What Europe's new AI content transparency rules mean for you

According to the European Commission, new transparency rules for certain AI systems and AI content have applied since 2 August 2026. If you only read the headlines, you might think that every text created with AI assistance now needs a warning label underneath it.

The rule is not that broad. Article 50 of the AI Act distinguishes between conversations with AI, content created or altered by AI, deepfakes, and text on matters of public interest. Who must disclose something, to whom, and how differs by situation.

Four situations under one article

Article 50 of the AI Act contains four types of transparency obligations.

1. Direct interaction with AI. A provider of, for example, a chatbot must inform people that they are communicating with AI, unless that is already obvious. This information must be visible to the user by the time of the first interaction.

2. Technical identification of AI content. Providers of systems that generate text, images, audio, or video must design their systems so the output is identifiable in a machine-readable format as artificially generated or manipulated. This is primarily a technical marking within the content itself.

3. Deepfakes. An organisation using a deepfake must make it clear to the people seeing or hearing the content that the image, sound, or video was artificially generated or manipulated.

4. Text on matters of public interest. An organisation publishing AI-generated or manipulated text to inform the public on a matter of public interest must disclose this to readers. The disclosure must accompany the publication. It must be clearly visible when the reader first sees the text.

For organisations mainly using AI to create blogs, news articles, reports, or other public texts, that fourth category is often the logical starting point. Their use mainly involves text published externally. The technical obligations for model providers and the separate rules for deepfakes are therefore less often relevant to their everyday work.

That does not automatically mean they need to add a disclosure. An important exception applies to human editorial review.

Human editorial review makes a legal difference

The transparency obligation for text on matters of public interest does not apply when the content has undergone human review or editorial control and a person or organisation holds editorial responsibility for the publication.

The exception therefore revolves around two elements. Someone must actually assess the content and adjust it where necessary. It must also be clear who is responsible for what is ultimately published. Automatically adding a name beneath a text produced entirely by AI is not a convincing way to meet that requirement.

A blog, a fully automated news feed, and a deepfake therefore fall under different rules within Article 50 of the AI Act. For a blog with real human editorial review, the text exception can apply. An automated news feed may lack that human assessment. The transparency obligation for public text can therefore apply.

A deepfake falls under a separate rule for images, audio, and video. Human review does not simply remove the obligation to disclose its artificial nature.

Other laws and sector-specific rules may also apply. Article 50 is therefore one part of the assessment, not a general exemption for all AI content.

What does the European code of practice cover?

Alongside the law, there is, for example, the Code of Practice on Transparency of AI-Generated Content. This code of practice is voluntary and provides practical measures organisations can use to show how they meet specific obligations around marking and disclosure.

The code consists of two parts. The first describes how providers can technically mark AI content and make it detectable. The second concerns clear labels for deepfakes and certain public texts. Among other things, the code covers machine-readable markings, information for readers, and the use of recognisable European AI labels.

In an assessment from July 2026, the European Commission and the AI Board called the code an appropriate tool. Participation does not automatically prove that an organisation meets all the rules. Organisations may also use other, equivalent measures to demonstrate how they fulfil their obligations.

What does this mean for a typical SME?

Does your team use AI to create a first draft of a blog, after which an employee checks the facts, edits the substance, and takes responsibility on behalf of the organisation? Then the exception for human review and editorial responsibility is directly relevant. For that specific text obligation, you do not need an AI disclosure as long as both conditions are met.

Does your organisation automatically publish texts on social, political, or other matters of public interest without anyone assessing the content? Then you probably need to state clearly alongside the publication that the text was generated or manipulated by AI.

Does your organisation create or use realistic images, audio, or video depicting existing people, places, or events differently? Then the deepfake rules may apply, and the artificial nature must be made clear to the public.

Do you offer a generative AI system yourself under your own name? Then you also face the technical requirements for machine-readable marking and detection.

The right approach therefore starts with your actual use case. Who creates the content, what type of content is it, who sees it, and where does human review take place?

Why we still explain how we work

Our articles receive human editorial review and have a responsible author. As a result, the specific transparency obligation in Article 50(4) for AI-generated public text generally does not apply to these blogs. Even so, we choose to state in every article that AI supports us.

That choice is about trust and openness. In the first article of our Small team, big impact series, we show how we create blogs and articles with AI assistance. AI helps with research, a first draft, and various checks. The author determines the content, performs the final edit, and takes responsibility for publication.

We believe readers should know that division of roles, even when the law does not require a disclosure. The legal minimum is therefore not our only criterion. Clarity about how we work is part of how we use AI.

What steps can you take now?

First, map where your organisation uses generative AI and which output leaves the organisation. Then record for each application:

• Does someone communicate directly with an AI system?
• Is it text, images, audio, or video?
• Does the content inform the public on a matter of public interest?
• Is there demonstrable human review or editorial control?
• Who bears responsibility for publication?
• Does your organisation offer an AI system itself, or use a system from another provider?

With that overview, you can determine which transparency obligation applies and which technical, editorial, or organisational measures are needed. Also document how human review works and who gives final approval. Where the legal application is uncertain, seek specialist advice.

A note on how this was made: our content is AI-assisted. We use AI for research and drafting a first concept. A human reviews and edits the text and takes responsibility for everything we publish. Read how we make our content.