EU AI Act, Article 50
The EU AI Act's transparency obligations, from the primary sources
Who must mark or disclose AI-generated content in the EU, what the text requires, and when it applies.
Regulation (EU) 2024/1689, Article 50, with the Commission's Guidelines and Code of Practice.
Which duty applies to me?
Article 50, by topic
Each part on its own page, with the source wording and common questions.
Who holds which duty
Providers build transparency into the system; deployers tell the people exposed to it. Paragraph by paragraph, with the article text.
Article 50 duties → Guidelines and CodeHow the Code of Practice marks content
The layered mark signatories commit to under 50(2), free detection, and the 2 February 2027 interoperability date.
The Code of Practice → ComparisonArticle 50 vs California SB 942
The EU and California marking rules row by row: who, what, labels, detection, dates and penalties. On sb942.com, the sister reference.
Side by side →Dates
Article 50 already applies. Systems on the market before August have until December to add marks.
Regulation enters into forceArt. 113
Final Code of Practice on Transparency of AI-Generated Content publishedEuropean Commission
Article 50 obligations apply. Systems placed on the market from this date must meet 50(2)Art. 113
- Today
Oct 2026 Transition ends for systems already on the market before 2 August 2026, following the 2026 amendment packageartificialintelligenceact.eu
Interoperable detection solution among Code signatoriesIPTC
Sources: Regulation (EU) 2024/1689, Article 113; artificialintelligenceact.eu on the December 2026 grace period; IPTC on the Code's interoperability date.
Penalties
- National market surveillance authorities
What 50(2) requires, and what it does not
Requires
- Machine-readable marks on synthetic audio, image, video and text
- Output detectable as artificially generated or manipulated
- Effective, interoperable, robust and reliable, as far as feasible
- Applied by the provider of the generating system
Does not require
- A specific marking technologyThe text names none
- A visible label from the providerA deployer duty under 50(4)
- Naming the person behind the system
- Application by the deployerDeployers hold the disclosure duties
Marking and detection technologies
Same fields for each, from each organisation's own materials. Not an endorsement.
SynthID
Google DeepMind
- Type
- Watermarking applied to outputs of Google generative models; detection portal described by Google
- Modalities
- Image, audio, video, text
- Availability
- Applied within Google products; SynthID Text published as open source (2024); text method described in a 2024 Nature paper
Verda
Verda Inc.
- Type
- Imperceptible watermarking with a verification service; company states it writes C2PA manifests alongside the watermark
- Modalities
- Image, audio, video
- Availability
- Commercial; company states verification is free to the public
Adobe Content Authenticity
Adobe Inc., San Jose, California; C2PA co-founder
- Type
- C2PA Content Credentials combined with an invisible watermark and fingerprint that link back to the credential, per company
- Modalities
- Image (JPG, PNG) in the public beta; company states video and audio are planned
- Availability
- Free web app in public beta since April 2025; Chrome extension for inspection
Meta AudioSeal, Video Seal
Meta FAIR
- Type
- Neural watermarking with published models and papers
- Modalities
- Audio (AudioSeal), video (Video Seal), image (Watermark Anything)
- Availability
- Open source under the licences stated in each repository
Updates
- Article 50 obligations begin to apply.Source: Regulation (EU) 2024/1689, Art. 113
- Final Code of Practice on Transparency of AI-Generated Content published by the Commission.Source: European Commission, Digital Strategy
- Regulation (EU) 2024/1689 enters into force. Article 50 applies from 2 August 2026.Source: Regulation (EU) 2024/1689, Art. 113
Glossary
- Provider
- Develops an AI system, or has one developed, and places it on the market or puts it into service under its own name or trademark (Article 3(3)).
- Deployer
- Uses an AI system under its authority, other than for personal non-professional activity (Article 3(4)).
- Deep fake
- AI-generated or manipulated image, audio or video that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful (Article 3(60)).
- Machine-readable marking
- A mark software can read without human interpretation. Not further defined; the Code of Practice describes metadata and watermarks as forms of it.
- Detectability
- Content can be identified as artificially generated or manipulated. The Code of Practice couples it with free detection.
About this site
Operated by Verda Inc., which offers one of the technologies profiled; its profile uses the same format as every other. No rankings, comparisons, reviews or recommendations. Corrections from any profiled organisation are applied on receipt of a public source.