In short
- Article 50 has applied since 2 August 2026. It splits across 4 substantive paragraphs: two bind providers, two bind deployers.
- Article 50(2) requires synthetic audio, image, video and text to be marked in a machine-readable format. Generative systems already on the market before 2 August 2026 have until 2 December 2026.
- That extension covers the marking duty only. The deployer duties in Article 50(3) and 50(4) had no transitional period and applied from day one.
- Disclosure must reach the person at the latest at the time of the first interaction, in a clear and distinguishable manner, and must meet accessibility requirements.
- Penalties for transparency breaches run to EUR 15 million or 3% of worldwide annual turnover, whichever is higher.
What do AI assistants say about your organisation?
Model answers are becoming a channel your communications team does not control. EntityRise audits how assistants describe your company and whose name they give instead. From $19.
Article 50 is the part of the AI Act that reaches organisations which were confident the Act was not about them.
It does not depend on high-risk classification. It depends on whether a person is talking to your system, or reading something it produced.
01Who does each paragraph of Article 50 bind?
The four substantive duties do not all fall on the same party, and confusing them is the most common error.
| Paragraph | Binds | Duty |
|---|---|---|
| Article 50(1) | Provider | Inform people they are interacting with an AI system |
| Article 50(2) | Provider | Mark synthetic output in a machine-readable format |
| Article 50(3) | Deployer | Inform people exposed to emotion recognition or biometric categorisation |
| Article 50(4) | Deployer | Disclose deep fakes, and AI generated text on public interest matters |
A company that buys a generation model and publishes its output is a deployer for paragraph 4 and not a provider for paragraph 2. A company that fine tunes and ships that model to customers is both, for different content, which is the provider question in another guise.
02What does machine-readable marking actually require?
More than a caption, and the text says so.
Article 50(2) requires providers of systems generating synthetic audio, image, video or text to mark the output in a machine-readable format and make it detectable as artificially generated or manipulated. The standard set in the text is that solutions be effective, interoperable, robust and reliable, as far as this is technically feasible.
Interoperable is the demanding word. A proprietary marker that only your own detector can read is arguably not interoperable, which is why the practical answers converge on provenance metadata and watermarking schemes with published specifications.
There are carve outs. Assistive editing tools and systems that do not substantially alter the input data fall outside paragraph 2, as do certain law enforcement uses.
03Which systems got until 2 December 2026, and which did not?
Only generative systems that were already on the market, and only for marking.
| Duty | Applies from | Transitional relief |
|---|---|---|
| Article 50(1) interaction notice | 2 Aug 2026 | None |
| Article 50(2) marking, new systems | 2 Aug 2026 | None |
| Article 50(2) marking, systems on market before 2 Aug 2026 | 2 Dec 2026 | Roughly 4 months |
| Article 50(3) and 50(4) deployer duties | 2 Aug 2026 | None |
The asymmetry matters in procurement. A deployer cannot rely on a vendor’s transitional period, because the deployer duties never had one. If you are disclosing deepfakes under paragraph 4, that duty has been live since August whatever your supplier is doing about watermarks.
04When does the disclosure have to appear?
At the latest at the time of the first interaction or exposure.
Article 50(5) sets the timing and the manner together: the information must be provided in a clear and distinguishable way, no later than first contact, and it must conform to applicable accessibility requirements.
That rules out the two designs teams reach for first. A notice in the terms of service is not at the time of first interaction. A grey line of small text under a chat window is arguably not clear and distinguishable, and is certainly not accessible if it fails contrast requirements.
Article 50(6) closes the obvious escape route: none of this displaces the Chapter III obligations for high-risk systems, and it is without prejudice to other transparency duties in Union or national law.
05What does non-compliance cost?
Up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher.
That tier sits below the ceiling for prohibited practices, including the one arriving on 2 December 2026, and above the tier for supplying incorrect information to authorities. It is enough that transparency should not be filed under minor administrative obligations, which is where it tends to sit on a risk register drawn up around high-risk classification.
The practical exposure is not a single large fine. It is that transparency breaches are visible from outside. Anyone can open your product and see whether the chatbot says what it is, and anyone can check whether your generated images carry provenance metadata. Unlike most of the Act, this part does not require an inspection to detect.
06Frequently asked questions
Do you have to tell users they are talking to a chatbot?
Yes, unless it is obvious. Article 50(1) requires providers to design systems that interact directly with people so that those people are informed they are interacting with an AI system, except where this is obvious to a reasonably well informed person.
What does machine-readable marking mean under Article 50(2)?
It means the marking has to be detectable by a machine, not only visible to a reader. The text requires solutions that are effective, interoperable, robust and reliable as far as is technically feasible, which points at watermarking and provenance metadata rather than a visible caption.
Does a visible label on a deepfake satisfy Article 50?
It satisfies the deployer duty in Article 50(4), which is about disclosure to people. It does not satisfy the provider duty in Article 50(2), which is about machine-readable marking of the output. The two duties bind different parties and both can apply to the same content.
Is journalism exempt from the AI generated text disclosure?
Partly. Article 50(4) exempts AI generated text published on matters of public interest where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication.
07References and method
- Regulation (EU) 2024/1689, Article 50, paragraphs 1 to 7. Primary source for who is bound, what must be disclosed or marked, and the exceptions.
- European Commission, Shaping Europe's digital future, quick facts and FAQ pages on transparency obligations, consulted 12 September 2026 for the 2 August 2026 application date.
- AI Omnibus provisional agreement of May 2026, as reported by law firm analyses consulted on 12 September 2026, for the transitional period to 2 December 2026 applying to Article 50(2) marking for generative systems already on the market.
- Regulation (EU) 2024/1689, Article 99, for the penalty ceiling of EUR 15 million or 3% of worldwide annual turnover.