Article 50 applies now: what actually changed on 2 August 2026

On 2 August 2026 the transparency obligations of Article 50 of the EU AI Act started to apply. On the same day, enforcement began for the general-purpose AI model obligations, the prohibited practices, the transparency duties and AI literacy. The preparation phase is over: this is law in force, not a date to work towards.

That difference matters more than it sounds. Until last week an organisation could say it was preparing. From now on the question is whether the disclosure is there. Below is exactly what changed, which part still has a transitional period, and which date now belongs at the top of the plan.

What started to apply on 2 August 2026?

Article 50 contains four transparency duties, split between providers and deployers. All four have applied since 2 August 2026.

Article 50 is not a high-risk regime. There is no conformity assessment, no Annex IV technical documentation and no registration in an EU database. That does not make it optional, but it does make the task bounded: four concrete behaviours, not a full management system.

Watch the split of roles, because that is where most organisations get it wrong. The duty to make a chatbot recognisable as AI sits with the provider, in the design of the system. The duty to label deepfakes sits with the deployer, at publication. Many organisations are the provider for one system and the deployer for another, so the analysis has to run per system.

What does it mean that enforcement started?

The enforcement structure the AI Act prescribes came into operation on 2 August 2026. National market surveillance authorities can use their powers for the obligations that apply, and for general-purpose AI models the powers of the Commission and the AI Office became active.

For the transparency duties, a breach can lead to a fine through the supervisor of up to 15 million euro or 3 percent of total worldwide annual turnover, whichever is higher. That is the ceiling, not the starting point: supervisors weigh the nature, gravity and duration of the infringement, and whether an organisation identified and remedied the gap itself.

The practical consequence is that demonstrability now counts. A supervisor asking how you meet Article 50 is not asking about your intentions but about your systems: which ones communicate with people, which generate synthetic content, where the disclosure sits, and who established that and when.

What still sits under a transitional period?

One provision has a transitional period, and it is narrower than often assumed. Only Article 50(2), the machine-readable marking of synthetic output, has time until 2 December 2026. That extension applies solely to systems already placed on the market before 2 August 2026.

Everything else applies in full. A generative system placed on the market after 2 August 2026 has to have the marking in order immediately. And the visible disclosure of deepfakes under paragraph 4 does not fall under the transitional period at all: that duty sits with the deployer and has applied since 2 August 2026.

The Commission published its final implementing guidelines on Article 50 on 20 July 2026. Alongside that, a voluntary code of practice on the transparency of AI-generated content has existed since 10 June 2026. Signatories can rely on it to demonstrate compliance with the second, third and fifth paragraphs of Article 50. Signing remains possible, even now that the list of initial signatories has closed.

Which date is next?

Not 2 December 2027, which is what many assume because the high-risk obligations moved in that direction. The next hard date is 2 December 2026, and it brings two things.

First, the transitional period ends for the machine-readable marking under Article 50(2) for systems already on the market before 2 August 2026. Second, the new prohibitions start to apply on the use of AI for child sexual abuse material and non-consensual intimate content, added to the AI Act by Regulation (EU) 2026/1744 and requiring technical safeguards.

After that come 2 August 2027 for operational AI regulatory sandboxes, 2 December 2027 for the core obligations on standalone Annex III systems including the Article 27 FRIA, and 2 August 2028 for AI embedded in regulated products under Annex I.

What do you do this week?

Start with the inventory, not with the wording of the disclosure. Without a view of which systems fall under which paragraph, you write disclosures for systems that do not need them and miss the ones that do.

The full breakdown per paragraph is in our practical guide to Article 50, and the article text itself is in the AI Act Explorer. Organisations that want to approach this in a structured way, from inventory through role determination to record keeping, start with Embed AI. For teams that need to recognise when a disclosure is required, LearnWize offers role-based training with an evidence file.