EU AI Act Transparency Requirements: Article 50 Explained
Article 50 of the EU AI Act requires chatbots to identify as AI, deepfake and synthetic content to be labelled, and emotion-recognition systems to notify the people they analyse. These obligations apply from August 2, 2026 — here is exactly what compliance looks like for each requirement.
Article 50 at a Glance
Applies from August 2, 2026. Fine for non-compliance: up to €15M or 3% of global turnover (Article 99(4)).
Why the EU AI Act Focuses on Transparency
The transparency chapter (Chapter IV) addresses a specific harm: people being manipulated, misled, or making significant decisions based on AI output without knowing that AI is involved. A person who knows they are talking to a chatbot makes different choices — about how much to trust the response, whether to escalate to a human, how much personal information to share — than one who believes they are speaking to a person.
Similarly, AI-generated synthetic media (deepfakes) can damage reputations, spread disinformation, and erode public trust unless audiences can identify it. And emotion-detection systems affect how individuals are assessed and treated without their knowledge.
Transparency obligations are among the least burdensome in the Act — they require a disclosure, not a full compliance programme. For most SMBs, implementing them takes hours, not months.
The Four Transparency Obligations
Chatbot and Conversational AI Disclosure
Applies to: Providers and deployers of AI systems that interact directly with natural persons
What it requires
Any AI system designed to hold conversations with people — chatbots, virtual assistants, AI support agents, AI sales tools — must clearly inform the person they are interacting with an AI system, at the start of the interaction.
Exception: No disclosure is needed if it would be obvious to a "reasonably well-informed" person that they are speaking to an AI — for example, a clearly robotic interface that no user would mistake for a human.
Examples
In practice
Add a single clear line to your AI interface: "I'm an AI assistant — how can I help you?" or "You're chatting with an AI." This is the simplest transparency obligation in the entire Act — one sentence at the start of an interaction, implemented once.
Emotion Recognition and Biometric Categorisation Disclosure
Applies to: Providers of AI systems that detect or infer emotions, or categorise people based on biometric data
What it requires
If your AI system detects, infers, or classifies emotions or uses biometric data to categorise individuals by protected characteristics, you must inform the natural persons being processed that such a system is operating.
Exception: Disclosure is not required if the system is used for AI safety or testing, where it is impossible to communicate with the person (for example, an automated quality assurance system in a factory).
Examples
In practice
If you use biometric analysis that infers characteristics (not prohibited outright), a written notice in your privacy policy and at point of data collection — "This system uses AI to analyse [emotion/biometric characteristic]" — satisfies the obligation. However, first confirm the system is not in a prohibited category under Article 5.
Deepfake and Synthetic Media Labelling
Applies to: Deployers of AI that generates or manipulates image, audio, or video content depicting real people, real-sounding voices, or real environments that did not originally exist as shown
What it requires
AI-generated or AI-manipulated audio, image, or video content — including voice cloning, face-swapped video, or generated imagery depicting real people — must be labelled as AI-generated or AI-manipulated in a way that is clearly perceivable.
Exception: Satire, parody, and creative or artistic expression are exempt, provided the AI-generated nature is clear from context or appropriate disclosure. A fully human narrator describing a fictional scene does not trigger this obligation.
Examples
In practice
For AI-generated video: add a visible on-screen label such as "AI-generated" or "AI-assisted." For audio: include a verbal or written disclosure. For images: add a caption or watermark. The EU Commission is expected to publish technical standards; current best practice is a clear, conspicuous label visible before consumption of the content.
AI-Generated Text on Matters of Public Interest
Applies to: Deployers publishing AI-generated text intended to inform the public on matters of public interest — including news, political commentary, public health, and civic information
What it requires
Text that is AI-generated (not merely AI-assisted with meaningful human editing) and published to inform the public on matters of public interest must be disclosed as AI-generated.
Exception: No disclosure required if there has been substantial human review, editing, and editorial responsibility — meaning a human editor has genuinely considered and owned the content, not just proofread it. Also exempt where the AI nature is evident from context.
Examples
In practice
Most SMBs publishing marketing or product content are not in scope here — this obligation primarily targets media and public-interest publishers. If you do publish public-interest content and use AI substantially, add a short disclosure: "Some content on this site is AI-generated" or a per-article AI label.
Who Is Responsible: Provider or Deployer?
Article 50 obligations sit on both providers (those who build and place AI systems on the market) and deployers (those who use AI systems in their operations or products). Understanding which you are — or whether you are both — determines your exact obligations.
You are a Provider if…
- You built the chatbot, AI content tool, or emotion-detection system
- You placed it on the market or made it available to other businesses
- You are the SaaS vendor supplying a chatbot product to customers
Provider obligation: ensure your system has the technical capability to support disclosure (e.g. an opening AI disclosure message).
You are a Deployer if…
- You use a third-party AI chatbot, content tool, or media generator
- You embed someone else's AI system in your product or website
- You use AI tools to generate content you then publish
Deployer obligation: ensure disclosure actually occurs for your users. If the vendor's tool does not support it, you must add it yourself or find a different tool.
Many SMBs are deployers, not providers. If you use Intercom, Drift, Zendesk AI, or any similar chatbot tool on your website, you are a deployer. Your vendor must make disclosure technically possible; you are responsible for making sure it is switched on.
What Does "Clearly Informing" Mean in Practice?
The Act uses the phrase "in a clear and distinguishable manner" but does not prescribe exact wording or format. The European Commission is expected to issue technical standards and guidance notes — in the meantime, these are the principles that will guide enforcement:
Visible before or at the start
Disclosure must occur before or at the very beginning of the interaction — not buried in a privacy policy or terms of service page.
Plain language
"I'm an AI assistant" is sufficient. Legal jargon or a reference to "automated decision-making" alone is not.
Not undermined by the rest of the interaction
If you disclose the system is AI at the start but the AI subsequently claims to be human or has a human name without any AI indication, that contradicts the disclosure.
Accessible to the person receiving it
For multilingual products, consider whether disclosure in one language is accessible to users interacting in others.
Practical minimum for a chat widget: Configure your chatbot's opening message to include a clear AI disclosure — for example: "Hi there — I'm an AI assistant for [Company Name]. I can help with [common queries]. How can I help you today?" The cost of this change is approximately 5 minutes of configuration.
When Do Transparency Obligations Come Into Force?
All organisations using AI must ensure staff have sufficient AI literacy. Not a transparency obligation, but often bundled with transparency training.
General-purpose AI model providers (those releasing LLMs) have transparency obligations under Chapter V. These are distinct from the Article 50 user-facing obligations.
Chatbot disclosure, deepfake labelling, emotion recognition notification, and AI-generated public text disclosure all become enforceable. This is the primary deadline for most SMBs.
Article 50 Compliance Checklist
Audit all AI systems that interact directly with customers or users
List every chatbot, virtual assistant, AI support agent, and any AI that users interact with as if it were a person.
Add AI disclosure to all customer-facing chatbots
Configure the opening message of each to clearly state it is an AI. This is the highest-volume task for most SMBs.
Review any AI-generated video, audio, or images you publish
If you produce AI-generated synthetic content depicting people or environments, add a label ("AI-generated") before August 2, 2026.
Check your emotion-recognition or biometric tool use
If any tool detects emotions, analyse faces, or categorises people — and is not prohibited outright — add a disclosure at the point of capture.
Confirm your chatbot vendor supports disclosure
Contact your chatbot or AI tool vendor and confirm the system is configured to make AI disclosure. Don't assume it is on by default.
Review AI-generated text published for public audiences
If you publish news, public health information, or civic content using AI without substantial human editing, add an AI-generated content disclosure.
FAQ
Do the Article 50 transparency obligations apply before August 2, 2026?
Article 50 follows the general 24-month transition timeline and becomes enforceable on August 2, 2026. The prohibited practices (Article 5) and AI literacy (Article 4) provisions have been in force since February 2025, but the full transparency chapter applies from August 2026. That said, implementing disclosures now has no downside — it costs almost nothing and demonstrates good faith.
Does every AI tool on my website need a disclosure?
No. The obligation targets AI systems 'intended to interact directly with natural persons' (Article 50(1)) and AI generating content that could mislead. A recommendation algorithm that surfaces content behind the scenes, an internal spam filter, or a translation tool used only by your staff does not trigger the chatbot disclosure. The test is whether a user is directly interacting with the AI as if it were a person.
What counts as "clearly informing" someone they're talking to an AI?
The Act requires the disclosure to be 'clear and distinguishable' but does not mandate specific wording. In practice: a visible notice at the start of the interaction (not buried in a privacy policy), in plain language, before the conversation begins. For a chat widget, this typically means an opening message from the AI itself stating it is an AI assistant. A small icon or fine-print footnote alone is unlikely to satisfy the obligation.
We use a third-party chatbot tool (e.g. Intercom, Zendesk, Drift with AI). Is the vendor responsible, not us?
Both parties have obligations. The chatbot vendor is the 'provider' and must make the AI disclosure feature available to you. You are the 'deployer' and are responsible for ensuring disclosure actually happens when users interact with the chatbot on your platform. You cannot outsource compliance by pointing to the vendor — if the disclosure does not appear for your users, you are the party exposed.
What is the fine for not disclosing AI in a chatbot?
Failure to comply with Article 50 transparency obligations is a Tier 2 violation under Article 99(4), carrying fines of up to €15 million or 3% of global annual turnover. For an SMB with €2M revenue, that is up to €60,000 — proportionate but significant. In practice, first-offence enforcement is likely to require remediation before fines, but the obligation is real from August 2, 2026.
Does Article 50 apply to AI-generated images or only video and audio?
Article 50(3) explicitly covers "image, audio or video content" — all three formats. A static AI-generated image depicting a real person in a false context requires labelling. An AI-generated banner image for a product (showing no real person) does not. The core test is whether the content depicts a real person or real environment in a way that could mislead about its authenticity.
Does Your AI System Need More Than Transparency?
Transparency obligations are just one layer. Use our free classifier to find out whether any of your AI systems are high-risk under Annex III — with a plain-English list of what you'd need to do if so.
Start Free Risk AssessmentRelated Guides
EU AI Act Prohibited Practices: Article 5 Already in Effect
Emotion recognition at work is banned outright — not just subject to disclosure. The 8 Article 5 prohibitions explained.
ChatGPT, Copilot & LLMs: Your Obligations
Provider vs deployer distinction, deployer obligations in 5 steps, GPAI model provider requirements.
EU AI Act Fines and Penalties: Up to €15M for Transparency Breaches
How the three-tier fine structure works and how authorities decide the actual amount.
SMB Compliance Checklist for August 2026
Step-by-step: from prohibited-practices check through to AI literacy. The complete action list.