The EU AI Act Article 50 marking and detection obligations apply from 2 August 2026, with separate disclosure duties for deepfakes and certain public-interest text. China’s explicit and implicit AI-content labelling measures have applied since 1 September 2025. In the United States, California’s AI Transparency Act has applied since 2 August 2026 to large generative AI providers, with duties for large platforms from 1 January 2027 and for camera and phone makers from 1 January 2028; there is still no general federal watermark mandate.
The strongest reading of the current sources
EU Regulation 2024/1689 requires providers of systems generating synthetic audio, image, video, or text to mark outputs in a machine-readable and detectable way where technically feasible. The European Commission’s final voluntary Code of Practice offers a compliance path but does not replace the binding Article 50 duties. China’s four-agency measures require visible and implicit labels, including file-metadata marking, and prohibit malicious removal, alteration, forgery, or concealment. California’s AI Transparency Act (SB 942, amended by AB 853) requires generative AI providers with over 1,000,000 monthly users to embed a latent disclosure (provider name, system name and version, creation or alteration time, and a unique identifier), offer an optional visible disclosure, and provide a free detection tool. From 1 January 2027, platforms with over 2,000,000 monthly users must detect and display provenance data and must not knowingly strip it where technically feasible; from 1 January 2028, capture-device makers must embed a latent disclosure by default. At federal level, the 2025 COPIED Act text proposes standards, provenance capabilities, and protections, but a bill is not law merely because it was introduced. China’s rules took effect on 1 September 2025; the EU Article 50 and California dates were 2 August 2026.
A legal disclosure can be visible, machine-readable, or both. Technical metadata removal is not the same as legal permission to remove a required label. Obligations depend on jurisdiction, role, content type, use, and exceptions; this research brief is general information, not legal advice.
See the source behind the answer
These are direct source links, not citations to other SEO summaries.
Regulation (EU) 2024/1689, Article 50
Article 50 requires technically feasible machine-readable marking and detectability for synthetic audio, image, video, and text outputs, plus specified deployer disclosures for deepfakes and public-interest text.
Code of Practice on Transparency of AI-generated Content
The Commission confirms that Article 50 transparency duties apply from 2 August 2026 and describes the final voluntary code as an adequate compliance tool for signatories.
Measures for Labelling Artificial Intelligence-Generated and Synthesized Content
China’s four-agency measures require explicit and implicit labels, including file-metadata markers, and prohibit malicious deletion, alteration, forgery, concealment, or tooling that enables those acts.
S.1396 — Content Origin Protection and Integrity from Edited and Deepfaked Media Act of 2025
The introduced COPIED Act text proposes NIST provenance standards, creator access to content-provenance capabilities, and protections against unauthorized removal, but the cited text is a legislative proposal rather than an enacted general federal mandate.
SB 942 — California AI Transparency Act
Covered providers must offer a free AI detection tool and embed a latent disclosure conveying the provider name, the system name and version, the creation or alteration time and date, and a unique identifier, permanent or extraordinarily difficult to remove where technically feasible. Violations carry a civil penalty of $5,000 each, with each day a separate violation.
AB 853 — California AI Transparency Act amendments
AB 853 makes the act operative on 2 August 2026; from 1 January 2027 it requires large online platforms to detect and disclose provenance data and not knowingly strip system provenance data or digital signatures where technically feasible, and from 1 January 2028 requires capture-device makers to embed latent disclosures by default.
What remains uncertain
- Application can turn on detailed definitions, exceptions, territorial reach, implementation guidance, and whether a system merely assists editing or substantially changes content.
- US federal and state proposals can change status, and sector-specific, election, impersonation, consumer-protection, and platform rules may apply even without a single general watermark statute.
- California’s duties fall on large providers, platforms, hosting services and device makers above stated thresholds, not on individuals who edit or share a file; its platform and device duties start on 1 January 2027 and 1 January 2028.
- Technical standards and provider implementations evolve faster than legislation; compliance may require more than finding or adding one metadata tag.
What to do before you share
- 01
Keep original files and provenance records when a disclosure or audit trail may be legally required.
- 02
Do not use metadata cleaning to evade a required AI label, provenance marker, platform rule, evidence hold, or contractual obligation.
- 03
For EU-facing generated content, map provider marking duties and deployer disclosure duties separately; they apply to different actors and outputs.
- 04
Get jurisdiction-specific legal advice before designing a compliance workflow or removing markers from regulated content.
Quick clarifications
Does the EU AI Act require watermarks on AI text?
Article 50(2) covers synthetic text as well as audio, images, and video, requiring machine-readable marking and detectability where technically feasible. Article 50(4) separately addresses disclosure for AI-generated or manipulated text published to inform the public on matters of public interest, subject to its conditions and exceptions.
Is it illegal to remove an AI watermark?
There is no universal answer. China’s cited measures prohibit malicious removal or concealment of required labels. From 1 January 2027, California bars large online platforms from knowingly stripping system provenance data where technically feasible; that duty applies to the platform, not to a person sharing a file. Other jurisdictions and contexts differ, and contracts, platform rules, evidence duties, copyright-related provenance protections, or anti-deception law may also matter. Obtain legal advice for the specific use.
Does any US law require AI watermarks?
There is no general federal mandate. California’s AI Transparency Act has applied since 2 August 2026: generative AI providers with over 1,000,000 monthly users must embed a latent disclosure in AI images, video and audio and offer a free detection tool. Large platforms must show provenance data from 1 January 2027, and camera and phone makers must embed disclosures by default from 1 January 2028.
Is the EU Code of Practice itself mandatory?
The Code is voluntary, but the underlying Article 50 transparency requirements are legal obligations. The Commission describes adherence as a way for signatories to demonstrate compliance.