AI Firms Tell Judges the Video Can No Longer Be Trusted

The people who build the tools for making synthetic video went before a federal judicial panel this month to deliver an uncomfortable message: the courts should stop assuming a recording is real just because it looks real. Representatives from OpenAI and Meta told the Advisory Committee on Evidence Rules, meeting in Boston, that AI-generated audio and video are now good enough to pass for authentic, and that reliable detection may never catch up.

The warning was reported by Reuters and carried widely. It is a striking thing to hear from the companies whose models created the problem, and the candour was the point.

No reliable test, maybe ever

Shira Anderson, policy counsel at OpenAI, told the panel she does not believe a scientific method to detect AI-generated images will exist in the future. Watermarking, the usual proposed fix, does not solve it either: each company uses its own system, there is no universal way to read them all, and the schemes will keep improving without ever becoming fully reliable, she said.

Devin Norder, a director of product management at Meta, put the other half of the problem plainly. Media made by recent AI models cannot be told apart from genuine recordings by a human viewer, he said, and producing it now takes very little time or money. Put those two statements together and you have evidence that is cheap to fake and hard to catch.

Rare in court, for now

The scale of the problem inside actual courtrooms is still small. A Federal Judicial Center survey from March found that only 15 of 931 responding judges had seen a case where a party challenged evidence as a deepfake. Judge Jesse Furman, who chairs the committee, noted that the issue has not yet shown up often in real trials.

That gap between what is technically possible and what has actually happened is the hard part for rule-writers. Move too early and you burden every ordinary case with authentication fights over video that is plainly genuine. Move too late and a fabricated clip sways a jury before the system notices. The panel has been working on deepfake authentication since 2023, and the testimony from Boston suggests the window for an easy answer has closed.

Some jurisdictions have reached for watermarking mandates. California and the European Union both require it in various forms, aimed squarely at making synthetic media detectable. Anderson's testimony is a warning that those mandates may be building on sand, since a watermark only helps if it survives and can be read, and neither is guaranteed. The courts spent a century treating photographs and recordings as close to self-authenticating. That assumption is the thing now being quietly retired, and nobody has settled on what replaces it. The question of whether a detector can save us is one researchers are still chasing, even as the fakes get better.