S. 1396: Content Origin Protection and Integrity from Edited and Deepfaked Media Act of 2025

S. 1396 does not ban deepfakes, and it does not require you to label anything you make. What it regulates is the record that travels with a file. The bill calls it content provenance information: machine-readable data documenting where a piece of digital content came from and what has happened to it since.

The first rule is for toolmakers. Two years after enactment, a company that sells a tool built mainly to generate or significantly alter content with algorithms would have to give users the option of attaching a record that marks the output as synthetic. Tools built mainly to create or substantially modify copyrightable work would have to offer a provenance record too. The choice stays with the user. When a user opts in, the toolmaker has to take reasonable security measures, to the extent technically feasible, so the record stays machine-readable and is not easily removed.

The second rule is for platforms. Large platforms could not remove, alter, or disable that record, or separate it from the content so that users can no longer see it. Large means $50 million in annual revenue, or 25 million monthly active users in at least 3 of the previous 12 months — a definition that takes in social networks, video sharing services, search engines, and content aggregators. The only exception is research that is necessary, proportionate, and limited to improving the platform's own security. Anyone else, of any size, would be barred from knowingly removing the record in furtherance of an unfair or deceptive practice.

The third rule is for AI developers. A company could not knowingly use work that carries a provenance record to train an AI system, or to generate synthetic content, for commercial purposes unless the owner gives express, informed consent. The company would also have to follow the owner's terms of use, including any terms on compensation. Removing the record first is not a way around it: the rule also reaches work the company knows, or should know, had its record stripped unlawfully.

Three parties could enforce all of this. The Federal Trade Commission would treat violations as unfair or deceptive practices. State attorneys general could sue on behalf of their residents. And if you own the work, you could sue directly for compensatory damages, a court order, and attorney's fees, with four years from the time you discover the violation to file.

The rest of the bill builds the technical groundwork. The Commerce Department official who leads the National Institute of Standards and Technology would set up a public-private partnership to develop voluntary, consensus-based standards for watermarking, provenance records, and detection of synthetic content, run a research program on how well those technologies hold up, and launch a public education campaign within one year. Only the tool requirements carry the two-year delay; the bill's text gives the platform and AI-training prohibitions no separate start date. The bill also states that it does not alter the rights copyright owners already have under other law.