AI Act Article 50

EU AI Act Article 50:
chatbots, deepfakes and the 2 December 2026 deadline

Article 50 of the EU AI Act has applied since 2 August 2026: people must be told they are dealing with an AI chatbot or agent from the first interaction, deepfakes must be disclosed, and AI-generated text on matters of public interest must be labelled unless a human reviews it. By 2 December 2026, generative AI systems already on the market must also mark their output in a machine-readable way.

Checked against the European Commission's texts as of 17 November 2026: who does what, what changes on 2 December, and what non-compliance costs.

By Cesare D'Adamo//10 min read

What does Article 50 of the EU AI Act require?

Article 50 of Regulation (EU) 2024/1689 sets four transparency obligations, applicable since 2 August 2026: two for whoever provides the AI system and two for whoever uses it professionally. The Digital Omnibus (Regulation (EU) 2026/1744) did not delay them; it only gave generative systems already on the market until 2 December 2026 to apply machine-readable marking.

SituationObligationWhoFrom
Chatbots, voice assistants and AI agents (para. 1)Tell people they are interacting with AI, unless it is obviousProvider2 August 2026
Systems generating text, images, audio or video (para. 2)Machine-readable marking of the outputProvider2 August 2026; 2 December 2026 for systems already on the market
Emotion recognition and biometric categorisation (para. 3)Inform the people exposedDeployer2 August 2026
Deepfakes and text on matters of public interest (para. 4)Disclose that content is AI-generated; not needed for text with human review and an accountable editorDeployer2 August 2026

Paragraph 5 sets the form: clear and distinguishable information, given at the latest at the first interaction or exposure, and accessible to people with disabilities. Non-EU providers are covered too when their system's output is used in the EU.

  1. 10 June 2026Final Code of Practice on AI-generated content
  2. 20 July 2026Commission guidelines on Article 50
  3. 27 July 2026Digital Omnibus enters into force
  4. 2 August 2026Article 50 applies
  5. 2 December 2026Grace period for marking by existing systems ends
Article 50 dates. Sources: European Commission and AI Act Service Desk (Articles 50 and 111), checked on 17 November 2026.

Provider or deployer: who has to comply?

It depends on your role, not on company size. A 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. A deployer uses an AI system under its authority in a professional activity; purely personal use is out of scope.

In its Article 50 FAQ, updated on 24 July 2026, the Commission confirms that employees using AI on the company's behalf are not separate deployers: the company is. For a chatbot, what matters is how it was built.

Vendor's productBuilt for you, your brand
ExampleSaaS customer service widget on your websiteChatbot built by an agency or your team on a model API
Who is the providerThe vendorYou, if you put it into service under your name
AI disclosure noticeDesigned by the vendor; you check it showsDesigned by you
Marking generated outputThe vendor's jobYour job; the model's marking is taken into account
Deepfakes and published textYour job as deployerYour job as deployer
Who does what under Article 50 and the Commission FAQ of 24 July 2026. The same company can be provider and deployer at once.

The same applies to AI agents: the Commission names chatbots, AI agents and avatars as examples of systems that interact directly with people. If an agent emails customers or answers the phone for you, Article 50 applies just as it does to a website chat.

How should a chatbot or AI agent disclose that it is AI?

The notice must come at the start of the first interaction, and be clear, distinguishable and accessible. According to the Commission FAQ, the duty applies when four conditions are met together:

  • it is an AI system;
  • it is designed for a genuine two-way exchange, not just data collection or automated replies;
  • it communicates directly, with no human intermediary;
  • the counterpart is a natural person.

The exception for cases where AI is obvious is judged from the point of view of an average person who is reasonably well informed, observant and circumspect, and must be read narrowly. Machine-to-machine exchanges and background systems are out of scope. Do not rely on a bot persona or a robot icon: say it in words.

  • Website or in-app chat: an opening message such as "You are chatting with a virtual assistant powered by artificial intelligence."
  • Phone and voice assistants: one sentence at the start of the call, before any question.
  • Emails written by an agent: a line in the message saying it was written and sent by an AI system.

A handover to a human is not required by Article 50, but it reduces complaints. Under the GDPR, the chatbot's privacy notice should also say what data it collects and for how long.

What changes on 2 December 2026: marking AI-generated content

On 2 December 2026 the grace period for marking ends: providers of systems generating text, images, audio or video that were placed on the market before 2 August 2026 must mark output in a machine-readable format, detectable as AI-generated (Article 111(4), added by the Digital Omnibus). Systems launched after 2 August are already covered.

Marking must be effective, interoperable, robust and reliable, as far as technically feasible. The Commission FAQ excludes short strings, source code, purely machine-to-machine output and assistive standard editing. Content created before 2 August 2026 does not need retroactive marking.

If you only use third-party tools, 2 December is your vendors' deadline. If you offer a generative feature under your own name (a product description generator, an assistant that drafts customer replies), you are the provider. The FAQ says techniques applied by the underlying model's provider will be taken into account, so check what your model already marks.

VendorWhat it says it doesSource
Anthropic (Claude)Invisible text watermark, extended to more models on 3 and 16 November 2026; C2PA credentials on generated files; watermark detection in private previewAPI release notes and help center, read 17 November 2026
OpenAIC2PA metadata and SynthID watermarks on images from ChatGPT, Codex and the API; public verification toolOpenAI post, 11 June 2026
GoogleSynthID watermarking and the C2PA standard; work with Apple, ElevenLabs, Kakao, NVIDIA and OpenAI on interoperable watermarksGoogle blog, 24 July 2026

Anthropic's release notes say text from some models might not carry the watermark when a request sets non-default values for temperature, top_p or top_k. If your chatbot changes them, have it tested before 2 December.

The reference tool is the Code of Practice on Transparency of AI-generated Content, published on 10 June 2026 and assessed as adequate by the Commission and the AI Board. It is voluntary: non-signatories must show compliance by other adequate means. About 190 organisations had signed by the end of July 2026 (Commission, 31 July 2026), including Anthropic, Google, Microsoft, Mistral and OpenAI as providers, and companies such as Getty Images, Lufthansa and Iberdrola as deployers.

When do deepfakes and AI-generated text need a label?

When the content could deceive. For the Commission, a deepfake is AI-generated or manipulated image, audio or video that meets three criteria together: high resemblance to its subject, a subject that exists or could plausibly exist, and content that could falsely appear authentic. Context and audience expectations matter.

A stylised product visual in a campaign is probably not a deepfake; a video of a spokesperson who looks like a real person is. The label must appear at the latest on first exposure, be perceivable without special tools, and the provider's invisible marking is not enough. For evidently artistic, satirical or fictional works, a disclosure that does not spoil the work is enough.

For text, the duty covers text published to inform the public on matters of public interest, such as politics, justice, public health, the environment, or economic and scientific developments under public debate. No label is needed if the text has had substantive human review by someone with professional expertise, or editorial control with a person accountable for it. A spell check does not count.

You can label with the three optional EU icons, downloadable from the Commission's transparency page. Human review and verifiable sources also help you get cited by ChatGPT and AI Overviews.

What are the fines and who enforces Article 50?

Breaching Article 50 can cost up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher (Article 99(4)). For SMEs, start-ups and, since the Digital Omnibus, small mid-caps, the lower of the two applies. See what the EU AI Act requires of companies for the other fine tiers.

National market surveillance authorities enforce Article 50. The AI Office steps in only in specific cases, such as systems built on a general-purpose AI model by the same provider, or systems inside very large online platforms and search engines under the Digital Services Act.

National laws add detail. Italy's Legislative Decree 179/2026, for example, names its cybersecurity agency ACN as market surveillance authority, repeats the EUR 15 million or 3% ceiling for Article 50, and allows non-financial measures, such as an order to end the breach, for minor infringements. Italy has also made spreading harmful AI deepfakes without consent a crime, punishable by one to five years in prison (Law 132/2025).

Article 50 checklist: what to do before 2 December 2026

Seven steps, and most of them have applied since 2 August 2026.

  1. 01List every chatbot, voice assistant, agent and generative tool in use, including marketing and customer service tools.
  2. 02Assign the role for each: a vendor's product (you are the deployer) or put into service under your name (you are the provider).
  3. 03Check the notice on every channel, from chat to phone, at the first interaction.
  4. 04Get written confirmation from vendors on marking and Code of Practice adherence, and put it in contracts.
  5. 05If you are a provider of a generative system live before 2 August 2026, complete marking by 2 December.
  6. 06Set the content rule: who labels deepfakes, with which icon, and who is the accountable editor for published text.
  7. 07Train and document: brief marketing, support and IT, and keep evidence of what you did and when.

For every other date, see the EU AI Act timeline. If you want a second look at your chatbots and vendor contracts before 2 December, the first consultation on AI transformation is free.

Free consultation

Want to apply this to your company?

In a free first call we look at your situation and I tell you where I would start, with which tools and what results to expect.

Frequently asked questions

Do I have to disclose that my blog posts are written with AI?

Not under the AI Act, if the text goes through substantive human review or editorial control with an accountable person. The duty covers public-interest text published without that control.

Our chatbot comes from a vendor: is compliance their problem?

If it is the vendor's product, the vendor must design the AI notice and marking. If the chatbot was built for you and you put it into service under your name, you are the provider under the AI Act. Either way, check that the notice is visible and the contract covers transparency.

What changes on 2 December 2026 if we only use ChatGPT, Claude or Gemini?

Very little for you: the deadline is for vendors whose generative systems were on the market before 2 August 2026. Your deployer duties already apply: label deepfakes and public-interest text published without human review.

Do AI images in advertising need a label?

Only if they are deepfakes, meaning realistic depictions of people, places or events that exist or could exist and could pass as authentic. For evidently creative works, a disclosure that does not spoil the work is enough.

Does Article 50 apply to companies outside the EU?

Yes. Providers are covered wherever they are established when their system's output is used in the EU. Signatories of the Code of Practice can rely on it wherever they are established or supervised.

Do we have to sign the Code of Practice on AI-generated content?

No, it is voluntary. Signatories get a recognised route to show compliance with marking and labelling; non-signatories must show it by other adequate means. Deployers can sign Section 2 only, which covers deepfakes and text.

What is the fine for a missing chatbot disclosure?

Up to EUR 15 million or 3% of worldwide turnover, whichever is higher; for SMEs and start-ups, whichever is lower. National market surveillance authorities enforce it.

Cesare D'Adamo
About the author

Cesare D'Adamo

Product and AI strategy consultant

15+ years in product, growth and AI, more than 20 AI-powered products built, and executive teams trained on EU AI Act compliance. Works from Berlin with companies across Europe, in English and Italian.

Let's put AI to work

Strategy, products and automation with AI that deliver measurable results, within the rules. It starts with a free call.

Book a free consultation

First call free, no commitment