Article 50 of the EU AI Act applies to ordinary AI products, not just high-risk ones: if your system talks to people, generates media or text, or labels people's emotions, it probably has a transparency duty. The obligations started to apply on 2 August 2026, and the European Commission's AI Act Service Desk gives providers whose generative systems were already on the market before that date until 2 December 2026 to meet the machine-readable marking duty in Article 50(2). This guide sets out who the article applies to, what each paragraph requires, how AI agents and machine identities are affected, how to implement and evidence compliance, and how Article 50 maps to other frameworks. It is for product, engineering, legal and compliance teams in organisations that build or use AI systems in the EU market. The wider Act is covered in the EU AI Act Compliance Guide.
Key takeaways
- Four duties, two roles. Providers must disclose AI interaction (50(1)) and mark synthetic outputs (50(2)); deployers must inform people exposed to emotion recognition or biometric categorisation (50(3)) and disclose deep fakes and certain public-interest text (50(4)).
- Dates are close and changed once. Article 50 applies from 2 August 2026; the Commission timeline, which it says reflects the Digital Omnibus on AI, sets 2 December 2026 for Article 50(2) for systems already on the market before 2 August 2026.
- Marking must be technical, not cosmetic. Article 50(2) asks for outputs marked in a machine-readable format and detectable as artificially generated, using solutions that are effective, interoperable, robust and reliable as far as technically feasible.
- Disclosure must arrive in time. Article 50(5) requires information that is clear and distinguishable, at the latest at the first interaction or exposure, and accessible.
- Fines are significant. Breaching Article 50 can attract fines of up to EUR 15 million or 3 percent of worldwide annual turnover, whichever is higher, with the lower of the two for SMEs (Article 99).
What it is and who it applies to
Article 50 sits in Chapter IV of Regulation (EU) 2024/1689, the AI Act, and creates transparency obligations for certain AI systems regardless of whether they are high-risk. Article 50(6) states that the obligations do not affect the high-risk requirements in Chapter III and are without prejudice to other transparency duties under EU or national law, so a high-risk system can carry both sets.
| Role | Definition in the Act (paraphrased) | Typical Article 50 examples |
|---|---|---|
| Provider (Article 3(3)) | A person or body that 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 | A company shipping a customer service chatbot; a model vendor offering an image generator; an organisation building its own voice agent |
| Deployer (Article 3(4)) | A person or body using an AI system under its authority, except for personal non-professional activity | A marketing team publishing AI-generated video; a news site publishing AI-written articles; an employer using a biometric categorisation tool |
| Deep fake (Article 3(60)) | AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic | A synthetic video of a real executive; a cloned voice in an advert |
The same organisation is often both. A bank that builds an AI assistant on a third-party model is the provider of the assistant for 50(1), and when its communications team publishes a synthetic video it is a deployer for 50(4). Which role you hold also determines who must act when you use a vendor's system: the vendor marks outputs, you label deep fakes you publish. See the EU AI Act Provider vs Deployer Obligations Guide.
Structure and key requirements
| Paragraph | Who | Obligation | Main exceptions |
|---|---|---|---|
| 50(1) | Providers | Design AI systems intended to interact directly with people so those people are informed they are interacting with an AI system | Where obvious to a reasonably well-informed, observant and circumspect person in context; systems authorised by law to detect, prevent, investigate or prosecute criminal offences (unless available for the public to report an offence) |
| 50(2) | Providers, including of general-purpose AI systems | Ensure synthetic audio, image, video or text outputs are marked in a machine-readable format and detectable as artificially generated or manipulated; solutions effective, interoperable, robust and reliable as far as technically feasible | Assistive functions for standard editing; systems that do not substantially alter the input data or its semantics; uses authorised by law for criminal offences |
| 50(3) | Deployers of emotion recognition or biometric categorisation systems | Inform the people exposed about the operation of the system and process personal data in line with EU data protection law | Certain uses permitted by law for criminal offences; ancillary biometric categorisation strictly necessary for objective technical reasons |
| 50(4) | Deployers | Disclose that deep fake image, audio or video content has been artificially generated or manipulated; disclose AI-generated or manipulated text published to inform the public on matters of public interest | Evidently artistic, creative, satirical or fictional works need only a disclosure that does not hamper the work; text under human review or editorial control with editorial responsibility; uses authorised by law |
| 50(5) | Providers and deployers | Information must be clear and distinguishable, given at the latest at first interaction or exposure, and meet accessibility requirements | None |
| 50(7) | Commission and AI Office | Encourage codes of practice on detecting, marking and labelling AI-generated content; the Commission may adopt implementing acts | Not an operator duty |
Timeline and status
| Date | Event | Source |
|---|---|---|
| 1 August 2024 | AI Act entered into force | European Commission |
| 19 November 2025 | Commission adopted its AI omnibus proposal (Digital Package on Simplification) | European Commission |
| 5 March 2026 | Second draft of the Code of Practice on marking and labelling AI-generated content published | European Commission |
| 7 May 2026 | Political agreement on the AI omnibus; draft Commission guidelines on Article 50 listed the same day | European Commission |
| 2 August 2026 | Article 50 transparency rules apply | AI Act Service Desk |
| 2 December 2026 | Deadline for Article 50(2) for providers of generative systems placed on the market before 2 August 2026 | AI Act Service Desk |
The Service Desk states that its timeline reflects amendments from the Digital Omnibus on AI. The sources checked for this guide on 27 October 2026 did not confirm the date of formal publication of the amending regulation, and the Code of Practice page still showed the code as in preparation. Check the Commission's pages before relying on these dates.
Identity and NHI requirements
Article 50 has no identity clause, but several obligations are delivered through identity and machine credential controls:
- AI agents that contact people: a voice agent placing calls, an email agent replying to customers or an agent posting in a chat channel is an AI system interacting directly with people. Build the disclosure into the agent's greeting, signature or display name, and give each agent a distinct, registered identity so you can show which agents are in scope and that each discloses. See the Voice AI Agent Security and Fraud Guide.
- Agent-to-agent channels: 50(1) concerns natural persons. Where an agent's output is relayed to a human by another system, decide where disclosure happens and document it.
- Marking keys are machine credentials: if you meet 50(2) with signed provenance such as C2PA Content Credentials, the signing key proves your outputs are yours. Hold it in an HSM or KMS, give the signing service a scoped workload identity and log every signing call. The AI Content Provenance and C2PA Guide covers this.
- API keys for generation services: outputs created through a leaked key still carry your marks, or none if the attacker uses an unmarked path. Protect generation API keys as part of your NHI programme.
- Audit trails: to evidence disclosure and marking, logs need to tie each output or conversation to the system and identity that produced it. See the AI Agent Audit Trails for Regulators Guide.
Implementation steps
- Inventory systems against each paragraph. For every AI system you provide or deploy, record whether it interacts with people (50(1)), generates synthetic content (50(2)), performs emotion recognition or biometric categorisation (50(3)), or produces content you publish as deep fakes or public-interest text (50(4)). Note that emotion recognition in the workplace and education is prohibited outright under Article 5, so 50(3) only covers permitted uses.
- Assign roles. For each system, record whether you are provider, deployer or both, and which supplier contract covers marking.
- Design 50(1) disclosures. Use a clear statement at the start of each interaction (text, voice prompt or interface element) rather than terms of service. Apply the "obvious" exception only with a documented rationale.
- Choose 50(2) marking techniques. Combine machine-readable metadata (for example C2PA manifests) with watermarks or fingerprints for durability, and decide how text outputs are marked. Follow the Code of Practice once final.
- Build detection. Article 50(2) also requires outputs to be detectable; provide a way for others to check (published verification tools, metadata readable by standard validators).
- Write 50(4) labelling rules for communications, marketing and editorial teams: when a label is required, approved wording and placement, and the editorial review route that removes the duty for public-interest text.
- Handle 50(3) notices alongside your GDPR transparency information and data protection impact assessment.
- Pilot and test. Test marking survival through your publishing channels, and test disclosures with accessibility tools, before the deadline that applies to you. Keep the previous release available so you can roll back a broken disclosure flow without removing the marking.
- Prioritise legacy systems. Generative systems on the market before 2 August 2026 have the 2 December 2026 deadline for 50(2); new systems must comply on placement.
Small versus large organisations: an SME that only uses vendor tools mostly has deployer duties (deep fake and public-interest text labels) plus contract checks that its vendors mark outputs. A large provider of generative features needs an engineering programme for marking and detection, version control of disclosure text and monitoring across products.
Evidence and assessment
| Obligation | Evidence to keep | How to test it |
|---|---|---|
| 50(1) interaction disclosure | Screenshots or recordings of each disclosure, design specification, rationale for any "obvious" exception | Scripted test conversations per channel and language; accessibility check |
| 50(2) marking and detection | Technical description of marking method, signing certificate details, sample outputs with validation results, survival test results | Automated sampling of outputs, validation through standard tools, survival after common transformations |
| 50(3) emotion and biometric notices | Notice text, placement, DPIA references | Review at each deployment site |
| 50(4) deep fake and text labels | Labelling policy, approval records, examples of labelled content, editorial review records for exempt text | Periodic review of published content against the policy |
| Code of Practice adherence (if adopted) | Gap assessment against the final code | Annual review |
Metrics: percentage of in-scope systems with a documented disclosure; percentage of generated outputs carrying a valid machine-readable mark (sampled); mark survival rate across publishing channels; number of published deep fakes without a label found in review (target zero); time to fix a disclosure defect.
Mapping to other frameworks
| Framework | Related element | Relationship to Article 50 |
|---|---|---|
| China: Measures for Labeling AI-Generated Synthetic Content (in force 1 September 2025) | Explicit labels and implicit metadata labels; platform detection duties | Similar split between visible labels and machine-readable marks; one technical approach can serve both markets |
| C2PA Content Credentials | Signed manifests with AI digital source type, plus soft bindings | A widely used technical means for 50(2) marking of media |
| GDPR Articles 13 and 14 | Information to data subjects | 50(3) notices sit alongside GDPR transparency |
| ISO/IEC 42001:2023 | AI management system | Provides the governance process to maintain Article 50 controls |
| MITRE ATLAS | AML.T0088 Generate Deepfakes; AML.M0034 Deepfake Detection | Threat-side view of the content Article 50 asks you to label |
| NIST CSF 2.0 | GV.OC-03 legal and regulatory requirements understood and managed | Places Article 50 in your compliance register |
For the global picture, see the Global AI Regulation Map.
Common pitfalls
- Disclosure buried in terms: a line in the terms of service is not information given clearly at the first interaction.
- Visible label only: a caption saying "AI generated" does not meet the machine-readable marking duty in 50(2).
- Metadata stripped in publishing: marks applied at generation are removed by image optimisation or social platform uploads.
- Assuming the vendor covers everything: the model vendor may mark outputs, but your deep fake labelling and your own chatbot's disclosure remain your duties.
- Treating "editing assistance" broadly: the exception covers standard editing and systems that do not substantially alter input; a tool that replaces a face or a voice is not standard editing.
- Agents nobody registered: a team deploys an outbound email agent without the disclosure, because it was never in the AI inventory.
Practitioner checklist
- Inventory AI systems against Article 50(1) to 50(4) and record your role for each.
- Confirm which deadline applies: 2 August 2026, or 2 December 2026 for 50(2) on systems already on the market, and re-check the Commission's pages for changes.
- Implement clear, accessible disclosures at first interaction for every chatbot and agent that talks to people.
- Mark generated media with machine-readable metadata plus a durable watermark or fingerprint.
- Protect marking and signing keys and generation API keys as privileged machine credentials.
- Publish or support a way to detect your marks.
- Issue a deep fake and public-interest text labelling policy to publishing teams.
- Collect evidence for each paragraph and test it on a schedule.
- Check supplier contracts for marking and detection commitments.
- Track the final Code of Practice and Commission guidelines and update controls.
Standards and references
- EU AI Act: Article 50 Transparency obligations for providers and deployers of certain AI systems
- EU AI Act: Article 99 Penalties
- European Commission AI Act Service Desk: Timeline for the implementation of the EU AI Act
- European Commission: AI Act regulatory framework
- European Commission: Code of Practice on marking and labelling of AI-generated content
- C2PA: Content Credentials Technical Specification
- Cyberspace Administration of China: Measures for Labeling AI-Generated Synthetic Content
Related NHI Mgmt Group resources: EU AI Act Compliance Guide · EU AI Act Provider vs Deployer Obligations Guide · AI Content Provenance and C2PA Guide · Deepfake and AI Impersonation Guide. This guide summarises Article 50 of the EU AI Act as checked on 27 October 2026 and is not legal advice.