Synthetic content labeling and provenance system
synthetic-content-labelingDomain: ai-transparencyType: mixedDescription
Synthetic-content labeling is the rare AI obligation where the regulation specifies a wire format. EU AI Act Article 50 reaches for machine-readable provenance and points implementers at C2PA-style content credentials by name in the recitals; California SB 942 splits the duty into a visible label and a latent disclosure embedded in the file itself; China's deep-synthesis regime requires both a perceivable marker and a registry-side filing. The labeling regimes are converging on the same architecture before the statutes have fully converged on the language. A working labeling system has three loosely-coupled pieces. The provenance layer is the cryptographic manifest that records what model produced what output, typically a C2PA content credential signed by the generator and carried in the file's metadata; the layer of the question is which signing identity is used (the model vendor's, the platform's, or the end user's) because each choice has a different downstream verification story. The disclosure layer is whatever user-visible surface the regulation actually demands: a watermark drawn into the image, an alt-text or caption that names the output as AI-generated, a terms-of-service section that informs users the platform produces or hosts AI-generated content. The pipeline layer is the export path that makes sure both the provenance manifest and the visible disclosure survive the round trip through re-encoding, social-sharing intermediaries, and the screenshot-and-repost shortcut that breaks most metadata-only schemes. Operators commonly get the provenance layer cleanest because the C2PA tooling is mature and well-documented; the visible disclosure tends to drift because product teams reasonably resist a permanent watermark on user-facing creative output; the pipeline layer is where the real engineering work hides. The deadlines are sharp. California SB 942 covers generative AI providers serving over one million California users and becomes operative on 2026-01-01, with both visible and latent disclosure requirements on covered output. EU AI Act Article 50 covers providers of generative AI systems and applies from 2026-08-02; the implementing technical standard work at CEN-CENELEC will define what counts as a sufficient machine-readable mark. The Take It Down Act gives platforms 48 hours from a verified non-consensual intimate imagery report to remove the content, regardless of whether it is AI-generated, which is a useful upper bound on how fast the takedown side of the labeling stack has to run. China's algorithm-recommendation and deep-synthesis provisions have been operative since 2022-03-01 and 2023-01-10 respectively, with perceivable-marker requirements on synthetic content distributed inside the PRC. What is still genuinely unsettled is which provenance signal regulators will actually inspect. The C2PA standard is the obvious convergence point and the EU recitals nod toward it, but no statute hard-codes a specific manifest format, and providers have begun shipping their own variants. The interoperability question (does a Sora output's C2PA manifest survive being imported into a third-party editor and re-exported?) is the operational question the labeling stack has to answer, and the answer today is mostly no.
Applicability
Applies when: features include ai-content-gen or synthetic-media.
Required by (8 regulations)
- EU AI Act
Article 50(2) requires providers of generative AI systems to ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated; Article 50(4) adds a deployer-side duty to disclose AI-generated text on matters of public interest and deepfakes.
Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (AI Act); transparency obligations under Article 50 apply from 2026-08-02
- California SB 942
The California AI Transparency Act requires covered generative AI providers (over one million California users) to offer a free AI-detection tool and to include both a manifest-style latent disclosure and an option for a clear and conspicuous visible disclosure on AI-generated image, video, and audio content.
California AI Transparency Act, SB 942 (2024); operative 2026-01-01
- Algorithm Provisions
The deep-synthesis provisions (read together with the algorithmic-recommendation provisions) require providers of deep-synthesis services to add perceivable marks to generated content that could cause public confusion and to keep non-perceivable identification information embedded in the file.
Provisions on the Administration of Deep Synthesis Internet Information Services (CAC, MIIT, MPS; effective 2023-01-10)
- TAKE IT DOWN
Covered platforms have 48 hours from a verified report to remove non-consensual intimate imagery, including AI-generated depictions, and to make reasonable efforts to remove identical copies; the labeling stack's takedown integration runs on this clock.
Tools to Address Known Exploitation by Immobilizing Technological Deepfakes On Websites and Networks Act (Take It Down Act), Pub. L. 119-12 (2025)
- California AB 1836
California AB 1836 restricts digital replicas of deceased personalities and supports labeling AI-generated likenesses.
California AB 1836 (2024), amending Cal. Civ. Code §3344.1
- China Content Review
China's deep-synthesis and content rules require labeling of AI-generated or synthetic content.
Administrative Measures for the Administration of Online Publishing Services (NPPA and MIIT Order No. 5, effective March 10, 2016); culture-side rules on online cultural products and online-game content self-examination and recordkeeping; NPPA content-review notices
- FTC AI Advertising
FTC guidance treats undisclosed synthetic or AI-generated endorsements as potentially deceptive.
15 U.S.C. §§41-58 (FTC Act §5); FTC business-guidance documents including 'Keep Your AI Claims in Check' (April 2023), 'Aiming for truth, fairness, and equity in your company's use of AI' (April 2021), Endorsement Guides (16 CFR Part 255, updated 2023)
- Tennessee ELVIS Act
The Tennessee ELVIS Act protects voice and likeness against unauthorized AI replicas, supported by synthetic-content labeling.
Tenn. Code Ann. §§ 47-25-1101 to 47-25-1108 (Ensuring Likeness Voice and Image Security Act of 2024, amending the Personal Rights Protection Act)
Evidence formats
- C2PA content credential signing implementation + manifest sample
- visible AI-generated content label / watermark UI specification
- metadata-embedding pipeline diagram (generation, export, re-encode survival)
- Terms of Service section disclosing AI-generated content presence
- latent disclosure (machine-readable mark) technical specification
- round-trip provenance survival test plan