Do I need to label AI? 7 everyday situations from Article 50

From 2 August 2026, Article 50 of the EU AI Act requires transparency about AI-generated content and about direct interaction with AI systems. Unlike the high-risk requirements, which the Digital Omnibus on AI is set to move to 2 December 2027 for stand-alone high-risk AI systems, Article 50 largely stays on 2 August 2026. But what does that mean for the employee who writes letters, press releases or presentations every day? Do you really have to put "Made with AI" on everything now?

The short answer: no. The labelling obligation for content focuses mainly on what organisations publish externally: deepfakes and AI-generated text about matters of public interest. Article 50 also contains a separate transparency obligation when people communicate directly with an AI system, such as a chatbot. AI used as a tool behind the scenes carries no labelling obligation. The line sometimes runs differently from what you would expect.

The rule of thumb that works

Ask yourself one question before you publish something, or before you let an AI system interact with people.

For images, audio and video: does it look real or authentic while it is actually AI? Then label it.

For text: is the text published to inform the public about a matter of public interest, and is there no genuine human editing or editorial responsibility? Then label it.

For direct interaction: is someone talking to an AI system, and is that not obvious? Then say so.

That is the heart of Article 50. The law aims to prevent readers, viewers or conversation partners from being misled about what they see, hear or who they are talking to.

Important: Article 50 mainly concerns what goes out to the public. Internal emails, meeting notes, your own Word file or a one-to-one conversation with a colleague fall outside it.

7 real-world situations

1. A letter to a citizen, written with AI help

A municipal officer has ChatGPT review a draft letter, rewrites the content, checks it, signs it and sends it.

Label it? No.

This is one-to-one communication between the authority and the citizen, not publication to a broad audience. The officer has reviewed the substance of the text and takes responsibility for it. Here AI was a tool, not an author. Compare it with a spellchecker: nobody expects a note saying "Text corrected by Word".

Note: if the citizen communicates directly with an AI system, or if a decision is reached in a fully automated way, other transparency obligations may apply, both under the AI Act and under the GDPR.

2. A press release on the local authority website

A communications officer has AI write a first version, revises the text and publishes the press release.

Label it? No, provided there is demonstrable human editing.

A press release about local authority policy falls under "text that informs the public about matters of public interest" and therefore within Article 50(4). The law makes an explicit exception for text that has been editorially reviewed by a human, where that person or the organisation carries editorial responsibility.

The June 2026 Code of Practice and the draft guidelines of May 2026 both point towards a practical approach: not every AI-assisted document needs a label, but human editing must be genuine and demonstrable. In practice that means an identifiable editor with a documented procedure: set out in an internal process who is authorised to sign off on publication. A communications officer who genuinely edits meets the bar. A copy-paste without reading does not.

3. A PowerPoint with AI text and an AI image

A manager creates an internal strategy presentation with AI-generated text and a photorealistic image of "a team in a meeting".

Label it? It depends on the audience.

In practice: a caption "Image generated with [tool]" is enough. For the text, the same rule applies as for a press release.

4. Excel or Word with an AI plugin

An HR employee uses an AI plugin in Excel to write a formula, or has Word rephrase a paragraph.

Label it? No.

Here AI is a productivity tool. Spellchecking, grammar correction and auto-formatting fall outside the rules because the human keeps control over the substance. Practical test: has the AI substantially changed the meaning, structure or content? If not, there is no labelling obligation. If it has, the same editing rule as in situation 2 comes into play.

5. A chatbot on the website

A healthcare organisation puts a chatbot on its website to guide visitors through frequently asked questions.

Label it? Yes.

This is the clearest obligation in Article 50. Anyone interacting with an AI must be aware that it is an AI. Under Article 50(1), the provider must design the chatbot so that people are informed they are interacting with AI. As the organisation deploying it on your website, you should make sure that notice is actually visible to visitors. A welcome message "Hello, I am the virtual assistant of [organisation]" is enough. The "obvious AI context" exception applies only when it is clear beyond doubt to a reasonable person. For a visitor to a hospital website, that is rarely the case.

6. Meeting minutes, summarised by AI

A colleague uses an AI tool to summarise a meeting automatically and sends the minutes to those who attended.

Label it? Not for the attendees, and not automatically when it goes wider.

Minutes that go only to the participants are internal communication. If the minutes are published to inform the public about a matter of public interest, and there is no genuine human review or editorial responsibility, disclose that the summary is AI-generated or AI-assisted. If they are merely shared externally in a closed context, Article 50 does not automatically require a label, although disclosure may still be good practice. Good practice regardless of Article 50: mention it at the top. People want to know whether a human has checked the facts.

7. A photo in the annual report, AI-generated

The annual report contains a photorealistic image of "employees in action". It was created by AI.

Label it? Yes.

An annual report is publication to the public. A photorealistic image that suggests it depicts real people or situations falls under the deepfake provision. A caption "AI-generated image" or "Image generated with [tool]" is mandatory. The editing exception does not apply here: it applies only to text, not to deepfake images.

A clearly stylised illustration that does not falsely appear authentic normally falls outside the deepfake rule. If the content is still a deepfake but forms part of an evidently artistic, creative, satirical or fictional work, disclosure is still required, but it may be given in a way that does not hamper the display or enjoyment of the work.

Important exceptions and nuances

Not every exception applies to every situation. Article 50 has exceptions per type of obligation, not one general list.

Outside these cases the rule is: when in doubt, label it. It costs nothing and it wins trust.

A word about the technical marking (Article 50(2))

Alongside the visible labels above, Article 50 has a technical side: providers of generative AI must mark their output in a machine-readable way and make it detectable, so that systems can automatically recognise that something was created by AI. For systems that were already on the market before 2 August 2026, the Digital Omnibus gives extra time for this until 2 December 2026. For most organisations that use AI and do not provide it themselves, this is mainly something to check with suppliers, not something you have to build in technically yourself.

The one rule that covers everything

If you have to remember this article in one sentence:

Anything that becomes public and looks human-made while it is AI should be labelled. Anyone who talks to an AI must know it. The rest does not.

No lawyer needed. Three short questions, three answers, done.

Written by Rob Ummels in collaboration with Claude (Anthropic). Responsibility for content: AIAdopt.