EU AI Act Article 50: what applies from 2 August 2026, and what got a grace period
Rasmus Kjaer Damgaard, Co-founder ·
Šis straipsnis parengtas su dirbtinio intelekto pagalba ir peržiūrėtas žmogaus. Už turinį atsakingas Rasmus Kjaer Damgaard.
TL;DR: Since 2 August 2026, four transparency duties from Article 50 of the EU AI Act apply to companies that build or use certain AI systems: tell people when they interact with an AI, mark AI-generated content in a machine-readable way, disclose emotion recognition and biometric categorisation, and label deepfakes and AI-written text on matters of public interest. The deadlines did not move, with one narrow exception. Below: each duty, who bears it, the exemptions that actually hold, and the dates, with citations. You can run the whole applicability analysis in a few minutes with our free Article 50 checker.
The four duties, in one table
| Duty | Who bears it | Applies from |
|---|---|---|
| Disclose AI interaction (Art. 50(1)) | Provider | 2 Aug 2026 |
| Machine-readable marking of AI output (Art. 50(2)) | Provider | 2 Aug 2026, or 2 Dec 2026 for systems on the market before 2 Aug 2026 |
| Emotion recognition / biometric categorisation notice (Art. 50(3)) | Deployer | 2 Aug 2026 |
| Deepfake and public-interest text disclosure (Art. 50(4)) | Deployer | 2 Aug 2026 |
The provider/deployer split matters more than most summaries admit. A "provider" is not only whoever trained the model: offering a third-party model under your own name or trademark makes you the provider in the Act's sense (Art. 3(3)). A company that white-labels a model behind its own branded chat widget carries the provider duties in rows one and two, not just the deployer duties.
The exemptions that actually hold
The Commission's final Guidelines, adopted on 20 July 2026, tightened and clarified the exemptions. The ones that survive scrutiny:
- Obviousness (50(1)): no disclosure needed when the AI interaction is obvious to a "reasonably well-informed, observant and circumspect" person. When in doubt, disclose.
- Standard editing (50(2)): grammar, spellchecking and, per the final Guidelines, machine translation are exempt from marking. Summaries and substantive rewrites are not.
- The business/industrial carve-out (50(2)) is narrower than most companies assume. Three cumulative conditions: the output is strictly technical, it only reaches a limited pre-defined group of professionals inside the organisation, and it is not intended for external sharing and is protected against foreseeable misuse. "We only use it internally" is not, on its own, sufficient.
- Editorial responsibility (50(4)): AI-written public-interest text needs no label when it has undergone human substantive review and a person or organisation holds editorial responsibility. Spellchecking-level review does not count.
- Artistic works (50(4)): evidently artistic, creative or satirical deepfakes get a reduced duty, not an exemption: disclose in a way that does not spoil the work.
The dates, precisely
The Digital Omnibus on AI (in force since July 2026) moved the high-risk Annex III regime to 2 December 2027, and that is where the "the deadline moved" confusion comes from. Article 50 stayed on 2 August 2026, enforcement included. The one transitional rule: generative systems already on the market before 2 August 2026 have until 2 December 2026 for the machine-readable marking duty only.
Penalties reach EUR 15,000,000 or 3% of worldwide annual turnover, whichever is higher; for SMEs the cap reverses to the lower of the two (Art. 99(6)). In Denmark, supervision is sector-based and the authority for Article 50 is not yet settled; the duties and fine levels follow the Regulation itself.
What to do this month
- Inventory where AI touches people or produces content in your company.
- Run the Article 50 checker: it applies the role branch and the cumulative exemption tests above and returns an evidence record with citations per finding.
- Put the result into a policy your employees can actually follow. Our AI policy generator assembles one from structured answers, with every section citing the obligation behind it.
Sources
- Regulation (EU) 2024/1689, Article 50
- Commission Guidelines on transparency obligations (adopted 20 July 2026)
- Commission FAQ on Article 50
- Digitaliseringsstyrelsen on the transparency obligations
Dažniausiai užduodami klausimai
Did the August 2026 deadline for Article 50 move?
No. Article 50 applies from 2 August 2026. The Digital Omnibus deferred the high-risk Annex III regime to December 2027, but the transparency duties stayed on schedule. Only one sub-duty moved: machine-readable marking of AI output, and only for systems already on the market before 2 August 2026, which have until 2 December 2026.
Does every chatbot have to say it is an AI?
As a rule yes, unless the AI interaction is obvious to a reasonably well-informed, observant and circumspect person. The disclosure must come at the latest at first interaction, and it is the provider's duty.
We only use AI internally in a business context. Are we exempt from marking?
Not automatically. The business/industrial carve-out requires three cumulative conditions: strictly technical output, a limited pre-defined group of professionals inside the organisation, and no intended external sharing plus safeguards against misuse. Internal use alone is not enough.
What are the penalties for an Article 50 violation?
Up to EUR 15,000,000 or 3% of total worldwide annual turnover, whichever is higher. For SMEs the rule reverses: the cap is the lower of the two amounts (Article 99(6)).