In Brief: As of August 2, 2026, the transparency requirements of Article 50 of the EU AI Act apply across the European Union. Users must understand when they are interacting with AI; providers of generative systems must ensure machine-readable labeling of synthetic content; and companies publishing deepfakes and certain public-interest texts must disclose AI use. The postponement for high-risk systems did not otmena these rules.
This article is intended for product owners, CTOs, marketers, editors, and compliance teams whose services are available to people in the EU. We focus specifically on transparency, labeling, and the nearest deadlines; we do not cover high-risk classification or sector-specific requirements in detail. This is an informational overview, not legal advice.
Contents
- What happened on August 2, 2026
- Who is covered by Article 50 of the EU AI Act
- The Four Transparency Requirements
- What applies now and what was delayed
- Do you need to label text written with AI?
- What businesses should do in 48 hours
- Fines and enforcement
- Frequently Asked Questions
- Conclusion
What happened on August 2, 2026
On August 2, 2026, the transparency rules of Article 50 of the European AI Act took effect. Around the same time, enforcement began at the national level and by EU authorities for the applicable provisions. The official AI Act timeline separately identifies this date for Article 50, oversight of general-purpose AI models, prohibitions, and AI literacy.
The key business takeaway is the EU AI Act was not postponed in full. In June, the Council of the EU moved the specific requirements for autonomous high-risk systems to December 2, 2027, and for high-risk AI embedded in regulated products to August 2, 2028. But the transparency requirements for chatbots, AI agents, and synthetic content are already in force. This is confirmed by the European Commission’s guidance on Article 50 , updated on August 5, 2026.
Practical meaning: if your website uses an AI chat, your ads include a realistic synthetic video, or your company launches a generative service for the EU market, you need to check transparency compliance now—not in 2027.
Who is covered by Article 50 of the EU AI Act
First, determine the company’s role. In one product, an organization can be both a provider of its own AI system and a deployer of a third-party model.
| Role | What the company does | Typical example | Primary area of responsibility |
|---|---|---|---|
| Provider | develops an AI system or places it on the market under its own name | a SaaS platform with its own AI assistant; a white-label generator | interface notice and technical output labeling |
| Deployer | uses an AI system professionally under its responsibility | a retailer connects a third-party chatbot; an agency publishes AI creative | notifying people and visible labeling of specific materials |
| Both roles | creates a product and uses it in its own operations | a company developed an AI avatar and deployed it in support | both sets of obligations |
The official guidelines explain the boundaries of the roles and give examples of what falls within and outside the scope of Article 50. For providers of general-purpose models, a separate set of rules applies: technical documentation, information for downstream developers, a copyright policy, and a public description of training data. As of August 2, 2026, the European Commission can enforce these requirements, including fines, as noted in the FAQ for GPAI providers.
If a company simply calls the OpenAI, Anthropic, or another model API, that does not usually make it a provider of the base model. But its own branding, white-label delivery, deep modification, and the specific distribution architecture can change the role—this requires a review of the contract and the product.
The Four Transparency Requirements
For implementation, it is useful to use the AI dawn model “4 layers of transparency”. This is not a European Commission term, but a working framework that translates Article 50 into product, content, and compliance tasks.
1. Interface: disclose interaction with AI
The provider must design the system so that a person is informed when they are directly interacting with AI. An exception is possible when this is obvious to a reasonably attentive user given the context. In practice, it is safer not to rely on “obviousness” if the bot has an employee name and avatar.
The notice must appear no later than the first interaction and be clear, distinct from the rest of the interface, and accessible to people with disabilities. For a corporate AI agent, wording like this works:
“You are chatting with the company’s AI assistant. It may make mistakes. To reach a specialist, click ‘Call an operator.’”
For a voice bot, it is better to say the message before collecting data or taking action: “Hello, I’m a virtual AI assistant...”
2. Content: clearly disclose deepfakes and certain texts
The deployer must disclose that realistic images, audio, or video that constitute a deepfake were created or altered with AI. For clearly artistic, satirical, and fictional works, the disclosure can be presented in a way that does not interfere with the audience’s experience, but it should not be omitted entirely.
A separate rule applies to text published to inform the public on matters of public interest. If such text is generated or substantially modified by AI and has not undergone human review or editorial oversight with a responsible person, it must be labeled as AI-generated. The exact wording is set out in Article 50 in the AI Act Service Desk.
3. File: add a machine-readable marker
The provider of a generative AI system must ensure that synthetic text, images, audio, and video are labeled in a machine-readable format so that their origin can be detected. The solution must be effective, compatible, robust, and reliable to the extent technically possible.
This is not the same as a visible “created by AI” label under an image. A visible label helps people; metadata, a provenance signal, or another machine-readable mark helps verification systems. The EU Code of Practice separates measures for providers and deployers exactly this way.
For generative systems already placed on the market before August 2, 2026, the transition period for technical labeling runs until December 2, 2026. It is not a general extension for all Article 50 requirements.
4. Evidence: keep records of who ensured transparency and how
Article 50 sets out the obligation, but during an audit companies will need evidence of compliance. The minimum package includes an AI system register, screenshots of notices, versions of disclosure text, an owner list, contractual information from suppliers, metadata review results, and a change log.
This continues standard AI governance practice: assign an owner, limit the system’s actions, and keep an auditable trail. We covered similar organizational principles in the article “Implementing AI in Business Processes: Where to Start”.
What applies now and what was delayed
| Date | What changes | Status as of August 5, 2026 |
|---|---|---|
| August 2, 2026 | Article 50: chatbots, deepfakes, certain public texts, emotion recognition; start of enforcement of applicable rules | in effect |
| December 2, 2026 | end of the transitional period for machine-readable marking for some generative systems that were on the market before August 2 | nearest deadline |
| December 2, 2027 | rules for standalone high-risk systems from Annex III | postponed |
| August 2, 2028 | rules for high-risk AI embedded in regulated products under Annex I | postponed |
The delay of the high-risk deadlines is set out in the EU Council release of June 29, 2026. It also sets a new deadline of December 2, 2026, for the transitional period for technical marking of older generative systems.
Do you need to label text written with AI
Not every text that used AI automatically requires a visible label under Article 50. The deployer obligation applies to AI-generated or AI-manipulated text that informs the public on matters of public interest. It does not apply if the material has undergone human review or editorial control and a natural or legal person bears editorial responsibility.
Use the following matrix as a first-pass triage, not as a substitute for legal review:
| Scenario | Likely action under Article 50 | Why |
|---|---|---|
| AI drafted a news story; an editor checked the facts, rewrote the piece, and the publisher is responsible for publication | document editorial review; a visible AI label may not be required | the human review + editorial responsibility exception applies |
| AI automatically publishes news without review | add a clear AI-generated label | the text informs the public and does not go through editorial review |
| AI wrote a product description | check the context and other rules; Article 50(4) on public-interest text may not apply | commercial copy is not always a matter of public interest |
| An ad shows a realistic nonexistent expert | label as AI-generated/manipulated; check whether it qualifies as a deepfake | the viewer may mistake the synthetic image for a real person |
| Customer support chat uses an LLM | tell the user about the AI before or at the first interaction | direct interaction between a person and an AI system |
A weak point in many processes is not the editing itself, but the lack of evidence. Record who checks the facts, what counts as a material change, where the pre- and post-review versions are stored, and who is responsible for publication.
What to do in 48 hours
Step 1. Build an AI touchpoint inventory
Document every point where an EU user may encounter AI: website, app, support, voice bot, HR tool, ad creatives, avatars, image generator, and automated editorial feed. For each point, list the model, vendor, owner, and audience.
Step 2. Define the role and obligation
Tag each system with two labels: provider/deployer and the requirement type — interaction notice, machine-readable mark, deepfake label, public-interest text label, or biometric/emotion notice. If the role is unclear, review the vendor contract and the description of whose name the system is marketed under.
Step 3. Fix the first screen
Make sure the AI notice is visible before the first substantive exchange, not buried in the privacy policy. Add a clear path to a human where an AI error could affect an order, payment, access, or complaint. Practical safeguards for conversational systems are collected in the article “8 Rules for Protecting AI Agents”.
Step 4. Separate label and metadata
The content workflow should include separate checks:
- Does a person see a clear label where one is required?
- Did the machine-readable marker survive export, resizing, editing, and upload to the CMS?
- Can the team prove file provenance and the edit history?
Step 5. Establish editorial responsibility
For texts on matters of public interest, require mandatory human review: fact-checking, sources, headlines, images, and data conflicts. Assign a responsible editor or legal entity and keep an approval log.
Step 6. Request evidence from vendors
Ask the supplier to explain how it complies with Article 50: when it shows the notice, what machine-readable marking it uses, what happens on export, and which product versions are covered by the transitional period. Signing the voluntary Code of Practice makes it easier to demonstrate compliance, but the code itself does not replace the binding law.
Penalties and enforcement
Article 50 is enforced by national market surveillance authorities, the AI Office for systems under its oversight, and the European Data Protection Supervisor for EU institutions. According to the European Commission Quick Facts, the maximum fine for Article 50 violations can reach €15 million or 3% of global annual turnover of the company. For small and medium-sized enterprises, the lower of the percentage and fixed caps applies; regulators also assess severity, duration, consequences, and remediation measures.
A maximum does not mean an automatic fine for every incorrect label. But lacking an inventory, an owner, and clear disclosure weakens the company’s position: it cannot show that it systematically managed the risk and corrected violations.
Frequently asked questions
Does the EU AI Act apply to companies outside the European Union?
Yes, in some scenarios — if the AI system is placed on the EU market, is used in the EU, or its output affects people in the Union. Geographic scope depends on the role, product, and supply chain, so companies from Russia and the CIS with European users need a separate applicability analysis.
Do you need to say “created by AI” on every image?
There is no universal rule to label everything. The provider is responsible for machine-readable marking of synthetic output, while the deployer is responsible for visible disclosure of deepfake content and other listed scenarios. Ordinary technical processing that does not change the meaning of the source data may fall under an exception.
Do you need to label an article if AI was used only for the draft?
For text on matters of public interest, visible labeling may not be required when there is genuine human review or editorial control and a responsible publisher has been designated. Keep proof of review; a simple click on “approve” without fact-checking is a weak position.
What should you write in a chatbot window?
Keep it short and before the first dialogue: “You are chatting with an AI assistant. It may make mistakes. To contact a human, click…”. The text must be distinguishable, accessible, and consistent with the system’s actual behavior.
Was the EU AI Act postponed until 2027?
No. Specific rules for high-risk systems were postponed: to December 2, 2027, for Annex III and to August 2, 2028, for AI inside regulated products. Article 50 on transparency applies from August 2, 2026.
Do you need to sign the Code of Practice?
The code is voluntary. Signatories may use the measures set out there to demonstrate compliance with labeling and disclosure requirements. Non-signatories may use alternative adequate measures, but they will need to justify them to the regulator.
Conclusion
As of August 2, 2026, the EU AI Act has turned transparency from a “best practice” into an enforceable requirement. The immediate business priority is not to rewrite the entire AI stack, but to do four things: show people where AI is being used; properly label content that requires it; preserve technical provenance markers; and document accountability and reviews.
Start with an inventory of AI touchpoints and the chatbot’s first screen. Then review the export chain for synthetic files and the editorial workflow. These are the actions that close the main gap between a polished policy and verifiable execution.