Which parts of EU AI Act Article 50 apply to a chatbot that also generates images and has been on sale in the EU since 2025? The notice telling people they are talking to an AI has applied since August 2, 2026. Machine-readable marking of the images it generates can wait until December 2, 2026, because the system was on the market before August 2.

The rest of this piece shows how to tell which of the four Article 50 duties your system triggers and when each one applies.

Checked November 12, 2026 against the Commission’s Article 50 guidelines, the Code of Practice page and Regulation (EU) 2026/1744. This is an explainer, not legal advice.

The four duties in one table

Article 50 of Regulation (EU) 2024/1689 is short, but it holds four separate obligations. The Commission’s guidelines (C(2026) 5054 final, dated July 20, 2026) lay them out side by side, and they can apply cumulatively to one system.

Paragraph Who must act What they must do Main exceptions
50(1) Provider of a system that interacts directly with people Design it so people are told they are interacting with an AI system The AI origin is obvious; systems authorized by law for criminal investigations (unless open to the public for reporting crimes)
50(2) Provider of a system that generates or manipulates audio, image, video or text Mark outputs in a machine-readable format so they are detectable as AI-generated or manipulated Assistive function for standard editing; no substantial change to the input or its meaning; law enforcement use
50(3) Deployer of emotion recognition or biometric categorization Inform the people exposed to it Law enforcement uses permitted by law
50(4) Deployer who publishes deepfakes, or AI-generated text informing the public on matters of public interest Disclose that the content is artificially generated or manipulated Text under human review or editorial control with someone holding editorial responsibility; law enforcement use authorized by law; a lighter regime for evidently artistic, satirical or fictional work

Paragraph 50(5) then sets the delivery rule for all four: the information must be given “in a clear and distinguishable manner at the latest at the time of the first interaction or exposure” and meet accessibility requirements, as the guidelines summarize it.

The Commission’s announcement of July 20, 2026 puts the split the same way: providers handle the “you are talking to an AI” notice and the machine-readable marks, deployers handle the visible disclosure for deepfakes, public-interest text and emotion recognition or biometric categorization.

Which deadline applies to you

Article 50 has applied since August 2, 2026. According to the guidelines, every in-scope system on the EU market had to comply from that date, whatever its launch date, with one exception added by the Digital Omnibus on AI, Regulation (EU) 2026/1744 of July 8, 2026.

That exception is narrow. It covers only the marking and detection duty in 50(2), and only for generative systems placed on the market or put into service before August 2, 2026. Those providers have until December 2, 2026. A system that is partly interactive and partly generative gets the transition only for the marking part: its “you are talking to an AI” notice had to be in place on August 2.

Your situation Notice under 50(1) Machine-readable marking under 50(2) Deployer labels under 50(3) and 50(4)
Generative system on the EU market before Aug 2, 2026 Since Aug 2, 2026 By Dec 2, 2026 Since Aug 2, 2026
Generative system launched on or after Aug 2, 2026 From launch From launch From first use
Interactive system with no generation (for example, a support chatbot that only retrieves answers) Since Aug 2, 2026 Not applicable Depends on use

Counted from today, November 12, 2026, the December 2 deadline is 20 days away. For the chatbot in the opening example, that is the window left to ship machine-readable marking on the images it generates. The notice to users was due three months ago.

What happens to content made before August 2

The guidelines say outputs and deepfakes generated before August 2, 2026 do not need to be marked or labeled retroactively, and the same holds for public-interest text published before that date. The catch is publication timing: text generated before August 2 but published on or after it does need a label.

The editorial exception for text

The second part of 50(4) is the duty most publishers ask about. A deployer who publishes AI-generated or manipulated text “with the purpose of informing the public on matters of public interest” must disclose it. The guidelines describe the exception as two cumulative conditions: the text has gone through human review or editorial control, and a natural or legal person holds editorial responsibility for the publication. Both have to be true. A human glance without anyone answerable for the publication does not meet the test.

The duty is also limited to text on matters of public interest. Deepfakes in image, audio or video are covered by the first part of 50(4) and have their own, lighter regime when they are part of an evidently artistic, creative, satirical or fictional work.

How big the fines can be: a worked calculation

The guidelines state that breaches of Article 50 can be fined up to EUR 15 million or, for an undertaking, up to 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. For small and medium-sized enterprises, including start-ups, the cap is whichever of the two is lower. EU institutions face up to EUR 750,000.

Company Prior-year worldwide turnover 3% of turnover Rule Maximum fine
Large provider EUR 2 billion EUR 60 million Higher of 15M or 3% EUR 60 million
Mid-size provider (not an SME) EUR 300 million EUR 9 million Higher of 15M or 3% EUR 15 million
SME EUR 20 million EUR 600,000 Lower of 15M or 3% EUR 600,000

These are ceilings, not expected amounts. The guidelines add that fines must be effective, proportionate and dissuasive and take account of smaller companies’ economic viability.

Where the Code of Practice fits

The Code of Practice on Transparency of AI-generated Content covers 50(2), 50(4) and 50(5). It has a provider section on marking and detection and a deployer section on labeling deepfakes and AI-generated text. Its closing plenary and publication were on June 10, 2026, and the Commission says about 190 companies and organizations had signed by the end of July 2026.

Signing is voluntary, the obligations are not. The Commission and the AI Board have confirmed the code as an adequate way to show compliance, so signatories can point to its measures. A provider that complies some other way has to show its own measures are adequate, and that is assessed by each market surveillance authority. The Commission has also published a set of icons that deployers may use to label AI-generated content.

A short checklist for the next 20 days

  1. List every product surface that talks to users, generates media or text, or reads emotions or biometric categories. Map each one to 50(1) to 50(4) using the first table.
  2. For each generative system, record whether it was on the EU market before August 2, 2026. Only those systems have until December 2, and only for marking.
  3. Confirm the interaction notice appears at or before the first interaction, not in terms of service.
  4. For published text on matters of public interest, write down who reviews it and who holds editorial responsibility. If either is missing, label the text.
  5. Decide whether to rely on the Code of Practice or to document an alternative approach.

If your system also retrieves documents to answer users, our guides on RAG data provenance and RAG document access control cover the records that make AI answers traceable.

Reporting & Sources

Checked November 12, 2026: Commission guidelines on the transparency obligations under Article 50, C(2026) 5054 final (July 20, 2026); Commission announcement of the guidelines (July 20, 2026); Code of Practice on Transparency of AI-generated Content; Regulation (EU) 2024/1689, the AI Act; Regulation (EU) 2026/1744, the Digital Omnibus on AI (July 8, 2026). The deadline matrix, day count and fine calculations are ASYA’s arithmetic from those texts; the turnover figures are hypothetical.