OpenAI and Meta warn US courts that deepfakes may outrun detection

On October 15, 2026, two of the biggest AI companies in the world walked into a room of federal evidence experts and said something blunt. OpenAI and Meta told the Advisory Committee on Evidence Rules of the US Judicial Conference that deepfakes may already outrun reliable detection. The meeting took place in Boston. The audience was not a product launch crowd. It was the group that writes the Federal Rules of Evidence that American courts use every day.

That warning matters because evidence law assumes judges and juries can still tell real from fake. If that assumption breaks, authentication fights stop being rare edge cases and become a core problem for trials, hearings, and discovery.

The three things you need to know

• OpenAI says detection will not save you. Policy counsel Shira Anderson told the committee she does not believe a scientific method will exist to reliably detect AI generated images. Company watermarks, she said, are not universal and will never be completely reliable.
• Meta says humans cannot tell. Devin Norder, who works in product management at Meta, said media from recent AI models cannot be distinguished by humans from authentic content, and creating it takes little cost or time.
• Courts are drafting a new rule. The committee has worked on deepfakes since 2023. Draft Rule 901(c) would require a challenger to show evidence of fabrication, then require the proponent to prove authenticity by a preponderance of the evidence. US District Judge Jesse Furman of Manhattan chairs the committee. A Federal Judicial Center survey in March found that only 15 of 931 responding judges had seen a deepfake challenge. California and the EU already have watermarking laws.

What OpenAI told the panel

Anderson's remarks go further than the usual industry talking point about "responsible AI." The common product pitch is that detectors will improve, watermarks will spread, and provenance standards will keep the information ecosystem intact. Anderson told the committee that she does not believe science will deliver a reliable detector for AI generated images. That is a different claim. It is not "detection is hard today." It is "do not plan on detection as the backstop."

She also undercut the idea that voluntary company watermarks can solve authentication for courts. Watermarks only work if everyone uses the same system, if the signal survives editing and compression, and if an adversary cannot strip or forge it. Anderson said company watermarks are not universal and will never be completely reliable. For a judge deciding whether a photo or video is authentic, that is a practical problem. A missing watermark does not prove fakery. A present watermark does not prove truth if someone can forge the mark or if the chain of custody is broken.

OpenAI has an obvious interest here. The company builds image and video generators. If courts treat AI media as presumptively unreliable, that raises the bar for any party that wants to introduce generated or AI assisted content. But the warning also lands on OpenAI's critics: detection tools marketed as forensic certainty may overpromise.

What Meta told the panel

Norder focused on the human side. According to reporting on the session, he said media from recent AI models cannot be distinguished by humans from authentic content. He added that creating such media takes little cost and little time. That combination is the real threat model for evidence law. If fakes are cheap, fast, and visually convincing, then volume matters as much as quality. A single perfect deepfake is a problem. A flood of good enough deepfakes is a different problem, because it can poison trust even when the media in a given case happens to be real.

Meta runs massive consumer platforms and also builds generative models. The company has pushed labeling and watermarking in other contexts. The message to the Evidence Rules committee was still sobering: do not assume human perception will catch AI media, and do not assume cost will limit who can produce it.

Draft Rule 901(c)

Federal Rule of Evidence 901 already covers authentication. The draft Rule 901(c) would create a special path for deepfake style challenges. Under the draft approach described in coverage of the Boston meeting, a challenger would first have to show evidence of fabrication. Only then would the proponent have to prove authenticity by a preponderance of the evidence.

That structure tries to stop fishing expeditions. Without a threshold showing, every photo and every video in every case could be attacked as "probably AI." With a threshold showing, the party that raises the deepfake alarm has to bring something concrete. Once that showing is made, the burden shifts in a way that puts authenticity front and center.

Chair Jesse Furman, a US District Judge in Manhattan, leads the committee's work. The committee has been working on deepfakes since 2023. That timeline matters. Courts did not wake up in 2026 and discover generative AI. They have been drafting for years while model quality kept climbing.

Reporting also notes a Federal Judicial Center survey from March. Only 15 of 931 responding judges said they had already seen a deepfake challenge in their courtroom. That number can be read two ways. Optimists say deepfake evidence fights are still rare. Pessimists say the survey measured the past, not the next two years, and that the Boston warnings from OpenAI and Meta are about the curve ahead.

Watermark laws already exist elsewhere

California and the European Union have watermarking laws on the books. Those rules push companies toward labeling or technical provenance signals. They do not magically make courtroom authentication easy. A statute can require a watermark at generation time. It cannot force every foreign model, every open weight checkpoint, or every offline edit to leave a trustworthy trail. Anderson's point about incomplete reliability sits next to those laws, not instead of them. Policy can mandate disclosure. Courts still need a workable test when disclosure is missing, forged, or stripped.

Why it matters

Evidence rules are the plumbing of justice. They decide what a jury may see and what a judge may rely on. If AI media becomes cheap and visually perfect, two failures can happen at once. First, fake evidence can slip through if authentication standards stay too loose. Second, real evidence can be drowned by doubt if every party can shout "deepfake" without a serious showing.

Draft Rule 901(c) is an attempt to balance those risks. It does not ban AI media. It tries to structure how a challenge works. OpenAI and Meta are telling the committee that the technical arms race for detection may not be winnable in the way lawyers hope. That is uncomfortable for anyone who wants a simple forensic test. It is useful for anyone who has to design rules that still work when detectors fail.

The Boston session also shows how fast the conversation moved from "will deepfakes arrive" to "how do courts live with them." The committee has been writing since 2023. The companies that build the generators are now saying out loud that reliable detection may never arrive. Judges, for the most part, have not yet seen these challenges in volume. The gap between industry capability and courtroom experience is exactly where evidence rules get rewritten.

Dany's take

I take the OpenAI and Meta warnings seriously, and I do not treat them as pure theater. If the builders of the models say detectors and watermarks will not fully solve authentication, courts should plan for that world instead of waiting for a magic detector. Draft Rule 901(c) looks like the right kind of response: force challengers to bring real fabrication evidence, then make authenticity a clear burden. What I would watch next is whether the final rule stays narrow enough that routine photos and videos do not turn into endless AI fights, and whether watermark laws in California and the EU actually create portable provenance that survives into US trials. For now, the takeaway for anyone following AI policy is simple. The detection story is weaker than the marketing, and the evidence rules committee knows it.

What this means for AI news readers

If you follow AI launches every day, the Boston panel can look like a side story next to model releases and chip deals. It is not. Authentication is where generative AI meets power. Courts decide liberty, money, and reputation. When OpenAI says reliable detection may never exist, and Meta says humans cannot tell recent AI media from real media, the story is not only about tech. It is about whether our public institutions can still agree on what counts as a fact in a file.

For creators and platforms, the same warning cuts another way. Watermarks and labels remain useful as signals. They are weak as absolute proof. California and EU rules will keep pushing disclosure. US federal evidence practice will keep asking a harder question: when the mark is gone or disputed, who has to prove what, and with what standard.

Stay with the primary sources linked above. The Yahoo Finance Investing.com writeup, the AI Affairs report, and the official US Courts agenda book are the cleanest public windows into what was said and what the committee is drafting. The optional Mad Robot post adds more context for readers who want a longer walkthrough.

Sources

• OpenAI, Meta warn U.S. judicial panel on deepfakes (Yahoo Finance / Investing.com)
• OpenAI and Meta tell US evidence panel deepfakes outpace authentication (AI Affairs)
• US Courts Evidence Rules agenda book (PDF)
• Optional deeper read: Mad Robot on the panel