News · Published 2 September 2026 · AI Act · Vastavus
The AI Act’s disclosure rules took effect. The high-risk deadline did not.
This post was written by AI from public sources. The sources are listed at the end.
Article 50 transparency duties have applied since 2 August 2026. The Annex III high-risk deadline moved to December 2027. Most companies are caught by the first one and spent the year preparing for the second.
Two things happened to the AI Act this summer and many companies heard only one of them. The Digital Omnibus pushed the Annex III high-risk obligations from 2 August 2026 to 2 December 2027, because national authorities and the harmonised technical standards were not ready. That delay got reported widely and plenty of people read it as a general reprieve.
It was not. The transparency duties in Article 50 started to apply on 2 August 2026 and national authorities can enforce them now. Those duties do not depend on a system being high-risk. They attach to ordinary things: a chatbot on your website, a generated product photo, a synthetic voice on your support line.
Who the rules reach
Article 50 splits duties between providers and deployers. If you build the system, you make sure a reasonably observant person can tell they are talking to software, and you mark generated text, audio, images and video so a machine can detect it. If you only use someone else’s system, you still carry duties of your own: disclose deepfakes, tell people when AI assesses their emotions or biometric categories, and label AI-generated text you publish to inform the public on a matter of public interest.
A second date matters for anything already shipped. Systems that were on the market before 2 August 2026 have until 2 December 2026 to meet the machine-readable marking requirement.
The exemptions are narrower than they sound
- Disclosure drops away when AI involvement is obvious in context, which is a judgement call worth writing down rather than assuming
- Assistive editing such as grammar correction falls outside the marking duty, but a tool that substantially rewrites your text does not
- Art and satire get lighter treatment, disclosed in a way that does not spoil the piece
- AI-generated text escapes the public-interest label only when a human reviewed it substantively and someone holds editorial responsibility
Penalties reach 15 million euros or 3 percent of worldwide turnover. The rules follow the user rather than the company address, so a provider outside the EU serving EU users is in scope.
Read it this way: the deadline that slipped covers systems most Estonian SMEs do not build. The deadline that held covers the chatbot and the generated marketing images many of them already run.
Key takeaways
- Article 50 transparency duties have applied since 2 August 2026 and are enforceable now
- The high-risk Annex III deadline moved to 2 December 2027, which covers a different set of systems
- Anything shipped before August has until 2 December 2026 to carry machine-readable marking
Sources
Related training
Custom AI Training Program for Your Company
We find where AI pays off in your company, then build the training, rollout and tools around it.
Book this training