Article 50 of the EU AI Act establishes binding transparency obligations for providers and deployers of certain AI systems, and these requirements applied as scheduled on 2 August 2026 – unaffected by the Digital Omnibus amendments that postponed other chapters of the Act. Businesses that interact with natural persons through AI systems, generate synthetic content, or deploy emotion-recognition and biometric categorisation tools must now meet disclosure and labelling standards enforceable by the AI Office. Non-compliance carries exposure to the penalty regime that also entered force on 2 August 2026, making early and structured preparation critical for any operator active in the EU market.
What the Digital Omnibus Changed – and What It Did Not
Before examining Article 50 itself, it is important to understand the regulatory landscape as it stands following the Digital Omnibus on AI (Regulation (EU) 2026/1744). Political agreement on that amending regulation was reached on 7 May 2026, the European Parliament voted it through in plenary on 16 June 2026, it was published in the Official Journal on 24 July 2026, and it entered into force on 27 July 2026.
The Omnibus restructured two major sets of obligations. High-risk AI system requirements linked to Annex III (systems deployed in sensitive areas such as employment, education, and critical infrastructure) were pushed back from 2 August 2026 to 2 December 2027. Obligations tied to Annex I product-embedded AI systems – those covered by existing product safety legislation – were extended further to 2 August 2028.
Critically, however, the Omnibus did not touch Article 50. Transparency obligations, AI Office enforcement powers, and the penalty regime all applied on 2 August 2026 as originally scheduled. Businesses that assumed the Omnibus bought them breathing room across the board were mistaken, and that misreading is one of the most common errors we observe among foreign clients entering the EU market.
For a detailed breakdown of all Omnibus deadline changes and their interaction with sector-specific rules, see our AI Act Omnibus Deadlines Alert.
The Scope of Article 50: Who Is Caught
Article 50 applies to four distinct categories of obligation, each attaching to a different type of AI system or output:
- AI systems interacting with natural persons: Providers of AI systems designed to interact directly with individuals – chatbots, virtual assistants, and automated customer service tools – must ensure those individuals are informed they are communicating with an AI, unless it is obvious from context.
- Emotion recognition and biometric categorisation: Deployers using such systems must notify the natural persons exposed to them at the time of operation.
- Deep fakes and AI-generated synthetic content: Providers and deployers must label audio, video, text, and image content that is artificially generated or manipulated using machine-readable marking and, where technically feasible, visible disclosure.
- General-purpose AI text generation: Providers of general-purpose AI systems that generate text published for the purpose of informing the public must ensure that output is machine-readable marked as artificially generated, unless the content has undergone substantial human editing.
Sector-specific guidance from bodies such as ESMA and the EBA is expected to clarify how Article 50 interacts with existing financial services disclosure frameworks, though that guidance was still pending at the time of writing. Operators in regulated industries should monitor developments closely rather than assume that existing sectoral disclosures satisfy Article 50 independently.
Step-by-Step Compliance Procedure
Step 1 – System Inventory and Role Determination
The starting point is a structured inventory of all AI systems deployed or made available in the EU. For each system, the business must determine whether it acts as a provider (placing the system on the market or putting it into service), a deployer (using a third-party system in a professional context), or both. This distinction governs which Article 50 obligations fall on whom, and failures here often produce duplicated or contradictory disclosure notices.
Step 2 – Classify Each System Against the Four Article 50 Categories
Each inventoried system should be mapped against the four obligation categories described above. Where a system produces synthetic content and interacts with natural persons, both sets of obligations apply concurrently. Document the classification rationale; the AI Office may request it during supervisory inquiries.
Step 3 – Design and Implement Disclosure Mechanisms
Disclosure mechanisms must be clear, timely, and technically appropriate. For systems interacting with persons, the disclosure must occur at the start of the interaction. For synthetic content, machine-readable labelling – such as the C2PA standard or equivalent technical metadata – must be embedded at the point of generation or at the latest before distribution. Visible disclosures where technically feasible are required in addition to machine-readable markers.
Contracts with technology vendors and content platforms should be reviewed to confirm who is responsible for embedding and preserving labels. Labels stripped or degraded by downstream processing create compliance gaps that regulators may attribute to the original provider.
Step 4 – Employment and Internal Governance
Article 50 compliance is not solely a technical matter. Employees who operate, review, or publish AI-generated content need clear internal policies on when disclosure is required and how to apply it consistently. Businesses operating in multiple EU member states should note that national supervisory arrangements for Article 50 may vary; the AI Office retains direct enforcement competence for general-purpose AI, while member-state authorities take the lead for other systems. Potential interaction with employment law frameworks – particularly where AI monitors workers – warrants separate analysis; our Employment Law practice can assist with that dimension.
Step 5 – Documentation and Ongoing Monitoring
Maintain records that demonstrate compliance decisions: system inventory, classification rationale, disclosure templates, vendor contractual provisions, and staff training logs. The European Commission retains delegated power to issue implementing acts refining technical labelling standards, and EUR-Lex should be monitored regularly for updates. The AI Office may also issue guidance and supervisory notices that affect best practice without formal legislative amendment.
Realistic Timelines and Cost Drivers
For organisations that began preparation before 2 August 2026, the principal work streams – inventory, classification, and disclosure design – typically required several months of cross-functional effort involving legal, technical, and communications teams. For those now conducting a retrospective review, the timeline to achieve a defensible compliance position will depend on the complexity of the system portfolio and the robustness of existing vendor contracts.
Cost drivers include technical integration of labelling frameworks into content generation pipelines, legal review of third-party provider agreements, training programmes for relevant staff, and, where gaps are identified, remediation of existing deployments. Organisations with complex or high-volume general-purpose AI use should allocate additional resource for ongoing monitoring as secondary legislation and guidance develop.
Common Errors by Foreign Clients
- Assuming the Omnibus postponed all obligations: As noted, Article 50 applied on 2 August 2026 without delay. Relying on the Annex III or Annex I postponements as a general reprieve is an error with live enforcement consequences.
- Conflating deployer and provider obligations: A business that uses a third-party chatbot as a deployer still carries direct Article 50 disclosure obligations toward users, even if the provider has its own parallel obligations.
- Treating machine-readable labelling as optional: Where visible disclosure is described as applying "where technically feasible", the machine-readable component is not similarly qualified – it is mandatory.
- Failing to address multi-format output: An AI system that produces both video and text requires labelling in each format; a single disclosure in one medium does not satisfy the obligation for the other.
- Overlooking contractual allocation: Provider agreements that are silent on who embeds and preserves synthetic content labels create ambiguity that neither party can afford in an enforcement context.
- Assuming existing financial services disclosures are sufficient: Sector-specific disclosure obligations under ESMA or EBA frameworks serve different purposes and do not substitute for Article 50 compliance, which remains independently enforceable.
If you are assessing your current Article 50 position or reviewing vendor agreements against the new requirements, our team can assist. Please write to us at info@ferrazwhitmore.com or visit our contacts page to arrange a discussion.
Decision Checklist
The following checklist summarises the key questions a business should be able to answer affirmatively to demonstrate a structured compliance approach:
- Have all AI systems interacting with EU persons been inventoried and classified under the four Article 50 categories?
- Is the business's role – provider, deployer, or both – determined and documented for each system?
- Are disclosure notices triggered at the correct point in time for each system type (start of interaction; at or before distribution for synthetic content)?
- Is machine-readable labelling embedded in all AI-generated content prior to distribution, regardless of whether visible disclosure is technically feasible?
- Do vendor contracts clearly allocate responsibility for labelling, preservation, and downstream processing?
- Are internal staff policies and training in place for employees who operate or publish AI-generated output?
- Is there a process for monitoring AI Office guidance and European Commission implementing acts on technical labelling standards?
- Where the business operates across multiple member states, has it assessed any variation in national supervisory arrangements?
Article 50 in Context: Interaction with High-Risk and General-Purpose AI Regimes
Article 50 sits within a broader compliance architecture. When Annex III obligations enter force on 2 December 2027, businesses deploying high-risk AI systems – for example, in recruitment or access to essential services – will face additional requirements that intersect with the transparency baseline already in effect. Early investment in algorithmic transparency infrastructure under Article 50 will reduce the incremental burden of high-risk compliance when that date arrives.
For businesses operating in Spain or other EU member states with active national implementation, our dedicated AI Act compliance practice in Spain provides jurisdiction-specific guidance on how Article 50 interacts with local supervisory arrangements and sector rules.
General-purpose AI model providers face a distinct but related set of obligations under other chapters of the AI Act, and compliance with those obligations does not substitute for the Article 50 framework applicable at the system-deployment level.
How Ferraz & Whitmore Can Assist
Our AI Act team advises providers, deployers, and importers across the EU on building compliance programmes that are proportionate, defensible, and designed to accommodate the regulatory evolution that will continue through 2027 and 2028. We work with legal, technical, and operational teams to assess system portfolios, review and negotiate vendor agreements, and develop internal governance frameworks.
To discuss your organisation's Article 50 position or broader AI Act compliance programme, please contact us at info@ferrazwhitmore.com or through our contacts page. We are glad to assist with an initial assessment of your current exposure and the steps most likely to reduce it.
Disclaimer: This publication is provided for informational purposes only and does not constitute legal advice. The information herein should not be relied upon as a substitute for professional legal counsel tailored to your specific circumstances. Ferraz & Whitmore assumes no liability for actions taken or not taken based on the contents of this material. For advice regarding your particular situation, please contact info@ferrazwhitmore.com.