Do you have an AI assistant on your site, or use AI to write copy and generate images? From 2 August 2026 there is one more thing to do, and it is simple: say so. Article 50 of Regulation (EU) 2024/1689 (the AI Act) requires you to make it transparent when content or a conversation goes through artificial intelligence. You do not need to rewrite your processes, you need to make clear what already happens. This article is a practical guide to what changes for an SME, without conference jargon and without alarmism.
The first thing to know is that the AI Act is a regulation, not a directive: it applies directly in every member state, without an Italian law to transpose it. So the 2 August 2026 date holds as it is, here too.
What the AI Act says, in short
Article 50 is about transparency: the core idea is that a person should be able to know when they are facing the output of an AI, instead of mistaking it for something made by a human. It does not ban using AI, it asks you to disclose it in the cases where it would not otherwise be evident. The full text is public on the AI Act portal.
The four transparency obligations
The article sets out four situations, and separates whoever builds the model (the «provider») from whoever uses it in their own processes (the «deployer», the user). In plain language:
50(1), interactive systems: if you have a chatbot or AI assistant that people talk to, you must tell them they are interacting with an AI, unless this is already entirely obvious from the context.50(2), synthetic content (on the provider): whoever produces AI-generated audio, images, video or text must mark it in a machine-readable format, so it can be detected as AI-generated.50(3), emotion recognition and biometric categorisation: whoever uses these systems must inform the people exposed to them.50(4), deepfakes and texts on matters of public interest (on the deployer/user): manipulated images, audio and video (deepfakes) and texts published on matters of public interest must be labelled as AI-generated or manipulated. This drops if there is human review with editorial responsibility, and is lighter for artistic or satirical works (a disclosure that does not spoil the experience is enough).
What really changes for an SME
This is the part that matters. Most SMEs are deployers, not providers: they use AI tools someone else built, they do not develop their own models. That narrows the field considerably. The obligations that affect a typical SME are two.
The first is 50(1): if there is an AI assistant on the site or in customer service, it must be disclosed as an AI. The second is 50(4): if you publish deepfakes or texts on matters of public interest generated by AI, they must be labelled, unless a person reviews them and takes editorial responsibility. The machine-readable marking under 50(2) is mainly the model provider's job, but for an SME there is still an indirect duty: to check that the tools you use actually guarantee it.
As for consequences, Article 99(4) of the AI Act sets fines for breaches of Article 50 of up to EUR 15 million or, for a company, up to 3% of total worldwide annual turnover for the preceding financial year (whichever is higher), in force from 2 August 2026. It is context, not a reason to panic: the practical steps, as we will see, are within reach.
The Code of Practice: concrete (and voluntary) instructions
On 10 June 2026 the European Commission published a Code of Practice on marking and labelling AI-generated content, aimed at providers and deployers of generative AI. It is useful because it turns the text of the rule into operational steps: how to mark, how to label. For labelling it points to the EU Icons, common European symbols, and it is open for signatures; it will then be complemented by the Commission's guidelines.
One clarification that avoids confusion: the Code is voluntary. Signing and following it helps demonstrate compliance, but it does not replace the law. The obligation comes from Article 50; the Code is the convenient road to meeting it well.
The 2 August deadline (and the only exception)
It is worth clearing up a point that causes confusion. A simplification package for the AI Act is under way, the so-called Digital Omnibus: the European Parliament approved it on 16 June 2026, but it is still awaiting formal adoption by the EU Council, so it is not yet definitively consolidated law. It should be read with caution.
Above all: the Omnibus does not move the general deadline of 2 August 2026. There is a single technical exception, and it concerns providers: for generative AI systems already on the market before 2 August 2026, compliance with the machine-readable marking only (50.2) shifts to 2 December 2026. For the obligations that affect a typical SME, namely disclosing the chatbot and labelling content, the date stays 2 August.
What to do now, in practice
The good news is that these are orderly actions, not a building site. A reasonable path to being ready:
- Map the AI systems you use: chatbots and assistants on the site or in customer service, tools that generate text, images, audio or video. Knowing where AI touches your content is the first step.
- Check the notices in the interfaces: where there is an AI assistant, make sure it discloses that it is an AI clearly and visibly, not buried at the bottom of a page.
- Formalise human review with traceability: for texts on matters of public interest, a person who reviews them and takes editorial responsibility removes the labelling obligation; but you need to be able to prove it, so keep a record of who reviews what.
- Review your supplier contracts: who marks what? Put it in writing that the tools you use guarantee the machine-readable marking required by 50(2), so the burden does not land on you by surprise.
- Choose solutions that mark and label by design: prefer tools that handle transparency and marking natively, rather than a patch added afterwards.
How we see it
Transparency is not a box to tick before a deadline: it is a design posture. When AI is placed where it is useful and disclosed from the start, Article 50 takes care of itself almost on its own. That is why the assistant on this site discloses that it is an AI, and in Verso Flow Flow AI is described as based on Claude by Anthropic: not to comply at the last minute, but because saying how things are made is how we build them.
Compliance is also a chance to put things in order upstream: to understand where AI is actually useful (we wrote about it in «AI in your SME: how to tell if it pays off»), and to make systems talk to each other, so the traceability of human review is not a separate sheet of paper. If you want native transparency in custom software, or you want the tools you already use connected through data integration, tell us how you work today: we start from your case, not from the rule in the abstract.