OpenAI and Meta Told Federal Judges Deepfakes May Be Undetectable: What It Means for Your Evidence (AI Trends, 18 October 2026)

What’s trending in AI on 18 October 2026: The companies that build the most realistic image and video generators have just told America’s federal judges, in effect, not to count on spotting their output. At a meeting of the U.S. Judicial Conference’s Advisory Committee on Evidence Rules in Boston on Thursday, an OpenAI policy counsel said she does not expect a scientific method to reliably detect AI images in future, and a Meta product director said people can no longer tell recent AI media from the real thing. The same committee has now paused both of its draft AI evidence rules, which means no federal deepfake rule can even be published for comment before August 2027. For businesses, any claim, investigation or dispute that relies on a photo, recording or video is now open to the question “how do you know that’s real?” Below: what was said, where the rules stand, what California and the EU already require, and a scenario playbook for proving your own evidence is genuine.

Key takeaways

The builders say detection won’t save us. OpenAI’s Shira Anderson told the committee detection will improve but never be fully reliable, and that each company’s watermark works differently, so no single tool catches them all.
Deepfake fights in court are still rare. A March Federal Judicial Center survey found just 15 of 931 responding judges had seen evidence challenged as a deepfake.
Both draft rules are on hold. Proposed Rule 707 (AI-generated evidence) was withdrawn as published, and the draft deepfake rule, 901(c), has not been finally approved. The earliest either could be republished for comment is August 2027.
Judges are split. Some argue existing authentication rules already handle fakes; others want an AI-specific test that shifts the burden onto whoever offers a challenged exhibit.
Your proof has to start at capture. With detection unreliable, provenance (where a file came from and who handled it) is becoming the practical way to show something is real.

The firms behind the most realistic generators say courts should not rely on detection tools to tell real from fake.

1. What happened in Boston

The Advisory Committee on Evidence Rules is the body that drafts changes to the Federal Rules of Evidence, the rulebook that decides what a jury in a U.S. federal court is allowed to see. It has studied AI since 2023, and on Thursday 15 October it opened its autumn meeting with expert panels on AI-generated evidence and deepfakes.

Shira Anderson, policy counsel at OpenAI, told the panel that AI images keep improving and that she does not expect a scientific method to detect them in future. She expects detection tools to get better but said they will never be completely reliable. She also pointed to a practical problem: each AI company uses its own watermarking system, so there is no universal method that catches every watermark. Devin Norder, director of product management at Meta, said media from recent frontier models cannot be told apart from authentic content by people, and that producing a convincing fake now takes little time or money.

The judges’ reaction was more measured. U.S. District Judge Jesse Furman of Manhattan, who chairs the committee, noted that while a great deal of academic commentary predicts a wave of deepfake disputes, such cases have not yet shown up much in practice. His Manhattan colleague Judge Valerie Caproni went further, arguing that courts already have the tools they need and that fake evidence is an old problem rather than a new one. The numbers support the caution: in a March survey by the Federal Judicial Center, the judiciary’s research arm, only 15 of the 931 judges who responded had handled a case in which a party challenged evidence as a deepfake.

The technical warning came from the AI companies; the scepticism about new rules came from the bench.

2. Why “detect it” is the wrong plan

It is unusual for vendors to tell a regulator their own safety tools have limits. Detection tools look for statistical traces or embedded watermarks. Both approaches have the same weakness: they only work if the fake was made by a tool that leaves a trace the detector knows about, and if nobody stripped the trace out. Open-weight image models, screenshots, re-encoding and simple cropping all erode those signals.

We have covered pieces of this before. OpenAI’s textGrain watermark for ChatGPT text showed how marking can work inside one company’s ecosystem, and Microsoft’s move to add deepfake detection to Teams after a €95 million voice-clone fraud showed real demand for it. But detection is a probability, not a proof. In a dispute, “our tool says 82% likely authentic” invites the obvious reply: whose tool, trained on what, and how often is it wrong? That is why the debate is shifting from spotting fakes to proving the real thing, which is a problem businesses can actually work on today.

There is also a second-order risk that lawyers call the “liar’s dividend”. Once everyone knows fakes are convincing, anyone caught on a genuine recording can claim it was generated. Our look at the proof-of-human problem described how quickly that doubt spreads online. In court, it means even honest evidence may need more support than it used to.

3. Where the federal rules actually stand

Much online coverage of these rules is out of date, so here is the status as reported from the committee’s own documents and by firms tracking them. There are two separate proposals.

Proposed Rule 707 deals with evidence that is produced by an AI system, such as an AI analysis of financial records or a machine-generated reconstruction, when it is offered without a human expert. The idea is to hold that output to the same reliability standards that Rule 702 applies to expert witnesses. It was approved for public comment in June 2025, and the comment period ran from August 2025 to 16 February 2026. At its 7 May 2026 meeting, the committee decided not to advance either the original text or any of three redrafts, and its October agenda book records that the published version has been withdrawn in favour of a modified proposal for further study. The leading redraft would normally require an expert to explain how the AI system produced its output.

Draft Rule 901(c) is the deepfake rule. It sets out a two-step process for exhibits accused of being fabricated by generative AI. First, the party making the accusation must produce enough evidence to support a finding that the item was faked; a bare claim is not enough. If it clears that bar, the burden moves to the party offering the item, who must show it is more likely than not authentic. The committee has kept this draft in reserve since late 2024 and has not finally approved it.

Proposal	What it covers	Key test	Status (October 2026)
Rule 707	AI or machine output offered as evidence without a human expert	Must meet Rule 702 reliability standards; leading redraft would usually need an expert to explain the system	Published version withdrawn; modified draft under study
Rule 901(c)	Exhibits challenged as AI-fabricated deepfakes	Challenger shows evidence of fakery, then proponent proves authenticity is more likely than not	Drafted, not finally approved; held for further study
Existing Rule 901	Authentication of all evidence	Enough evidence to support a finding the item is what its proponent claims	In force today and used for deepfake disputes now

Sources: Redgrave LLP (July 2026); RegLayer rule tracker (checked 26 Sept 2026); UIC Law Library. This is general information, not legal advice.

According to Redgrave LLP, the earliest that any AI evidence amendment could now be published for public comment is August 2027. After that, a new rule still needs approval from the Standing Committee, the Judicial Conference and the Supreme Court, followed by a review period in Congress. On the normal federal rulemaking calendar, our estimate is that a final rule would be unlikely to take effect before late 2029. If you read an earlier explainer on Rule 707 and the deepfake defence, this is the update: the rule is not law and its published wording is gone.

The rulebook is years away. Disputes over AI-faked exhibits are being decided case by case in the meantime.

4. The burden shift that matters for business

Even unfinished, draft Rule 901(c) is a useful map of where disputes are heading, because it shows how judges are thinking. Today, getting an exhibit admitted is usually easy: a witness says “that’s the photo I took” or “that’s the call I recorded”, and any doubts go to how much weight the jury gives it. The draft would raise the stakes once a credible challenge is made. At that point the side relying on the file has to prove, on the balance of probabilities, that it is genuine.

Think about what that means if you are the one holding the evidence. A security camera clip of a break-in, dashcam footage after a delivery-van crash, a recorded sales call, a photo of damaged goods on arrival: each is the kind of exhibit companies rely on in insurance claims, employment tribunals and commercial disputes. If the other side can produce an expert who says the file shows signs of manipulation, you will need more than a witness’s memory. You will need metadata, an unbroken chain of custody, original files and ideally a record of provenance created at the moment of capture.

Once a credible challenge lands, the side holding the file has to prove it is real.

5. California and the EU are moving faster

While the federal courts deliberate, lawmakers have gone ahead with labelling rules, which Anderson also referred to in her testimony. They won’t settle a courtroom dispute, but they change what evidence will exist.

In California, the AI Transparency Act took effect on 2 August 2026 after AB 853 pushed it back from 1 January. Covered generative AI providers must offer a free tool that lets people check whether content came from their system and that surfaces any provenance data it finds. From 1 January 2027, large online platforms with more than 2 million monthly users must detect provenance data that follows widely adopted standards and show users whether content was AI-generated or altered. From 1 January 2028, makers of cameras and other capture devices sold in California must give users the option to embed a hidden disclosure recording the manufacturer, the device and the time of capture. That last step is the important one for evidence: it builds proof of authenticity into the original file.

In the EU, Article 50 of the AI Act has applied since 2 August 2026. Providers must mark AI output in a machine-readable way so it stays detectable as generated, though under the AI Omnibus deal, systems already on the market before that date have until 2 December 2026 to comply. Businesses that publish deepfakes must disclose them, and that duty has no grace period. As our coverage of the EU’s first AI Act enforcement moves noted, Brussels is now acting on these deadlines.

Authenticity is easiest to prove when it is recorded at the moment of capture, not argued about later.

6. Who is most exposed

A low number of court cases does not mean low risk. Breitbart, citing FBI figures, reported more than 22,000 AI-related scam complaints in 2025 with nearly $893 million in losses. Voice cloning has moved from novelty to legal question too, as Japan’s courts showed with their landmark voice-rights ruling. The businesses most exposed to an “it’s fake” argument are those that settle disputes using media:

Insurers and claims handlers, who rely on photos of damage and injury and now face both faked claims and genuine claims dismissed as fakes.
HR and employment teams, whose misconduct investigations often turn on a screenshot, a message or a recording.
Retail, logistics and property firms, whose CCTV and dashcam footage is routine evidence in theft, accident and liability cases.
Finance teams, where a cloned voice or video call can authorise a payment, and the later dispute is over who really said what.
Anyone using AI to analyse evidence, such as AI transcription, enhancement or summaries of records, which is the territory Rule 707 is meant to cover. As we argued in “The AI did it” is no longer a defence, courts expect you to explain the tools you rely on.

7. The deepfake evidence playbook: five scenarios

Here are the situations most businesses will actually face, and what to do in each. Treat it as a starting point for your lawyers, not legal advice.

If you capture footage or recordings routinely (CCTV, dashcams, call recording, body cams): check whether your devices or software can sign files or embed content credentials, keep originals in write-once storage, and record a hash value whenever a clip is exported. Ask vendors now what provenance standard they support, because California will push device makers toward it from 2028.
If an incident happens and media may become evidence: preserve the original file, not a forwarded copy or a screenshot. Note who collected it, when and from which device. Avoid “enhancing” or AI-upscaling the original; if you need an enhanced version, keep both and document the tool used.
If someone claims your evidence is a deepfake: do not rely on a detection score alone. Assemble the proof stack: original file, metadata, hash, custody log, the capture system’s own logs and any witness who saw the event. Under the draft federal approach, those are what would carry the burden.
If you suspect evidence against you is fake: bring in a forensic examiner early and ask for the original file and its metadata, not a copy. Remember the draft rule’s first step: an allegation alone will not get far, so you will need specific signs of manipulation.
If a voice or video request moves money or data: require a call-back on a known number or a second approver before acting, and log the verification. That record protects you whether the request was real or cloned. For broader controls, see our guide to shadow AI in the workplace.

8. The bottom line

The most telling moment in Boston was not a new rule, because there wasn’t one. It was two of the world’s largest AI companies telling the judiciary that the problem they helped create cannot be solved by better detectors. The judges are right that deepfake fights are still rare, but formal rules are years away and fraud, insurance and employment disputes will not wait. The practical response is the same for every organisation: stop assuming a file speaks for itself, and start keeping the records that prove where it came from.

Frequently asked questions

What did OpenAI and Meta tell the federal judicial panel?

At the Advisory Committee on Evidence Rules meeting in Boston on 15 October 2026, OpenAI’s Shira Anderson said she does not expect a scientific method to reliably detect AI images in future and that company watermarks are not interoperable. Meta’s Devin Norder said people cannot tell recent AI media from authentic content.

How often do deepfake challenges come up in federal court?

Rarely so far. A March 2026 Federal Judicial Center survey found that 15 of 931 responding judges had handled a case where evidence was challenged as a deepfake.

Is Federal Rule of Evidence 707 in effect?

No. Proposed Rule 707 was published for comment from August 2025 to February 2026, but the committee did not advance it in May 2026 and the published version has been withdrawn in favour of a modified draft. The earliest any AI evidence amendment could be republished for comment is August 2027.

What is draft Rule 901(c)?

It is a proposed deepfake rule. A party claiming an exhibit was fabricated by generative AI would first have to produce evidence supporting that claim; if it did, the party offering the exhibit would have to show it is more likely than not authentic. The committee has not finally approved it.

Can AI detection tools prove a video is real or fake?

Not on their own. The AI companies themselves say detection will never be fully reliable and that watermarks differ between providers. Detection scores are best treated as supporting evidence alongside original files, metadata, hashes and chain-of-custody records.

What do California and the EU require for AI-generated content?

California’s AI Transparency Act, in force since 2 August 2026, requires covered AI providers to offer free detection tools, with platform duties from 2027 and a capture-device disclosure option from 2028. EU AI Act Article 50 has required machine-readable marking and deepfake disclosure since 2 August 2026, with a grace period to 2 December 2026 for marking by existing systems.