# EU AI Act Article 50: AI Labelling Rules and Deadlines

> Article 50 of the EU AI Act has applied since 2 August 2026. Chatbots must tell users they are talking to AI, generative AI providers must mark output in a machine-readable way, and businesses that publish deepfakes or AI-written public-interest text must label them. Systems already on the market before August get until 2 December 2026 for the marking duty.

- URL: https://fake-ai.com/ai-text/eu-ai-act-article-50/
- Language: en
- Updated: 2026-10-19
- Publisher: Fake AI (OnlineBuilders LLC, https://onlinebuilders.biz)

The EU AI Act entered into force on 1 August 2024, but most of it arrives in stages. The transparency rules in Article 50 are the stage that reaches the most companies, because they don't depend on whether an AI system is "high risk". If you run a customer-service chatbot, sell an image generator or publish AI-made marketing videos to European audiences, they apply to you. They have applied since 2 August 2026.

This summer also brought a late change. Regulation (EU) 2026/1744, the so-called Digital Omnibus on AI, gave one narrow part of Article 50 more time. Everything else stayed on schedule.

## The four duties in Article 50

Article 50 splits obligations between providers, meaning those who develop an AI system and place it on the market, and deployers, meaning businesses and public bodies that use one ([Regulation (EU) 2024/1689](https://eur-lex.europa.eu/eli/reg/2024/1689/oj)).

**Who must do what under Article 50**

| | Applies to | What it requires |
|---|---|---|
| Art. 50(1) Interactive AI | Providers | Design chatbots and voice agents so people know they are dealing with AI, unless obvious |
| Art. 50(2) Synthetic content | Providers | Mark AI-generated audio, image, video and text in a machine-readable, detectable way |
| Art. 50(3) Emotion recognition, biometric categorisation | Deployers | Inform the people exposed to the system |
| Art. 50(4) Deepfakes | Deployers | Disclose that the content was generated or manipulated |
| Art. 50(4) Public-interest text | Deployers | Disclose AI generation, unless a human with editorial responsibility reviewed it |

Article 50(5) adds the how: the information must be clear and distinguishable, given at the latest at the first interaction or exposure, and accessible to people with disabilities.

### What counts as a deepfake

The Act defines a deepfake in Article 3(60) as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic. Fantasy dragons are not deepfakes. Realistic video of your real CEO announcing a product she never presented is.

For evidently artistic, creative, satirical or fictional works, the duty shrinks: you only have to disclose the existence of generated content in a way that doesn't spoil the work. Films can use a credit line instead of a permanent overlay.

### The text rule and its exemption

AI-generated or manipulated text published to inform the public on matters of public interest must be disclosed. The exemption matters for publishers: if a human reviewed the text and someone holds editorial responsibility for the publication, no AI label is required. Newsrooms that use AI for drafts but keep editors in charge are covered by that exemption. Sites that auto-publish unreviewed AI news are not.

## The deadlines, including the Omnibus change

**Article 50 and related dates**

- **1 Aug. 2024:** AI Act enters into force.
- **2 Feb. 2025:** Bans on prohibited AI practices apply.
- **2 Aug. 2025:** Obligations for general-purpose AI models apply.
- **10 June 2026:** Commission publishes the final, voluntary **code of practice** on marking and labelling AI content.
- **8 July 2026:** Digital Omnibus on AI adopted as Regulation (EU) 2026/1744, in force from 27 July.
- **2 Aug. 2026:** **Article 50 applies**: disclosure, marking, deepfake and text labelling.
- **2 Dec. 2026:** Deadline for Art. 50(2) marking for systems placed on the market **before** 2 Aug. 2026. New Omnibus ban on AI that generates non-consensual intimate images of real people applies.

Source: Regulation (EU) 2024/1689; Regulation (EU) 2026/1744; European Commission

The Omnibus inserted a new Article 111(4): providers of systems generating synthetic content that were placed on the market before 2 August 2026 "shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026" ([EUR-Lex](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32026R1744)). The recitals describe it as a four-month transition. It covers only machine-readable marking by providers of existing systems. Chatbot disclosure and deepfake labelling were not postponed, and new systems launched after 2 August 2026 get no grace period.

## Fines

Breaching Article 50 falls under Article 99(4): fines of up to €15 million or, for companies, up to 3 percent of total worldwide annual turnover in the previous financial year, whichever is higher. For SMEs and start-ups the same caps apply, but whichever is lower.

**Example: maximum fine for a company with €800m worldwide turnover**

- Fixed cap: €15,000,000
- 3 % of €800,000,000 turnover: €24,000,000
- **Maximum (higher of the two): €24,000,000**

For an SME the lower of the two figures is the cap. Actual fines depend on the circumstances of the breach.

National market surveillance authorities enforce the rules, and each member state decides how its authorities are organised.

## How providers can meet the marking duty

The Act asks for marking that is machine-readable and detectable, as far as technically feasible, taking into account the type of content, the costs involved and the state of the art. It does not name a technology. Providers usually combine methods:

- signed metadata such as C2PA Content Credentials for images and video, and for audio
- invisible watermarks hidden in pixels, in audio or in word choice, such as Google's SynthID
- fingerprinting or logging that allows a later lookup

Each method has weak spots. Metadata gets stripped by platforms, and text watermarks fade when text is rewritten or translated. Our guides to [Content Credentials](https://fake-ai.com/images-and-video/content-credentials-c2pa/) and the [AI text watermark](https://fake-ai.com/ai-text/ai-text-watermark/) explain the trade-offs. The voluntary code of practice published in June 2026 sets out practical steps for providers in one section and for deployers' labelling in another.

## What a good label looks like

The Act asks for disclosure that is clear and distinguishable, no later than the first exposure. It doesn't prescribe wording. Alongside the code of practice, the Commission published EU icons for labelling AI-generated content, which gives companies a common visual option.

Sensible patterns, depending on the format:

- Chatbot: a first message such as "You're chatting with an AI assistant. You can ask for a human at any time."
- Voice agent: a spoken line at the start of the call, before any data is collected.
- Image or video: a visible label on or directly beside the content, not only in a caption that gets cut off on mobile.
- Satirical or artistic work: a credit line or note in the description.

Avoid labels hidden in terms and conditions or revealed only on hover. They fail the "at the latest at first exposure" test, and they invite exactly the complaints the rule is meant to prevent.

## A checklist for deployers

Most companies are deployers, not providers. They use ChatGPT or Midjourney, maybe a voice bot, but don't build them.

**Article 50 checklist for businesses using AI**

- [ ] List every customer-facing chatbot, voice agent and AI avatar, and check that each tells users it is AI at the start
- [ ] Find out where your marketing or comms team publishes realistic AI images, audio or video of real people, products or places
- [ ] Add a visible disclosure to deepfakes, or a credit line for clearly artistic or satirical work
- [ ] For AI-assisted articles on public-interest topics, document human review and who holds editorial responsibility
- [ ] Keep metadata intact: don't strip C2PA data from generated files in your CMS
- [ ] Ask your AI vendors how they meet Art. 50(2) and by which date
- [ ] Write the rules into your AI policy and train the people who publish

## What this doesn't cover

Article 50 is about transparency, not about whether a use is allowed. Prohibited practices sit in Article 5, high-risk systems in Chapter III, and copyright questions outside the AI Act. It also doesn't create a duty to detect AI content, so platforms and schools that use AI detectors do so under other rules. Why those detectors are a weak basis for decisions is covered in our analysis of [whether AI detectors are accurate](https://fake-ai.com/ai-text/are-ai-detectors-accurate/).

If you build or operate generative AI products for EU users, the marking and logging work is engineering, not just policy. Studios such as [OnlineBuilders](https://onlinebuilders.biz/solutions/ai-systems) build AI systems with provenance and review built in. For everyone else, the next step is the inventory in the checklist above: one afternoon, one spreadsheet, every AI touchpoint your customers see. Ask a legal adviser to check the edge cases. Our [AI text section](https://fake-ai.com/ai-text/) has the background on detection and watermarks.

## FAQ

### Does Article 50 apply to companies outside the EU?

Yes, if they place AI systems on the EU market or if the output of their systems is used in the EU. A US company offering a chatbot to European customers has to meet the disclosure duty.

### Do I have to label every image I edited with AI?

No. Marking duties don't apply to assistive editing that doesn't substantially alter the input, and the labelling duty for deployers concerns deepfakes: content resembling real people, objects, places or events that would falsely appear authentic.

### Does Article 50 apply to private individuals?

Not for purely personal, non-professional use. The AI Act excludes deployers who are natural persons acting in that capacity. Providers of the tools they use still have to mark output.

### Is the EU code of practice on AI labelling mandatory?

No. The code published in June 2026 is voluntary. The Commission and the AI Board have confirmed it as an adequate tool, so signatories can rely on its measures to show they comply. Companies that don't sign have to show by other means that their measures are adequate.

Photo: Romaine, CC0 (https://commons.wikimedia.org/wiki/File:Brussels-Berlaymont_building_(1).jpg)