On 2 August 2026 Article 50 of the AI Act became directly applicable law, and on 31 July the Commission confirmed it would start enforcing it from that date. In July we wrote about a deadline to prepare for; now it is the rule that surveillance authorities measure your product against. The good part: for the average SME it is still half a day of work. The awkward part: that half day has to happen now, not in September.
What went live
The transparency duties hit anyone providing or deploying four categories of system, regardless of whether they are high-risk.
- Chatbots: the user must know they are talking to an AI no later than the first interaction. A disclosure hidden in the FAQ or the terms is not enough, it has to reach the first message.
- Synthetic content: providers of generative systems must mark output in a machine-readable format. Systems already on the market before 2 August have until 2 December 2026 to meet the marking requirement, but that is the only extension.
- Deepfakes and public-interest text: the labelling duty sits with whoever publishes them.
- Emotion recognition and biometric categorisation: the people exposed must be informed.
The Commission's guidelines, adopted in late July, are not binding but are exactly the text authorities will read when they assess a case. Worth keeping to hand.
What getting it wrong costs, and who enforces
Non-compliance with Article 50 sits in the penalty band up to 15 million euro or 3% of annual worldwide turnover, whichever is higher, capped at the lower of the two amounts for SMEs. It is not the 35 million band reserved for prohibited practices, but it is still wildly out of proportion to the cost of fixing it. In Italy the enforcer is ACN, under law 132/2025: the same agency many SMEs already know from NIS2, not an abstract office in Brussels.
What we fixed, in practice
- One line in the chatbot's first message: "You are talking to an automated assistant". Not in the footer, not on the terms page, in the first conversation turn.
- A note on generated content: where a model produces published text or images, machine-readable marking plus a visible label where the rule requires one.
- A clause in SaaS contracts: who owns the Article 50(2) marking obligation, the model provider or you integrating it. It is the question that sets client and vendor arguing in September.
- An inventory of AI features: without knowing which functions call a model, you cannot know which obligations apply to you.
Verdict
Anyone who read "transparency" in July and put it off is today non-compliant on a feature they probably already run in production. This is not a project, it is closing four points in one morning and writing down who owns the marking obligation. The expensive part is not the remote fine: it is discovering in a tender audit that your site's chatbot discloses nothing and that the vendor assumed you were handling it.