NIST surveyed synthetic-content tools without mandating any

The shot

NIST's Editorial Review Board approved NIST AI 100-4, 'Reducing Risks Posed by Synthetic Content', on 18 November 2024, and the document's cover dates its publication to November 2024. Subtitled 'An Overview of Technical Approaches to Digital Content Transparency,' it sits in NIST's Trustworthy and Responsible AI series, separate from the agency's AI Risk Management Framework Generative AI Profile, which NIST's own AI Risk Management Framework page dates to 26 July 2024, and from the base AI RMF 1.0, dated on the same page to 26 January 2023.

What the documents show

The report states its own scope directly: it surveys existing standards, tools, methods and practices across six areas — authenticating content and tracking provenance, labeling synthetic content such as through watermarking, detecting synthetic content, preventing generative AI from producing child sexual abuse material or non-consensual intimate imagery, testing software used for those purposes, and auditing and maintaining synthetic content over time. It adopts the definition of 'synthetic content' from Executive Order 14110: 'information, such as images, videos, audio clips, and text, that has been significantly altered or generated by algorithms, including by AI.' NIST's framework page separately confirms the Generative AI Profile and base framework are distinct, differently dated publications a reader should not conflate with AI 100-4's synthetic-content survey.

The workflow

The report organizes its material into two tracks: provenance data tracking, meaning recording watermarks or signed metadata about a file's origin and edit history, and synthetic content detection, meaning classifying whether content is synthetic, whether by reading recorded provenance or other signals. It states the two overlap in practice — a covert watermark is only useful if a detector can read it — and that pairing a watermark with signed metadata can offer complementary protection. NIST frames this as a landscape survey rather than a required control, stating the report 'focuses on technical approaches' and that 'normative, educational, regulatory, and market-based approaches not described in this report' may also be necessary.

What the tool does not change

NIST's own text states digital content transparency 'may contribute to trustworthiness but does not guarantee it, and in some cases may undermine it' — for example when legitimate content is later presented out of context. That is NIST's own caution: no watermarking or provenance scheme the report describes substitutes for a publication's own editorial verification, and the report does not present itself as a rule any platform is required to follow.

• Does the cited NIST document mean AI 100-4's synthetic-content survey, the Generative AI Profile, or the base Risk Management Framework, given all three carry different dates?
• Is a vendor's watermarking claim checked against one of NIST's named technique categories, or only asserted in general terms?
• Does a workflow relying on these techniques also address the non-technical approaches NIST says the report does not cover?

AI 100-4 is NIST's own survey of what content transparency can and cannot do today, not a mandate. Reading it against the agency's other, differently dated AI publications matters before citing 'NIST guidance' for any specific practice.