights and appeals to a human decision-maker.

The bill also leans hard on transparency. Individuals must get a short-form notice capped at 500 words, and they must be notified before any material changes are implemented. Version histories of disclosures must also be retained for 10 years. On top of that, the FTC would have to launch a consumer-rights web page within 90 days of enactment, send Congress a report on whether useful algorithmic explanations are feasible within 18 months, begin publishing an annual algorithmic trends report starting 2 years after enactment, and then create a public repository for evaluations and assessments within 180 days after that first annual report. "Covered language" is defined as the 10 languages with the most speakers in the United States, based on Census data, signaling an effort to make rights and disclosures accessible beyond English.

Enforcement is where the bill gets unusually aggressive. State attorneys general could seek civil penalties of $15,000 per violation or 4% of the defendant’s average gross annual revenue over the previous 3 years, whichever is greater. Private plaintiffs could win treble damages or $15,000 per violation, whichever is greater, plus attorney’s fees. People would first have to notify the FTC and their state attorney general, who then get 60 days to respond or intervene. The bill also wipes out pre-dispute arbitration agreements and joint-action waivers for cases under the Act, making it easier to sue in court. To support enforcement, the Office of Personnel Management would have 270 days to create a new federal occupational series for algorithm auditing, and the FTC could hire up to 500 additional personnel.