Two of the companies that build the tools behind synthetic media told a US judicial panel that courts should not count on technology to sort real recordings from fake ones. At a meeting of the Advisory Committee on Evidence Rules in Boston on October 15, 2026, OpenAI and Meta representatives said deepfake audio and video is getting harder to authenticate, according to Reuters reporting carried by Investing.com. The original Reuters story is behind a login, so this post relies on the syndicated text and on the committee's own public materials.
TL;DR: what was said and what is decided
| Question | Answer |
|---|---|
| Who spoke? | OpenAI policy counsel Shira Anderson; Meta director of product management Devin Norder |
| Main message | AI media is increasingly indistinguishable from authentic content; detection tools are imperfect |
| Is there a new rule? | No. Draft Rule 901(c) has not been adopted |
| How often do deepfake challenges arise? | 15 of 931 responding judges (about 1.6%) in a March survey |
| Committee chair's view | Judge Jesse Furman: the issue "has not yet appeared frequently in actual cases" (as reported) |
What did OpenAI and Meta actually say?
Per the Reuters text, Anderson said AI-generated images keep getting better and that she doubted a scientific way to detect them will emerge. She said each company uses its own watermarking and verification system, so no universal method reliably detects all watermarks or verifies visual content, and that these systems will improve but "will never be completely reliable." Norder said humans cannot distinguish media from recent AI models from authentic content, and that producing it now takes minimal cost and time. The article paraphrases both, so treat the wording as the reporter's summary.
That is notable from companies that sell or support provenance tools. OpenAI has been involved in content credentials work, and Meta has been labeling AI content; our explainers on C2PA content credentials and Google's SynthID verification site describe what those tools can and cannot prove. Their own representatives are now saying that none of it is a complete answer for a courtroom.
What is Rule 901(c)?
The committee has worked since 2023 on authenticating deepfakes. A draft Rule 901(c), described by a National Law Review summary, would address "altered or wholly fabricated AI-generated images, audio, or video" with two steps:
- Opponent's threshold. A party challenging an item as AI-fabricated must present evidence sufficient to support a finding of fabrication to warrant an inquiry by the court. A bare assertion that something is a deepfake would not be enough.
- Proponent's burden. If the threshold is met, the item is admissible only if the proponent shows it is more likely than not authentic, a higher bar than the ordinary prima facie showing under Rule 901.
The same summary says the committee decided an amendment was not necessary at that time, given existing authentication methods and the small number of deepfakes in court so far, and kept 901(c) in reserve. The committee's October 15, 2026 agenda book lists the topic for discussion. We could not read the agenda book text, so we do not describe its contents beyond that.
A companion proposal, Rule 707 on machine-generated evidence, would hold machine output offered without an expert to the reliability standard for expert testimony; it is a separate question from deepfakes.
How common is the problem?
Rarely, so far. A March Federal Judicial Center survey found 15 of 931 responding judges had encountered a case in which a party challenged evidence as a deepfake. That is about 1.6 percent. Furman said academic commentary predicts courts will face the issue but it has not yet appeared frequently.
Low incidence does not mean low risk. Two failure modes matter:
- Fakes admitted as real. A fabricated recording or image gets through.
- The liar's dividend. Genuine evidence is attacked as fake. The National Law Review summary says the committee acknowledged 901(c) does not directly address jurors distrusting genuine evidence and pointed to Rule 403 and judges' gatekeeping instead.
Real cases already show the stakes of synthetic media in court. In Arizona, an appeals court vacated a sentence after an AI-generated video of a victim was used at sentencing, a different issue from forged evidence but part of the same debate. For the fraud side, see the $25.6 million deepfake video call scam.
Why watermarks are not proof
Watermarks and credentials help when they are present and intact. They fail in three common ways:
- Absence proves nothing. Open models and many tools add no mark, and marks can be stripped; see our posts on watermark removal tools and why C2PA on cameras can be broken.
- Fragmentation. Anderson's point is that each company uses its own system, so a court would need many detectors, each with its own error rates.
- Presence of a mark is not authenticity. An authentic photo can be edited, and a mark can attest to origin without attesting to truth.
How did the committee get here? A short timeline
| When | What happened | Source |
|---|---|---|
| 2023 | The advisory committee begins work on authenticating deepfakes and on AI-generated evidence | Reuters text via Investing.com |
| May 2, 2025 | A draft Rule 901(c) for altered or fabricated AI media is considered; the committee decides no amendment is needed yet and keeps it in reserve | National Law Review summary |
| Early 2026 | Public comment on proposed Rule 707 closes February 16, 2026; critics include defense lawyers and a digital-rights group, supporters argue the rule asks the right reliability question | Published comment letters and commentary |
| May 2026 | Per a committee summary, Rule 707 is revised and deferred for study together with the deepfake problem; 901(c) stays a study item, and DOJ's representative said the department did not think it was ready for publication | Committee materials as reported in search summaries |
| October 15, 2026 | Boston meeting; OpenAI and Meta present on authentication difficulty | Reuters via Investing.com |
The pattern is deliberate caution. Evidence rules apply to every federal case, so the committee tends to wait for demonstrated need and then draft narrowly. The industry testimony adds a data point on the other side of the ledger: the people who build the generators say the technology to catch them is not on a path to certainty.
What questions does this raise for AI builders?
Should generators do more? The testimony implies the answer is "provenance, but do not oversell it." If you ship an image, audio or video model, embedding content credentials and invisible marks is a reasonable default, and publishing a public verifier helps. But documentation should state plainly what a missing or present mark does and does not prove. Our explainer on why watermarks are a good case for provenance lays out that argument.
Will detection tools get sold to courts? Probably, and buyers should ask for error rates on realistic, adversarial samples rather than lab benchmarks. Anderson's reported doubt that a scientific detection method will emerge is a caution against any vendor claiming certainty.
Who bears the cost of proving authenticity? Under the draft 901(c) structure, a party making a serious fabrication claim must first offer evidence, and the side relying on the item must then show it is more likely than not authentic. In practice that moves attention toward device logs, original files and custody records, which are cheap to preserve early and impossible to rebuild later.
Does this affect non-US readers? The federal rules bind US federal courts only, but other jurisdictions face the same problem. The Reuters text notes that California and the EU require watermarking to help detect deepfakes; those laws govern generators, not courtroom admissibility, so they are a different layer of the same defense.
What the testimony does not tell us
The Reuters piece does not say whether OpenAI or Meta proposed specific rule language, and we could not find a published text of their remarks. It does not describe how other committee members reacted beyond the chair's comment, and it gives no timeline for a decision. The 1.6 percent survey figure counts judges who encountered a challenge, not confirmed fakes, so it does not tell us how many deepfakes were actually admitted. Readers who need the committee's reasoning should read the agenda book and the minutes once they are posted.
What should people do?
- Preserve originals. Keep the first-generation file, device, and metadata, not a forwarded copy.
- Document custody. Note who handled a file and when.
- Check provenance data where present and record the result, but do not treat a clean check as final.
- Corroborate. Pair video and audio with logs, witnesses, and other records.
- For lawyers: under the draft, a challenge needs concrete evidence of fabrication, and a proponent facing a serious challenge should be ready to show authenticity affirmatively.
What to watch next
The committee has not adopted a rule and the draft text may change before any publication for comment. Watch for the minutes of the October meeting, any move to publish 901(c) or revise 707, and further input from other AI companies. The companies' testimony also feeds the wider policy debate on watermarking laws, such as the California and EU requirements mentioned in the Reuters piece.
Related reading
- What is C2PA content credentials, explained
- Google SynthID website: public AI media verification
- Arizona appeals court and the AI victim video
- The $25.6 million deepfake video call scam
- AI watermark removal tools and C2PA
- C2PA on Android cameras can be forged
Sources: Reuters via Investing.com, National Law Review, and the US Courts committee agenda book. Discussion: Hacker News.
Accurate as of October 16, 2026. Rule drafts may change.