On Thursday, representatives from OpenAI and Meta told a federal judicial panel in Boston something unusual for tech companies to admit about their own products: the fakes are getting too good to catch.

The panel is the U.S. Judicial Conference’s Advisory Committee on Evidence Rules, the group that shapes how evidence works in federal courts. It has spent three years trying to figure out what to do about AI-generated evidence, including how to make sure it meets the same reliability bar as testimony from human expert witnesses.

Shira Anderson, policy counsel at OpenAI, told the committee that AI images keep improving and that she doesn’t expect a scientific method for detecting them to exist in the future. Meta’s Devin Norder, a director of product management, said humans can no longer tell media from the latest AI models apart from the real thing, and that making it now takes almost no money or time.

What about watermarks? California and the EU now require them, and both companies have built verification tools. But Anderson pointed out the catch: every company uses its own system. There’s no universal way to check all watermarks, and she said these systems will improve but never be fully reliable.

Here’s the twist. Right now, this is mostly a future problem. A March survey by the Federal Judicial Center found only 15 of 931 responding judges had seen a party challenge evidence as a deepfake. Judge Jesse Furman of Manhattan, who chairs the panel, noted that academics have predicted a deepfake wave in courtrooms, but it hasn’t really shown up in actual cases yet.

The crazy part of this story is the source. When the actual companies who building the AI tell judges, on the record, that detection may never be reliable, it’s worth taking seriously. The good news is that courts have time: the problem is still rare, and judges are working on it before it becomes common.

The harder problem isn’t just catching fakes. It’s what happens to trust in real recordings once everyone knows fakes are undetectable.

How worried?: 🟡 3/5 — Meaningful (The tools to fake evidence are here; the rules to handle it aren’t yet)

Picture a custody hearing in 2028. A mother plays a doorbell-camera clip of her ex-husband shouting threats at her front door. It’s real. His lawyer doesn’t argue about what was said. He simply asks: “Can anyone in this room prove this wasn’t generated by AI?”

No one can. There’s no watermark, because doorbell cameras don’t add one. The forensic expert says the clip shows no signs of manipulation, but admits, under questioning, that the newest models don’t leave any. The judge, unable to tell either way, gives the video little weight.

Researchers call this the “liar’s dividend”: once everyone knows fakes are possible, people caught on tape can claim the tape is fake. It doesn’t take a single deepfake to work. It just takes doubt.

The fix probably isn’t better detection. It’s proof of origin: cameras and phones that cryptographically sign footage the moment it’s recorded, so courts can verify where a file came from instead of guessing whether it’s fake. Some of this technology already exists. Whether it becomes standard before the doubt does is the open question.

The signal isn’t a more realistic deepfakes. It’s the first high-profile case where a real recording is dismissed, or a fake one is admitted, and changes the outcome. Also watch whether the evidence committee moves from studying the problem to proposing an actual rule.