OpenAI and Meta tell a federal evidence panel deepfakes will outrun reliable detection

What the committee put on the record for 15 October

The primary document is the Advisory Committee on Evidence Rules agenda book for 15 October 2026, published by the federal judiciary. The agenda for the Boston meeting schedules a morning panel on deepfakes and Federal Rule 901 authenticity, and a second panel on artificial intelligence and proposed Rule 707. Chair U.S. District Judge Jesse M. Furman appears throughout the opening materials.

On deepfakes, the agenda states that the committee has spent its last seven meetings on a possible amendment and has prepared, but not finally approved, a new Rule 901(c). Under that draft concept, an opponent would need evidence sufficient to support a finding that a challenged item was a deepfake; if that showing is made, the proponent would then have to show the item is more likely than not authentic—a preponderance gate rather than the usual prima facie authenticity standard.

The same agenda book includes reporter memoranda, case digests and panelist biographies. Those bios list Shira Anderson as OpenAI policy counsel working on synthetic media, deepfakes and digital evidence, including engagement with C2PA, and Devin Norder as a Meta director of product management for generative-AI trust, safety and risk, including watermarking and content credentials work. Their presence on the panel is therefore a primary-source fact even before secondary outlets quote what they said aloud.

What OpenAI and Meta were reported to tell the panel

Secondary coverage of the Boston session, including an Investing.com / Yahoo Finance write-up dated 15 October 2026, attributes specific statements to the two company representatives. According to that reporting, Anderson said AI images keep improving and that she does not believe a scientific method will exist in the future to detect them. Norder was reported to say media from recent AI models cannot be distinguished by humans from authentic content, and that producing it takes little time or money.

The same secondary accounts say Anderson warned that companies use different provenance systems, so there is no universal way to detect every watermark, and that those systems will improve without becoming completely reliable. California and EU watermarking requirements are cited as policy context in that reporting, not as new statutes announced at the meeting.

Those spoken quotes are not reproduced as a verbatim transcript inside the public agenda PDF we reviewed. Treat the company bios and Rule 901(c) draft as judiciary primary evidence, and treat the Anderson/Norder sentences as attributed secondary reporting of the panel discussion unless and until the committee publishes a transcript or minutes that repeat them.

Draft Rule 901(c) and why the burden shift matters

Today's authenticity practice under Rule 901 generally asks whether there is evidence sufficient for a jury to find an item is what its proponent claims. The working Rule 901(c) idea described in the agenda would change that sequence after a credible deepfake allegation: the opponent must first muster enough evidence to support a fabrication finding; then the judge, not only the jury, would decide authenticity under a more-likely-than-not standard before the item goes further.

Committee materials in the agenda book debate whether that higher gate is premature. An FJC survey summarised in the record found a 45% response rate among federal trial judges; of 931 respondents, only 15 reported encountering a deepfake challenge. Among judges who had not seen one, 82% said they would want some initial showing before entertaining a fabrication claim—the same structural instinct as draft Rule 901(c). A narrow majority of that group also favoured amending the rules.

The reporter's memorandum collects emerging case anecdotes, including a 2026 Kentucky appellate decision (Adams v. Anderson) where a party faced contempt after submitting altered or AI-created documents in a family-law fight, and older federal sanctions for fabricated filings. Those examples are offered to show the problem exists even if large-scale deepfake trials are still rare.

Separately, the committee is still wrestling with proposed Rule 707 on AI-generated evidence offered without an accompanying expert. The agenda notes the proposal was withdrawn after public comment and is being restudied, with a second expert panel scheduled the same morning. Deepfake authentication and machine-generated expert-substitute evidence are related but not identical rule problems.

Detection tools, watermarks and why vendors sound cautious

Industry detection products already exist. Google has a public SynthID Detector, and watermarking debates continue in policy forums such as OpenAI's EU text-watermarking discussion. The Boston panel remarks, as reported, are not a claim that those tools are useless; they are a claim that no universal, scientific detector will catch every synthetic image as models improve.

That distinction matters for courts. A watermark that works when intact can fail after screenshots, recompression or adversarial stripping. A classifier that scores well on yesterday's generator can miss tomorrow's. Judges deciding authenticity under a preponderance standard would still need usable evidence—chain of custody, device forensics, corroborating witnesses—not a magic "AI or not" meter.

The caution also sits beside broader containment debates, including the scientists' open letter on air-gapping cyber-capable models. Deepfake evidence is a courtroom-facing slice of the same authenticity crisis: synthetic media that is cheap, persuasive and hard to certify.

How often deepfake fights actually reach federal judges

The FJC numbers are the cold water in the agenda book. Fifteen of 931 responding judges is a small incidence rate for a topic that dominates academic and press speculation. Judge Furman, per secondary reporting of the session, noted that academic commentary predicts deepfake litigation while the docket experience so far is limited.

Limited is not zero. The agenda's case digest and the obstruction anecdote in the FJC summary show fabricated audio-visual or documentary submissions already drawing contempt, sanctions or guilty pleas. State family courts and administrative forums may see more of this before federal criminal trials do. Rulemaking lead times are long: even a published proposal would not take effect overnight, which is why the committee keeps the issue on the agenda while debating whether to publish.

For news readers, the accurate near-term takeaway is institutional, not apocalyptic. A federal evidence committee has a draft deepfake authenticity rule, invited frontier labs to a public panel, and still sees few federal deepfake challenges in survey data—even as those labs warn that detection will not become a solved science.

What to watch next

Watch for published minutes or a transcript of the 15 October Boston panels that would lock Anderson's and Norder's remarks into the judiciary record. Watch for any committee vote to publish Rule 901(c) for public comment, or a decision to hold another mini-conference instead. Watch for Rule 707's revised draft if the committee republicises machine-generated evidence standards alongside deepfake authentication.

Until those documents appear, the story that can be sourced tightly is this: the Advisory Committee on Evidence Rules put OpenAI and Meta on a deepfake panel, kept a burden-shifting Rule 901(c) draft in play, and confronted survey data showing the courtroom problem is still rare—while secondary reports of the session say the companies themselves do not expect a universal detector.