OpenAI and Meta have told the US federal judiciary that deepfake evidence may soon be impossible to tell apart from the real thing, by people or by software. Their representatives spoke on 15 October at a meeting of the Judicial Conference's Advisory Committee on Evidence Rules in Boston, the body that drafts changes to the Federal Rules of Evidence.
The warning lands as the committee weighs whether courts need a new rule for authenticating audio, video and images that a party claims are AI fakes. Members have prepared a draft, but some judges remain unconvinced the problem is big enough to justify one.
What OpenAI and Meta told the panel about deepfake evidence
Shira Anderson, policy counsel at OpenAI, said advances in generative AI could defeat efforts to spot fabricated images. "AI images are only getting better, and I do not think that there will be a scientific way going forward to detect them," she said, according to International Business Times.
Anderson also said each company uses its own watermarking system, so there is no universal way to check for all of them, and that such systems would never be "bulletproof."
Devin Norder, director of product management at Meta, told the committee that people cannot reliably distinguish media made by recent AI models from authentic material. He said producing convincing fakes now takes little money and little time.
Those statements matter because they come from two of the companies whose tools generate this content. Both firms have promoted watermarking and labelling as part of their safety approach. Their message to the judges was that those measures help, but cannot carry the weight of proving authenticity in court.
The draft rule on the table
The committee has studied deepfake authentication since 2023. According to its October 2026 agenda book, it "has prepared, but not finally approved, a new Rule 901(c)."
Under the draft, as described by IBTimes:
- A party that challenges a piece of evidence as a deepfake must first produce enough support for that claim.
- If the challenge clears that bar, the party offering the evidence must show it is more likely than not authentic.
That would add a step to the existing Rule 901, which already requires a party to show that evidence is what it claims to be. The two-stage design is meant to stop lawyers from crying "deepfake" at every inconvenient recording, while giving courts a way to deal with real disputes.
Why the committee has not voted to publish it
The draft has not been released for public comment. Minutes from the committee's May 2026 meeting, included in the agenda book, record that "there would be no need for a Committee vote regarding public comment now," and that members planned to hold a mini-conference instead. In May, the chair noted that even if the rule had been published that summer, it "would not go into effect until December 1, 2028." That timetable has now slipped.
A separate proposal, Rule 707, would cover machine-generated evidence more broadly. The agenda book says the committee voted to withdraw the version that went out for public comment and to submit a modified proposal for further study.
Judges are split on how big the problem is
The panel's chair, US District Judge Jesse Furman of Manhattan, acknowledged that predicted deepfake problems had not yet produced many reported disputes. He has appointed a subcommittee to study the issue further.
Data backs that up for now. A Federal Judicial Center survey found that 15 of 931 responding judges had dealt with a case in which a party challenged evidence as a deepfake.
Judge Valerie Caproni, a committee member, argued that courts already handle false evidence and that existing law already forbids submitting fraudulent material. She cautioned against an AI-specific rule without stronger evidence that the problem is widespread.
The disagreement captures a real dilemma for rule-makers:
- Act early, and courts may lock in a procedure built around today's technology.
- Wait, and the rules may lag behind cases where neither side can prove what is real.
How other jurisdictions are responding
The US federal courts are not the only ones grappling with synthetic media. New laws in California and the European Union require watermarking or labelling of AI-generated content, though OpenAI's own testimony suggests that alone will not settle authenticity disputes in a courtroom. In Europe, Denmark has gone further with a bill on realistic digital imitations, which we summarised in our policy register entry on Denmark's Bill L 49.
For US litigators, the practical upshot is that nothing has changed yet. Authentication still runs through the current Rule 901, and challenges to suspected fakes are handled case by case. What the Boston meeting showed is that the companies building generative tools no longer expect detection technology to rescue the courts. If they are right, the burden will fall on procedure, chains of custody and expert testimony.
The subcommittee's work, and any decision to publish a revised Rule 901(c), will be the next signal of how far the federal judiciary is willing to go.
