Chatbots and AI content in your company — Article 50 transparency obligations in practice

From 2 August 2026 the transparency obligations under Article 50 of the EU AI Act stop being theory. Many companies already know THAT they must label chatbots and AI-generated content. Far fewer know HOW to do it correctly. So we take apart the three most common SME scenarios: a chatbot on the website, marketing content and email communication. We show examples of good labelling, explain what is not enough, and describe how we implemented these requirements ourselves at eskom.ai.

Scenario 1: a chatbot on the company website

What the provision says

An AI system intended to interact with people must be designed so that the user knows they are talking to artificial intelligence — unless this is obvious to a reasonably well-informed person in the given context. In practice: don't count on "obviousness". Bots are now natural enough in conversation that context is rarely enough.

How to label correctly

• A notice before or at the start of the conversation, e.g.: "You are talking to an AI assistant. The answers are generated by artificial intelligence."
• Visibility that is permanent, not one-off: a notice in the chat window header works better than a single bubble that disappears once you scroll.
• Plain language. "AI assistant", "artificial intelligence", not "a conversational system based on machine learning models".
• A clear path to a human. This is not strictly an Article 50 requirement, but it's good practice that builds trust and reduces frustration.

What is NOT enough

• Calling the bot "Assistant" or giving it a name, because a name suggests a human rather than a machine.
• A notice buried in the terms of service or privacy policy: the user should know at the moment of interaction, not after reading documents.
• A robot icon with no word of explanation (the symbol can be read as a graphic element, not a message).

Scenario 2: AI-generated marketing content

What needs labelling

Content (text, images, audio, video) generated or substantially modified by AI, particularly content published to inform the public. A separate, stricter category is deepfakes: material that deceptively imitates real people, places or events. These must always be disclosed, and clearly.

How to label correctly

• An annotation next to the content, e.g. at the end of an article: "Content prepared with the involvement of AI". Short, visible, in the same language as the content.
• Consistency: label by one rule across the whole website, not "case by case".
• AI-generated graphics and video: a caption or note next to the material, and for realistic likenesses of people, unmistakably clear disclosure.
• The technical layer: Article 50(2) provides for marking content in a machine-readable format (e.g. metadata, watermarking). New generative systems must have it from 2 August 2026; the transition period until 2 December 2026 applies only to systems placed on the market earlier. Another reference point is the voluntary Code of Practice on the transparency of AI content (AI Office, June 2026): two-layer labelling, meaning metadata plus a visible notice for humans.

What is NOT enough

• A general declaration that "our company uses AI" on the "About us" page. The label has to sit next to the specific content.
• Labelling only some of the material "so as not to hurt the image". Such inconsistency is easy to spot and hard to defend.
• Hiding the annotation in light grey 8-pixel type in the footer: formally present, practically uninformative.

One distinction matters here. Content written by a human and merely checked by AI for typos is a different thing from an article generated entirely by a model. A sensible internal policy distinguishes these cases and sets the threshold above which you label.

Scenario 3: email communication

Emails written with the help of an AI assistant but verified and sent by a human are, as a rule, not the main target of Article 50. The situation looks different when:

• replies are sent automatically (e.g. AI answers customers with no human involved): then the recipient should know they are corresponding with an AI system;
• a newsletter or mass communication is generated by AI, which you label consistently with your policy for published content.

Good practice: a short note in the footer of automated replies, e.g. "This reply was generated by an AI system. If you need to reach a consultant, reply to this message."

How ESKOM AI implemented this in-house

We're not writing about theory. We implemented these requirements on our own website, eskom.ai, back in spring 2026:

• The chatbot on the site displays a visible notice that the user is talking to artificial intelligence before the conversation starts.
• Content prepared with the involvement of AI carries a clear annotation directly next to the material, following one consistent rule for the whole site.
• In the technical layer, pages with AI-generated content additionally send an X-AI-Generated header, and exported documents contain the relevant metadata. That way the label is read not only by people but also by machines, in line with the direction of Article 50(2).

This combination (a notice for humans, a machine signal, a consistent policy) turned out to be simpler to implement than we expected. The most work went not into the labelling itself, but into tidying up the process: who decides that a piece of content needs an annotation, and who checks it before publication.

Where to start in your own company

1. Take stock of the touchpoints: where customers talk to AI, where you publish AI content, where an automated system answers emails.
2. Prepare standard label wording (chatbot, article, graphic, auto-reply) and write it into your internal policy.
3. Implement, test on real users, train the team. And write down what you did: the documentation is your evidence of due diligence.

FAQ

Do I have to label texts that AI only edited or shortened? What decides is whether the content was generated or substantially modified by AI. Fixing typos is not the same as generating an article. Adopt an internal, consistently applied threshold, and in borderline cases, label.

Does the chatbot notice have to be in Polish? The notice should be understandable to the user, so in practice in the language you serve the content in. A multilingual site should carry the labels in every language version.

What about content published before 2 August 2026? The priority is content published from the date the provisions apply and content still in active use (e.g. product pages, ongoing campaigns). Review archival material as you go, starting with the most visited pages.

Check your labels before someone else does

ESKOM AI offers advisory and implementation services in AI compliance: from a review of your customer touchpoints, through label design, to an internal AI usage policy. We also develop tools that support compliance, including data anonymisation and monitoring of regulatory changes.

Book a free consultation via the form at eskom.ai/pl/kontakt. Together we'll check whether your labels will survive the 2 August test.

This article is for information purposes and does not constitute legal advice. Consult a lawyer before making decisions on EU AI Act compliance.