On November 20, 2024, the National Institute of Standards and Technology (NIST) announced the publication, by the U.S. Artificial Intelligence (AI) Safety Institute (AISI), of the report NIST AI 100-4: Reducing Risks Posed by Synthetic Content - An Overview of Technical Approaches to Digital Content Transparency.

The report identifies voluntary approaches to address risks from AI-generated content.

The report examines existing standards, methods, tools, and practices, in light of the development of generative AI, for:

• the authentication of content and tracking its provenance;
• labeling synthetic content through watermarking;
• detecting synthetic content;
• preventing generative AI from producing child sexual abuse material (CSAM) or non-consensual imagery;
• testing software used for the above; and
• auditing and maintaining synthetic content.

The report clarifies that the impact of transparency depends on the effectiveness of the technical methods used and how people interact with digital content. The report uses the same definition for 'synthetic content' as provided under Executive Order 14110 on the Safe, Secure, and Trustworthy Development and Use of AI. The report also applies concepts given under the NIST AI Risk Management Framework (RMF).

The report frames the risks and harms stemming from synthetic content within the pipeline of creation, publication, and consumption of synthetic content. Digital content transparency tools are broken down into:

• provenance tracking tools that record information about the origins and history of digital content:
  • such as covert or overt digital watermarks; or
  • cryptographically signed or unsigned metadata; and
• synthetic content detection tools that classify whether content is synthetic or not, such as:
  • provenance data detection methods;
  • automated content-based detection; or
  • human-assisted detection methods.

The report considers the benefits of different provenance methods and tools in detail. For instance, regarding covert digital watermarking, the report outlines:

• technical methods for covertly watermarking synthetic content;
• technical trade-offs for watermarking;
• the robustness and security of watermarking;
• scale considerations for watermarking; and
• privacy considerations for watermarking.

Likewise, for metadata recording, the report provides information on:

• linking content to external metadata;
• authenticating metadata through cryptography;
• example provenance metadata specifications;
• privacy considerations and security considerations for metadata recording; and
• scale considerations for metadata recording.

You can read the press release here and the report here.