OpenAI and Meta warned federal judges that tools for identifying AI-generated audio and video content will probably never be completely reliable, even though only 15 of the 931 judges surveyed said they had encountered this type of case.
Key points:
OpenAI and Meta told a federal panel in Boston that authenticating AI-generated images and videos is becoming increasingly difficult.
An OpenAI lawyer said that neither watermarks nor detection tools will be able to offer complete reliability.
Judges are studying a proposed rule on deepfakes, but have not yet adopted it.
OpenAI’s warning about deepfakes
Representatives from both companies presented their concerns on October 15 at a meeting in Boston of the Judicial Conference of the United States’ Advisory Committee on Evidence Rules. The working group has been looking into the issue since 2023.
Shira Anderson, OpenAI’s public policy counsel, told the committee that the quality of AI-generated images is constantly improving and that she does not expect a foolproof scientific method for detecting them to emerge. New laws in California and the European Union already require watermarks. But each player uses its own technology, she noted, so there is no universal system for verifying all visual content. In her view, these tools are unlikely ever to become “foolproof.”
Devin Norder, Meta’s director of product management, added that the public can no longer distinguish content generated by the latest AI models from authentic images and videos. Producing these fakes now takes little time or money, he said.
Also read: Bitcoin holds $80,000: Just a pause before a test of $87,000?
Jesse Furman’s examination of deepfakes
Federal Judge Jesse Furman of the Manhattan District Court, who chairs the committee, noted that academics’ alarming predictions of deepfakes flooding courtrooms have not yet materialized in actual cases. The figures bear this out: in a survey conducted in March by the Federal Judicial Center, only 15 of the 931 responding judges reported that a party had challenged evidence on the grounds that it was a deepfake.
Furman appointed a subcommittee to examine the issue further. Its members drafted—but have not yet approved—an amendment to the rules called “Rule 901(c).” It would require a party alleging that evidence is fake to first substantiate the challenge, before the opposing party is required to show that the content is more likely authentic than fabricated.
One committee member urged caution, however. Judge Valerie Caproni warned against rushing to enact legislation specifically targeting AI, in the absence of solid evidence that deepfakes are already a systemic problem in the courts.
She noted that the justice system has long dealt with falsified evidence and that existing laws already prohibit the creation and submission of fraudulent documents.
In May 2025, the committee declined to adopt proposals from several legal experts to establish rules specifically for deepfakes, with most members believing that current authentication standards are sufficient to address AI-generated content.
A few months later, Alameda County Superior Court Judge Victoria Kolakowski dismissed a California housing case after finding that the plaintiffs had submitted an AI-generated video of a witness as evidence, passing it off as a real recording.
Coming up: The 39.6% crypto rally has a problem: spot buyers aren’t following