Since Sunday, 2 August 2026, the transparency obligations in Article 50 of the European regulation on artificial intelligence — the AI Act — apply in earnest. From the same date, the full penalty regime and the enforcement powers for general-purpose models are in operation. We are no longer talking about an approaching deadline: the rules are in force.
If your company has a chatbot on its website, uses AI to generate texts, images or clips for marketing, or publishes materials featuring artificially generated or altered people, this article concerns you directly. You do not have to “do AI” to be in scope — using AI tools in your dealings with the public is enough.
The good news: for most SMEs, complying with Article 50 is a short list, solvable in a few working days. The less good news: many people have mixed up the deadlines and are left with the impression that “the AI Act was postponed”. Something else was postponed. Let us put things in order.
What changed on 2 August 2026
Three things came into force on this date:
- The transparency obligations in Article 50 — detailed below: chatbots that declare themselves, generated content marked, deepfakes labelled.
- The full penalty regime — non-compliance with the regulation can from now on be penalised, not merely noted.
- The enforcement powers for general-purpose models — aimed primarily at the providers of large models, but they consolidate the framework your obligations are part of too.
A detail that often slips by: the obligations do not depend on the company's size, nor on who built the technology. The chatbot taken on subscription from an external vendor remains, in the visitor's eyes, your chatbot — and the disclosure must exist in your concrete implementation, not just in the vendor's documentation. If it is not clear which part falls to you and which to them, ask for the answer in writing; it is, in passing, a good test of the vendor's seriousness too.
Obligation 1: the chatbot must say it is AI
If there is a conversational assistant on the company website, the user must be informed that they are interacting with an artificial intelligence system, not with a human. The regulation accepts one exception — the situation where this is obvious to a reasonably informed person — but “obvious” is slippery ground on which it is not worth building your defence. A clear message at the start of the conversation closes the subject without argument.
In practice, this means:
- the chatbot's first message says explicitly that the user is talking to an AI-based virtual assistant;
- the information is visible in the conversation window, not hidden in a terms-and-conditions page;
- if the bot hands the conversation to a human operator, the moment of the handover is clearly signalled.
If you are only now planning an assistant for your website, it is cheaper to build it correctly from day one — we have written separately about what a company chatbot costs and what the AI Act requires of it.
Obligation 2: AI-generated content gets a machine-readable marking
Texts, images, sounds and clips generated or substantially altered by AI must be marked in a format computer systems can detect automatically — not just a cosmetic mention in the footer. The marking must exist in the file or in its metadata, so that platforms and verification tools can read it without human intervention.
For your company, the useful question is: where does AI-generated content go out? Newsletter, social media posts, product images, descriptions generated in series for the online shop. Serious generation tools already apply such markings; your job is to check that the option is switched on in the tools used by your team or agency and that your publishing flow does not strip the markings along the way — for example by re-saving or recompressing the images.
Obligation 3: deepfakes get a visible label
If you publish materials in which a real person appears saying or doing something generated or manipulated by AI — or realistic images of people, places or events that never existed in that form — you must declare visibly that the material is artificially generated or manipulated. For marketing, the practical rule is simple: if a clip or an image could be taken for an authentic recording, label it. The cost of the label is zero; the cost of its absence, from 2 August onward, no longer is.
What was actually postponed — and what was not
On 16 June 2026, the European Parliament approved the amendments in the “Digital Omnibus” package, which postpone the obligations for high-risk AI systems: the standalone ones (Annex III) from 2 August 2026 to 2 December 2027, and those embedded in products (Annex I) from 2 August 2027 to 2 August 2028. That is where the “everything was postponed” confusion was born.
| Obligation | Status in August 2026 |
|---|---|
| Transparency (Article 50): chatbots, content marking, deepfakes | In force since 2 August 2026 |
| The penalty regime | In force since 2 August 2026 |
| Standalone high-risk systems (Annex III) | Postponed to 2 December 2027 |
| High-risk systems embedded in products (Annex I) | Postponed to 2 August 2028 |
The short conclusion: transparency applies now; high risk, later. The chatbot's obligation to declare itself AI was not postponed by a single day.
The checklist for your company
- Inventory the AI contact points: chatbot on the website, in-app assistant, content generation for marketing, product images, video clips. A one-page list is enough.
- Chatbot: add the disclosure message at the start of the conversation and check it on the phone too, where space is tight and messages get cut off.
- Generated content: confirm that the tools used apply machine-readable markings and that the publishing process keeps them to the very end.
- Deepfakes and realistic images: set an internal rule of visible labelling, applied before publication, not after the first complaint.
- Vendors and agencies: ask them for written confirmation that the deliverables comply with Article 50 — towards your public, the responsibility remains yours.
If nobody on the team masters the technical side of these checks, the AI and digital marketing services cover exactly this area, and the frequently asked questions about artificial intelligence projects clear up the usual uncertainties in advance.
The next step: in the coming days, take the inventory of the AI contact points in the company. It is a one-page list, not a project — but it is the difference between knowing where you stand and finding out from others. Neoxis, an IT services company founded in 2015 in Pitești, works with SMEs across Romania on exactly this type of implementation.