OpenAI and Meta Tell US Judges Deepfakes Won’t Be Reliably Detected

What OpenAI and Meta Told US Judges About Deepfake Evidence

OpenAI’s Shira Anderson doubted that AI images will be scientifically detectable going forward and noted that each company uses its own watermark and there is no universal way to verify them, while Meta’s Devin Norder said people can no longer tell recent AI media from the real thing.

The deepfake panel opened the committee’s fall meeting, a mini-conference set up in May 2026 to hear experts before any vote; the committee’s agenda book lists academics, NIST, a forensic examiner and federal defenders alongside the two companies.

OpenAI’s Shira Anderson: No Scientific Way Ahead to Detect AI Images

Shira Anderson is policy counsel at OpenAI, where she works on AI regulation and content provenance and collaborates with the C2PA, the coalition behind Content Credentials. On images, she was direct: “AI images are only getting better, and I do not think that there will be a scientific way going forward to detect them.”

She also pointed to a structural problem: each company uses its own watermarking system, and there is no universal method to verify them, Investing.com reported. On watermarks themselves she offered a guess rather than a forecast: “My guess is that this absolutely will continue to get better. I do not think it’ll ever be bulletproof.”

Meta’s Devin Norder: Indistinguishable to People, Trivial to Produce

Devin Norder is a director of product management at Meta, where she leads trust, safety and risk for generative AI and has worked on C2PA and on watermarking for Meta’s models. Her point was about human perception: “Media generated from frontier models, at least in the last few generations, are indistinguishable by humans from authentic or AI-generated media.” She added that “the cost and the time it takes to create this media really are quite trivial at this point.”

Earlier research points the same way. In a 2022 study in PNAS co-authored by Hany Farid, who sat on the same Boston panel, 315 participants tried to tell AI-synthesized faces from real ones and averaged 48.2% accuracy, slightly below chance. Those were faces made with the models of the time.

Software is not closing the gap. One recent benchmark found that detection accuracy falls by up to 49% over a single year of AI video progress.

Why an AI Watermark Is Not Enough to Prove What Is Real

An AI watermark proves something only when it is present and readable, and its absence says nothing about whether a file is authentic.

An AI watermark is a signal a generator embeds in its output, usually invisible to the eye, and laws on both sides of the Atlantic now require one, placing the duty first of all on the companies that provide the generators. California’s AI Transparency Act (SB 942, as amended by AB 853) became operative on August 2, 2026: providers of generative AI systems with more than 1,000,000 monthly users in California must embed a latent disclosure in the images, video and audio those systems create. Text is not covered. AB 853 also extends obligations to large online platforms from January 1, 2027, and to makers of capture devices such as cameras and phones from January 1, 2028.

In the EU, Article 50(2) of the AI Act requires providers of systems that generate synthetic audio, images, video or text to mark outputs “in a machine-readable format and detectable as artificially generated or manipulated,” starting August 2, 2026. Under Regulation (EU) 2026/1744, systems already on the market before that date have until December 2, 2026 to meet the marking duty. Article 50(4) separately requires deployers who use an AI system to create deepfakes to disclose that the content was artificially generated or manipulated.

A mandatory mark is still not a universal one. Each provider picks its own technology, and a reader built for one mark cannot see another. The memo the committee’s reporter prepared for the Boston meeting says it plainly: “The EU Act does not command the use of one universal watermark, nor does it establish a forensic presumption of machine authorship.” Google’s SynthID check shows the limit in practice, because it finds Google’s own watermark and nothing else.

A file with no watermark may be authentic, may come from a generator that does not mark its output, or may have had its mark degraded by editing. TrueScreen treats the absence of a watermark or of Content Credentials as missing information, never as proof of authenticity or of tampering.

For evidence you create yourself, the question shifts from how to unmask a fake to how to demonstrate what is real. TrueScreen acquires photos, videos and audio with a forensic methodology at the moment they are created and certifies them with a qualified electronic seal and a qualified timestamp.

Feature

Capture, analyze and certify any online content.

TrueScreen’s Forensic Browser certifies web pages at the moment of capture, and Media Inspector analyzes files that already exist.

Discover more →

Rule 901(c) and Deepfake Challenges: Who Must Prove What

Draft Rule 901(c) would make a party claiming a deepfake show real grounds first, then require the party offering the evidence to prove it is more likely than not authentic, but it is still an unapproved working draft.

The Two-Step Mechanism

Proposed Rule 901(c) is a working draft amendment to Rule 901 of the Federal Rules of Evidence for items challenged as fabricated, in whole or in part, by generative AI. First, the challenger must present evidence sufficient to support a finding of fabrication; as the committee’s reporter puts it, “a simple claim of ‘fake’ is insufficient.” If that showing is made, the item is admissible only if its proponent demonstrates that it is “more likely than not authentic,” a higher bar than ordinary authentication.

The reporter, Professor Daniel Capra, has also proposed adding to the committee note a list of factors a court could consider at the first step: missing or anomalous metadata, inconsistent creation dates, no native file, gaps in the chain of custody. That list is his proposal, not the text of the rule or a position the committee has taken.

None of it is law. Rule 901(c) is a working draft that the committee “has prepared, but not finally approved,” and it has never been published for public comment. No official record of what happened at the Boston meeting has been published yet. On May 7, 2026, a divided committee deferred 901(c) and the revised Rule 707 on machine-generated evidence to the fall mini-conference that became the Boston meeting. Judge Jesse Furman, who chairs the committee, proposed the deferral even though it meant “a delay of another year,” and noted that even a rule published in summer 2026 would not have taken effect before December 1, 2028.

Why Judges Are in No Hurry

Few federal judges have actually seen deepfake evidence challenged in court. In a Federal Judicial Center survey reported in the committee’s May 2026 minutes, 931 federal trial judges responded (a 45% response rate) and only 15 had faced a deepfake challenge. Among the rest, 82% said they would require an initial showing before examining a claim of fabrication, which is what the draft’s first step does.

Furman said as much in Boston: “There’s a lot of academic commentary and other commentary predicting that courts will be dealing with deepfakes, but we haven’t actually seen a whole lot of that yet.”

Judge Valerie Caproni, a committee member, urged colleagues not to rush into an AI-specific rule without stronger evidence that deepfakes are a widespread problem in court, according to IBTimes UK. “We’ve dealt with bad actors presenting fake evidence since Cain and Abel,” she said, as quoted by Reuters. “That’s a longstanding problem.”

What Changes for Anyone Who Produces Digital Evidence

Whatever happens to Rule 901(c), digital evidence that can show how, when and with what it was captured is easier to document and verify in a challenge than evidence whose only protection is that nobody spots a fake.

In US federal courts, Rule 901(a) already requires the proponent to produce evidence sufficient to support a finding that an item is what the proponent claims. Among its examples, Rule 901(b)(9) lists “evidence describing a process or system and showing that it produces an accurate result.” That is context for how a documented capture process can support authentication, not a guarantee of admission.

In the EU, the eIDAS Regulation gives a qualified electronic seal “the presumption of integrity of the data and of correctness of the origin of that data” (Article 35(2)), and a qualified electronic time stamp “the presumption of the accuracy of the date and the time it indicates and the integrity of the data” (Article 41(2)). These presumptions cover the integrity of the file, its origin from whoever sealed it and its date; neither attests that what the file shows is true, and courts weigh both case by case.

AI Act Article 50 deals with marking synthetic content and making it recognizable; it does not ask anyone to prove that authentic content is real. For the capture itself, ISO/IEC 27037 sets international guidelines for identifying, collecting, acquiring and preserving digital evidence.

1. 1

   Capture With a Verifiable Method

   Record photos, video and audio with a process you can describe and someone else can check, instead of pulling files later from a phone gallery or a chat. A missing native file or odd metadata, two of the factors the reporter proposes courts look at, is what undocumented capture tends to leave behind.

2. 2

   Fix the Date and Integrity at the Moment of Capture

   A SHA-256 hash calculated at capture, bound to a qualified timestamp and sealed, lets anyone check later when the file existed and that it has not changed. A timestamp added months afterward fixes the content only from that later date.

3. 3

   Keep What a Third Party Needs to Verify

   Preserve the native file, its metadata, the capture context (time, location, device, network) and a record of who handled it. An opposing expert should be able to check integrity and origin without taking your word for it.

How Do You Prove a Photo or Video Is Authentic Before It Is Challenged?

TrueScreen certifies photos, videos and audio at the moment they are captured: forensic acquisition prevents alteration while the content is recorded, and a qualified electronic seal with a qualified timestamp makes any later change detectable.

TrueScreen is the Data Authenticity Platform, and in its methodology the forensic work comes before the seal. Content is acquired in a controlled environment, so that no person, software, system or AI can alter it during capture, and its context (time, location, device, network) is collected with it. The data is then verified and a technical report is produced. Only at the end is the package sealed with an official digital seal and a qualified timestamp, recognized internationally. The chain of custody stays unbroken, and the process is aligned with the ISO/IEC 27037 guidelines.

That order matters. Anyone can seal a file, and sealing proves a file has not changed since the seal went on; it does not prove the file was true before it arrived. Anyone who receives a certification can verify its integrity and seal with standard tools, without trusting the person who delivered it.

The TrueScreen app acquires photos, videos, audio and screen recordings in real time and certifies them with this methodology. The Forensic Browser on macOS and Windows does the same for web pages, social media and browsing sessions, while on a computer the Portal certifies files and the screen, then stores the certifications for verification and sharing.

For content it did not acquire, TrueScreen also offers deepfake detection for photos, video and audio, available on every TrueScreen plan and reachable today in the Forensic Browser. It reports how likely a manipulation is: detection is an additional signal; certification at the source documents how and when the content was captured.

There is a limit: TrueScreen guarantees the how, not the what, so a staged scene is certified as an authentic recording of a staged scene. Certified content carries full legal and evidentiary value worldwide, yet the court weighs it case by case, and no certification can promise the outcome of a proceeding.

Take a construction company that photographs site damage with the app before handing over a building. A year later, the other side claims the photos were made with AI. The certification package shows how and when each image was acquired, and the opposing expert can examine that instead of a file of unknown origin. A file of unknown origin produces an opinion; a file sealed at acquisition produces a defensible result.

Conclusion