HomeAI Build SprintProposition 01AI employeeProposition 02Use casesCasesInsights
About us
Plan an introduction

Insights

Your chatbot and AI content: what the transparency rules now require

16 Nov 2026Xavier Fung-A-Jou8 min read
AI-verordeningTransparantieChatbotAI-inhoudCompliance
Your chatbot and AI content: what the transparency rules now require

On 3 September 2026, Digitaal Vlaanderen published a message for Flemish government bodies: Article 50 of the European AI Act has come into effect. Anyone who deploys a chatbot or digital assistant must let citizens know they are talking to AI. Anyone who creates images or texts with AI must, in certain cases, disclose that too.

These rules do not only apply to governments. Since 2 August 2026, they apply to every organisation that uses AI to talk to people or to create content for the outside world. And on 2 December 2026 the last transition period ends: from then on, generative systems that were already on the market before August must also mark their output in a machine-readable way, according to the amendment to the AI Act adopted this summer.

Does your organisation have a chatbot on its website, an assistant that answers customer emails, or a team that creates news items and images with AI? Then this article explains what Article 50 requires, what falls outside it, and which 4 steps to take now.

What Article 50 actually requires

Article 50 contains 4 obligations. They apply to all AI systems within its scope, not just high-risk systems.

  • Talking to AI. Anyone who provides a system that communicates directly with people must make sure those people know they are dealing with AI. Unless that is already obvious to a reasonably well-informed person.
  • Marking generated content. Providers of systems that create text, images, audio or video mark that output in a machine-readable format, so it can be recognised as created or manipulated by AI.
  • Emotion recognition and biometric categorisation. Anyone who uses such a system informs the people exposed to it.
  • Deepfakes and texts on public matters. Anyone who uses AI to create images, audio or video that appear real must disclose this. The same applies to AI-generated text published to inform the public on matters of public interest, unless a human has reviewed that text and someone holds editorial responsibility.

The information must be clear, distinguishable, accessible, and provided no later than the first interaction. Under Article 99, non-compliance can lead to a fine of up to 15 million euros or 3% of worldwide annual turnover, whichever is higher. For SMEs, the lower of the two applies.

On 20 July 2026, the European Commission published guidelines explaining how it reads these rules. They are not binding, but they are intended to guide the supervisory authorities that enforce the rules. Earlier, on 10 June, the code of practice on marking and labelling AI-generated content was published. According to the Commission, about 190 organisations had signed that code by the end of July.

Provider or deployer: which rules apply to you

The AI Act divides the obligations between two roles. The provider develops an AI system or puts it into service. The deployer uses a system under its own authority. An organisation can be one, the other, or both.

That distinction determines what you have to do. If you buy an off-the-shelf tool, marking the output is mainly the supplier's job. If you have your own chatbot or assistant built and put it into service under your own name, you may be a provider yourself, and therefore also responsible for the notice to your customers. According to an analysis of the guidelines by law firm Stephenson Harwood, an organisation that builds on an existing model may rely on the marking that model already applies, but must be able to show that it works.

So make a short list for each application: what does the system do, who built it, who deploys it, and does anyone outside your organisation see the output?

Step 1: make your chatbot say it is a chatbot

The most common situation is a chatbot or assistant that talks to customers, citizens or job applicants. The guidelines ask for the notice to be built into the design and to appear at the first interaction. Examples Stephenson Harwood cites from the guidelines: a short plain-language sentence at the start of the conversation, a persistent label on screen, or a spoken statement for a voice assistant.

What is not enough: a vague name such as "assistant", a notice that only appears in the terms and conditions or the website footer, or only a technical mark the user cannot see. The summary in The National Law Review also calls references in terms, footers and vague labels unlikely to be sufficient.

This also matters for assistants with their own persona. In the culture chatbot we built for DEN, each assistant has its own role, tone and scope. The more human an assistant sounds, the more important it is that the visitor knows from the first sentence that they are talking to AI.

If an AI agent acts independently on someone's behalf, for example by booking an appointment or sending a message, the guidelines say it must also make clear on whose behalf it acts, and repeat this at key moments. Read more about keeping control of agents in our article on agents with access to your systems.

Step 2: decide which AI content needs a label

Not everything you create with AI needs a label. The obligation for deployers mainly concerns two types of content.

Deepfakes. Images, audio or video created or manipulated with AI that resemble real people, places or events. Think of a video with a voice that sounds like an existing employee, or a photo of a situation that never happened. The label must be visible or audible without special tools, the first time someone encounters the content. A lighter obligation applies to evidently artistic, satirical or fictional work.

Text on matters of public interest. A news item, public information or a press release written with AI must carry a notice, unless a human has reviewed the text on substance and the organisation holds editorial responsibility for it. According to Stephenson Harwood, the guidelines require genuine substantive review, not a quick glance. And if AI still materially changes the text after that review, the exception no longer applies.

So for an ordinary customer email, an internal summary or a quote, this labelling obligation usually does not apply. Note, however: if an AI assistant writes that email on its own in a conversation with the customer, Step 1 may apply.

If you want to label visibly, you can use the EU icons for AI content: a basic icon, an icon for fully AI-generated and an icon for partially AI-modified. Using them is optional, the Commission writes, but the labelling obligation itself is not.

Step 3: know what falls outside the rules

The guidelines also list exceptions to marking by providers. According to Stephenson Harwood, these include short outputs such as alt text and captions, source code, AI translations, and standard editing that does not change the substance or meaning. Output that stays within a closed environment is also excluded, but only if it is not shared outside the organisation and is properly protected, for example with access controls.

Two points need extra attention. Content you created before 2 August 2026 does not need to be labelled retroactively. But a text you wrote with AI before that date and only publish afterwards is covered, unless the editorial control exception applies.

An internal tool is therefore not a free pass: as soon as the output goes to customers, citizens or the public, the rules apply. That ties in with what we wrote earlier about AI used outside the view of IT: you can only label what you can see.

Step 4: make it part of your process

Transparency is not a one-off change to your website. Every new application, every new campaign and every new channel raises the same question. So record:

  • In your overview of AI applications: for each application, whether Article 50 applies, which role you have and what the notice looks like.
  • In your editorial process: who does the substantive review of texts for the public, and where that is recorded.
  • In agreements with agencies and suppliers: who labels, and what happens if they use AI for content that appears under your name.
  • In your delivery checklist: a check on the notice for every new chatbot or assistant, just like a check on accessibility.

How NewWorks approaches this

For every solution we build, whether it is a knowledge base agent, a document agent or a full workflow, we determine in the first weeks who uses the solution and who sees the output. If the solution talks to customers or publishes content, we build the notice and the human review into the design, not afterwards. In an AI Build Sprint, that is part of validation, before anything is built.

What you can do this week: open every chatbot and assistant your organisation uses externally, and ask the first question the way a customer would. Can you clearly see in the first answer that you are talking to AI? If not, you know where to start. Want to think through your own situation with us? Get in touch.

Share

Curious how this would work for you?

Plan an introduction