news

OpenAI and Meta tell a US evidence-rules panel that no method reliably detects AI video

At a Boston meeting on October 15, OpenAI's policy counsel said she does not expect a scientific method for detecting AI images to exist, while a March survey found 15 of 931 federal judges had ever seen a deepfake challenge.

By Daniel Ochoa · 16 October 2026

Two of the companies that build the largest image and video generators spent Thursday telling a federal rulemaking body that their own output cannot be reliably identified. Representatives from OpenAI and Meta Platforms appeared before the US Judicial Conference's Advisory Committee on Evidence Rules at a meeting in Boston on October 15, and the message was that verifying the authenticity of audio and video evidence will keep getting harder rather than easier.

OpenAI's policy counsel does not expect a scientific detector to arrive

Shira Anderson, policy counsel at OpenAI, told the committee that AI image quality keeps improving and that she does not believe a scientific method capable of detecting such images will exist in the future. Devin Norder, director of product management at Meta, told the same panel that media produced by recent AI models cannot be distinguished from authentic content by humans, and that producing it now takes minimal time and cost. Those are not hedged positions from outside critics, they are the vendors describing the limits of their own category.

Watermarking laws exist, but each company runs its own system

New laws in California and the European Union require watermarking aimed at making synthetic media detectable, and both OpenAI and Meta have built tools to verify visual content. Anderson's caution was about the gap between those two facts: every company uses its own system, so there is no universal way to reliably detect all watermarks or verify the authenticity of a given file. She said the systems will keep improving and will never be completely reliable.

That is the same wall anyone scoring generated media runs into from the other side. In our own benchmark we score face rendering across models precisely because that category is where synthetic output used to give itself away, and the visible tells have been shrinking release over release across Veo 3.1 and its competitors.

15 of 931 judges have actually seen a deepfake challenge

The courtroom pressure is still mostly theoretical. A survey conducted in March found that 15 out of 931 responding federal judges had encountered a case where a party challenged evidence on the grounds that it was a deepfake. US District Judge Jesse Furman of Manhattan, who chairs the panel, noted that academic commentary predicts courts will have to deal with deepfakes, but that the issue has not yet appeared frequently in real cases. The committee has been working on the question since 2023, including on whether AI-generated evidence should meet the same reliability standards as testimony from human expert witnesses.

Detection is just judging with a ground truth attached

A deepfake detector is an evaluator with a correct answer available: real or generated. What the labs described in Boston is an evaluator that cannot hold a stable accuracy figure as generators improve, and whose fallback, provenance metadata, is fragmented across vendors rather than standardized. For a leaderboard that mislabels a clip, the cost is a wrong rank. For an evidentiary hearing, the cost is different, which is why the panel is drafting rules rather than waiting for a classifier.

The volume side of the problem is already measurable elsewhere. Detection work at scale has grown fast enough that one charity assessed 6,310 AI-generated abuse images in six months, 40% above its total for all of 2025, with human analysts doing the classification that no model reliably automates.

Sources: Investing.com (Reuters), Yahoo Finance.

Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.