OpenAI and Meta tell U.S. advisory panel no universal deepfake detector exists

OpenAI and Meta told a U.S. judiciary advisory panel that there is no universal, court-ready way to determine whether every image was made by artificial intelligence, underscoring the difficulty courts face in authenticating digital evidence. An OpenAI lawyer told the panel she did not believe a scientific method would ever reliably detect AI-generated images, while Meta representatives said their company-specific tools did not amount to a test that could establish authenticity across the board. The discussion took place Oct. 15 in Boston, where the Advisory Committee on Evidence Rules convened experts to consider deepfakes and possible amendments to the Federal Rules of Evidence. The committee had spent seven previous meetings developing a response to the evidentiary problems posed by manipulated material. Its draft proposal, Rule 901(c), had not been finally approved. Under the version described in the committee’s agenda, once an opponent offered sufficient evidence that an item was a deepfake, the party seeking to use it would have to establish that it was more likely than not authentic. The testimony addressed the limits of image detection rather than the adoption of that proposed rule, Tech Insider reported. The committee’s deliberations were still under way and no final rule had been adopted. The problem extends beyond courtroom procedure. The National Institute of Standards and Technology has said generative AI can create realistic images, text, audio, video and multimodal material, posing risks to trust, safety, transparency and credibility. NIST defines digital-content transparency as the ability to document and access information about a file’s origins and history, but cautions that transparency does not itself guarantee trustworthiness. NIST identifies provenance tracking, detection, labeling and authentication as potential technical responses. None provides a comprehensive solution alone, it said, and the effectiveness of many approaches has not yet been fully assessed. For high-risk uses requiring high accuracy, including investigations into non-consensual intimate imagery, the agency said a defense-in-depth approach may be useful. That layered model can combine provenance metadata, invisible watermarking and forensic analysis. C2PA, or the Coalition for Content Provenance and Authenticity, uses cryptographically signed manifests that can record such details as the originating device, software, time and declared AI involvement. The specification reached version 2.3 in January. Watermarks take a different approach. Google DeepMind’s SynthID embeds a signal within media, rather than preserving a provenance record as metadata does. Google has deployed it in its Imagen image, Veo video and Lyria audio models, and launched a detector portal for uploaded images, audio, video and text. In May, OpenAI said it would add SynthID to images generated through ChatGPT, Codex and its API, while describing it as a complement to metadata-based provenance rather than a replacement. Several of the technical measures have limits. Metadata can be stripped when a file is exported, watermark signals can deteriorate under compression, and classifiers can lag new generation techniques. Open or lightly restricted image-generation models may also circulate without watermarks or provenance manifests.