Everything here is dated and sourced. Laws in this area move monthly; check the changelog at the bottom for when this page was last updated.

The EU AI Act, Article 50: the core obligation

Article 50 of the EU AI Act (Regulation (EU) 2024/1689) is the transparency article. Its central marking duty, quoted verbatim from the statute [EUR-Lex, CELEX:32024R1689, accessed Aug 16, 2026]:

“Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated. Providers shall ensure their technical solutions are effective, interoperable, robust and reliable as far as this is technically feasible…” — Article 50(2)

Note the words: text is explicitly included, the marking must be machine-readable, and the outputs must be detectable as AI-generated. The statute is technology-neutral — watermarks, signed metadata, and fingerprinting can all qualify — but “we didn't mark it” is no longer a lawful option for providers of generative AI serving the EU.

Who owes what: providers vs. deployers

The article splits duties between the companies that build and supply AI systems (providers) and the organizations that use them (deployers):

Providers (Article 50(1)–(2)) must ensure that people interacting directly with an AI system know it is an AI (unless it's obvious), and must mark all synthetic audio, image, video, and text output machine-readably and detectably. The marking duty does not apply where the AI performs only an “assistive function for standard editing” or does not substantially alter the deployer's input — so a grammar checker is out of scope; a text generator is squarely in it [EUR-Lex, accessed Aug 16, 2026].

Deployers (Article 50(4)) must disclose deepfakes — AI-generated or manipulated image, audio, or video — with a lighter-touch regime for evidently artistic or satirical works. For text, deployers must disclose AI generation when the text “is published with the purpose of informing the public on matters of public interest” — think news, civic information — unless the content “has undergone a process of human review or editorial control” and a person or company “holds editorial responsibility” for it [EUR-Lex, accessed Aug 16, 2026]. In plain terms: a newsroom that uses AI drafting but has a human editor who signs off does not have to label every article; an unreviewed AI-written news feed does.

All disclosures must reach people “in a clear and distinguishable manner at the latest at the time of the first interaction or exposure” (Article 50(5)).

The three dates that matter

  • August 2, 2026 — Article 50 became applicable. New generative AI systems must mark their output now [European Commission FAQ, accessed Aug 16, 2026].
  • December 2, 2026 — end of the grace period. Under the May 2026 AI Omnibus agreement, generative systems already on the market before August 2, 2026 have until this date to meet the machine-readable marking requirement. (Interactive-disclosure and deployer deepfake duties applied from August 2 regardless.) [CSA research note, Jul 2026; ReedSmith, accessed Aug 16, 2026]
  • February 2, 2027 — the interoperable-detection deadline under the Code of Practice (below): providers must give the outside world a working route to check for their marks — a public access point, an embedded “signpost” in content, or a shared consortium solution [TechPolicy.Press; Stibbe, accessed Aug 16, 2026]

Penalties

Non-compliance with Article 50 carries administrative fines of up to €15,000,000 or 3% of total worldwide annual turnover, whichever is higher (for SMEs, whichever is lower), enforced by national market surveillance authorities [AI Act Article 99(4)(g), accessed Aug 16, 2026].

Scope clarifications from the Commission Guidelines (July 2026)

Final Commission Guidelines on Article 50 (published ~July 20, 2026) carve out AI-assisted translations, source code, very short outputs, and machine-to-machine or purely internal outputs; they define deepfakes broadly, read the creative-works exception narrowly, and confirm that content merely used in the EU can trigger territorial reach [ReedSmith, accessed Aug 16, 2026].

The Transparency Code of Practice: how industry is complying

Article 50(7) told the EU AI Office to facilitate a code of practice — and it did. The Code of Practice on Transparency of AI-generated Content was published June 10, 2026, and the Commission and AI Board found it adequate on July 8–9, 2026, making it the EU-recognised instrument for demonstrating Article 50 compliance [European Commission; IPTC, Jun 2026; ReedSmith, accessed Aug 16, 2026].

By July 31, 2026, roughly 190 organisations had signed. The provider section includes Anthropic, Google, Meta, Microsoft, Mistral, and OpenAI, alongside Aleph Alpha, Cohere, Black Forest Labs, and Synthesia; the deployer section spans companies from Getty Images to Lufthansa [European Commission signatory announcement, accessed Aug 16, 2026]. The one conspicuous absence: xAI, reported as the only major language-model maker that declined to sign — though it remains bound by the AI Act itself [European Commission signatory list; CityAM, Aug 12, 2026].

The Code accepts three marking techniques — digitally signed, time-stamped metadata (in practice, C2PA Content Credentials); imperceptible watermarking; and optional fingerprinting with a registry — and requires signatories to make detection mechanisms available, culminating in the February 2, 2027 interoperability deadline [IPTC; TechPolicy.Press, accessed Aug 16, 2026].

Two honest admissions inside the Code are worth quoting in spirit: no single marking technique currently meets all four statutory requirements (effective, interoperable, robust, reliable), and forensic detection of unwatermarked text — what AI detectors like this site do — “is not yet considered reliable enough” [TechPolicy.Press, accessed Aug 16, 2026]. We agree, and we say so on our methodology page.

California: SB 942 and AB 853 — and the text-shaped hole

California's AI Transparency Act (SB 942), as amended by AB 853 (signed October 13, 2025), became operative on August 2, 2026 — deliberately aligned with the EU date. Covered providers (over 1 million monthly users, publicly accessible in California) must offer a free public AI detection tool, optional visible disclosures, and embedded latent disclosures where feasible. AB 853 extends duties to large online platforms and GenAI hosting platforms (January 1, 2027) and capture-device makers (January 1, 2028). Violations run $5,000 per violation per day [aicomplianceatlas.com; Hintze Law, Oct 2025; Troutman, Oct 2025; accessed Aug 16, 2026].

The crucial nuance almost every summary misses: SB 942 covers image, video, and audio — not text. A student essay, a marketing blog post, a fake product review: none of it falls under California's marking or free-detection-tool mandate. In the United States, as of August 2026, our review found no federal or state law requiring AI-generated text to be marked — a conclusion from the absence of such a law in our searches, in a fast-moving legislative landscape, rather than from any authority saying so. Text marking is, for now, a European and Chinese legal reality and an American voluntary one.

China: labeling in force since September 2025 — text included

China's “Measures for Labeling of AI-Generated Synthetic Content” (issued by the Cyberspace Administration of China and three co-regulators, March 14, 2025) took effect on September 1, 2025, backed by mandatory national standard GB 45438-2025. They require explicit labels — visible indicators on AI-generated text, images, audio, video, and virtual scenes — plus implicit labels (embedded metadata, optionally watermarks) inside files. Platforms must verify labels, and users must self-declare AI-generated uploads [Loeb & Loeb, Mar 2025; ChinaLawTranslate; Bird & Bird, accessed Aug 16, 2026]. Unlike California, China's regime squarely covers text — and unlike the EU, it demands labels a human can see, not only machine-readable marks.

What this means for you

If you write — as a student, freelancer, or author. No law anywhere requires you to watermark your own writing, and nothing lets you prove a text is human-written. But the flip side matters: text you generate in the Gemini app or on the Gemini web experience (since 2024), or with new Claude models (since August 2026), carries a provider watermark that the provider can detect — plausibly forever. If you publish AI-generated text presented as your own in a context where that matters, assume it may one day be provable. And if you are accused of AI use based on a statistical detector score alone, know that even the EU's own Code of Practice deems such detection “not yet reliable enough” — a fact worth citing; our response playbook walks through exactly how.

If you deal with chatbots as a customer. Since August 2, 2026, Article 50(1) gives you the right to know when you are interacting with an AI system. What that right covers — and the behavioral tests that work when disclosure is missing — is in Am I talking to a bot?

If your business publishes content. If you operate in or into the EU and publish AI-generated text “to inform the public on matters of public interest,” you must disclose it — unless a human editor reviews it and someone holds editorial responsibility. That editorial-control exception is your practical compliance path: put a named human review step in your pipeline and document it. Deepfake-style image, audio, and video must be disclosed regardless. If you serve California at scale with image/video/audio generation, the SB 942 detection-tool and disclosure duties may reach you. And check your vendors: from December 2, 2026, every major model provider serving the EU should be marking text output — ask yours how.

If you build with AI models. If your product generates text, image, audio, or video for EU users, Article 50(2) applies to you as a provider — machine-readable marking is your duty, not just OpenAI's or Google's. If you build on a provider's API, do not strip their marks: preserve C2PA metadata through your pipeline, don't re-encode files unnecessarily, and don't run output through transformations that destroy watermarks if EU compliance matters to you. Watch February 2, 2027: the interoperable-detection mechanisms providers must ship by then will define the APIs your compliance tooling can call. The exemptions (assistive editing, source code, short outputs, internal machine-to-machine traffic) are genuine but narrow — read the Commission Guidelines before relying on one.

Sources

  1. Regulation (EU) 2024/1689, Article 50 (verbatim) and Article 99 — EUR-Lex, CELEX:32024R1689 (accessed Aug 16, 2026)
  2. European Commission, FAQ on Article 50 transparency obligations, digital-strategy.ec.europa.eu (accessed Aug 16, 2026)
  3. European Commission, Code of Practice on Transparency of AI-generated Content — policy page and “Strong backing” signatory announcement (Jul 2026; list updated Aug 12, 2026)
  4. IPTC, “EU AI transparency Code of Practice” (Jun 2026)
  5. TechPolicy.Press, “The EU's AI Transparency Code of Practice, Explained” (accessed Aug 16, 2026)
  6. ReedSmith, “Transparency obligations for AI-generated content: the Code of Practice adequacy” (accessed Aug 16, 2026)
  7. Cloud Security Alliance, research note on Article 50 and the AI Omnibus grace period (Jul 29, 2026)
  8. Stibbe, “The AI Act's transparency obligations” and “(Water-)marking the machine” (accessed Aug 16, 2026)
  9. CityAM, “ChatGPT might follow Claude's watermark pledge — but Grok to swerve it” (Aug 12, 2026)
  10. California SB 942 / AB 853: aicomplianceatlas.com; Hintze Law (Oct 19, 2025); Troutman Amin (Oct 2025)
  11. China labeling measures: Loeb & Loeb (Mar 2025); ChinaLawTranslate; Bird & Bird (accessed Aug 16, 2026)