Key Notes
- OpenAI and Meta told a US judicial evidence panel that no universal system can reliably authenticate every AI-generated image or video.
- Watermarks and provenance tools can add useful clues, but their absence or presence does not conclusively establish a file's origin, accuracy or context.
- The judiciary is considering authentication changes for suspected deepfakes, while those disputes remain relatively rare in court.
OpenAI and Meta have warned a federal judicial advisory committee that courts may not get a universal scientific answer to one of the hardest questions raised by generative AI: whether a photograph, video or audio clip is authentic. The message was not that every AI-made file is impossible to investigate. It was that detection systems, watermarks and visual inspection may not deliver a single conclusive test that works across every tool and every piece of evidence.
The warning came as the US judiciary considers how its evidence rules should respond when a party says that an apparently ordinary recording is a deepfake. As reported, OpenAI policy counsel Shira Anderson told the committee that she does not expect a scientific method to emerge that can reliably identify all AI-generated images. Meta’s Devin Norder similarly warned that current AI media can already be difficult for people to distinguish from authentic material.
A warning about authentication, not a magic detector
That distinction matters. A detector can be useful evidence without being a final verdict, particularly after an image has been cropped, compressed, screenshotted or passed through another editing tool. Provenance signals also have a narrower job than many users assume: OpenAI says its own provenance tools can identify supported OpenAI signals, but they cannot reliably judge material from every other generator or prove that a file’s surrounding claim is true.
The companies are therefore describing a fragmented technical landscape, not announcing that courts must give up on verification. A challenge to digital evidence can still turn on where it came from, who preserved it, the testimony around it, metadata, account records and the consistency of the underlying facts. But that process is slower and more fact-specific than pressing a button labeled “AI check.”
Why watermarks do not settle a courtroom dispute
Watermarking and content credentials can make useful clues available when they survive distribution. They can show that a supported system attached a particular signal to a file, and they can help platforms or investigators prioritize material for review. But a missing signal does not prove that a clip is human-made, while a surviving signal does not establish authorship, accuracy or context.
That is why the question reaches beyond the latest detector benchmark. Google recently opened its SynthID detector to wider public use, while OpenAI has outlined invisible text provenance measures for the EU. Those projects can improve transparency for their respective systems, yet neither creates a shared verification layer for the entire internet.
The evidence rules are still being written
The Advisory Committee on Evidence Rules has been examining fabricated digital evidence for several years. Its agenda materials describe a proposed Rule 707 that would address the reliability of acknowledged machine-generated evidence. The committee has also considered a more direct authentication procedure for an item challenged as a potential deepfake, but that work remains under consideration rather than final law.
For now, deepfake disputes are still relatively uncommon in courtrooms. Reuters reported that only 15 of 931 judges who responded to a March Federal Judicial Center survey had encountered a party challenge evidence as an AI deepfake. The legal system is trying to build its guardrails before the volume of convincing synthetic evidence makes those disputes routine.
The practical takeaway is not that video and images have become worthless. It is that their reliability will increasingly depend on a fuller record: provenance where available, careful preservation, corroborating evidence and an adversarial process that can test a claim before a judge or jury sees it. That may be less elegant than a universal detector, but it is closer to how courts already decide whether contested evidence deserves trust.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
