AI content must be recognisable

Since 2 August 2026, your audience has to be able to tell when AI is involved. What that means in practice for a chatbot, an advertisement using generated models or a film full of artificial imagery is worked out in new Commission guidelines. They also show where the room to manoeuvre lies.

Article 50 of the AI Act has applied since 2 August 2026. It sets out the transparency obligation for providers and deployers of certain AI systems. Anyone using AI to create or distribute content now has a duty to inform people that AI has been used. The obligation is broadly worded in the Regulation itself. On 20 July the European Commission therefore published guidelines on how to comply in practice, alongside a Code of Practice on the transparency of AI-generated content. Below I work through the obligation and consider the exceptions the Commission develops in the guidelines.

What does the transparency obligation involve?

In the AI Act, the European legislature has laid down what the development and use of AI systems must satisfy. Those rules take effect in stages. The transparency obligation in Article 50 belongs to the stage that began on 2 August 2026. The thinking behind it is that people should be able to recognise when they are dealing with AI, so as to counter deception and manipulation. An important stage, then.

Article 50 covers four situations. It applies first to AI systems that interact directly with people and to AI systems that generate content. It also covers those who use systems for emotion recognition or biometric categorisation, and finally those who publish deepfakes or AI-generated text. The first two obligations rest with the provider of the AI system. The other two apply to the deployer, the party that ultimately puts the AI-generated content before an audience.

Obligations for providers of AI systems

Two obligations apply to providers.

First, AI systems that interact directly with people, such as chatbots, avatars and AI agents, must be designed so that users know they are dealing with AI. That information has to be given from the start of the first interaction. The obligation does not apply where it is already obvious to a reasonably well-informed person that they are communicating with AI.

Second, AI-generated or AI-manipulated content (audio, images, video and text) must carry a machine-readable marking, so that it can be detected as such. The guidelines require that marking to be effective, reliable, robust and interoperable. How to implement it technically, through watermarking or metadata for instance, is developed further in the Code of Practice.

Obligations for deployers of AI systems

Information obligations also apply to those who deploy AI systems.

For deepfakes, it must be made clear that the content has been artificially generated or manipulated. A machine-readable watermark from the provider will not suffice on its own. A label that people can see or hear is needed, at the latest when the public is first exposed to the content.

The AI Act defines a deepfake as AI-generated or AI-manipulated image, audio or video content resembling existing persons, objects, places, entities or events, which would falsely appear to a person to be authentic or truthful. The Commission goes on to give the concept a broad reading in the guidelines. An AI character need not actually depict a specific existing person or place. It is enough that the image or video imitates a realistic person, object or place that exists, could exist or could reasonably have existed. On that approach, almost every AI-generated ‘person’ or place will qualify as a deepfake under the AI Act.

Anyone publishing AI-generated text to inform the public on matters of public interest must say so where that text is published without human review. And anyone using systems for emotion recognition or biometric categorisation must inform those concerned about how the system works.

The exceptions: where is the room to manoeuvre?

The guidelines also work out the exceptions, in other words when Article 50 does not apply.

The marking obligation does not apply where the AI system performs no more than an assistive function in standard editing. Using AI to crop a photograph or tidy up a text does not in itself produce AI content that has to be marked. Where AI forms part of the post-production of a film or a podcast, the duty to disclose that AI has been used therefore does not automatically follow. Short sequences of numbers or characters, source code and output that never reaches people also fall outside the obligation.

For deepfakes forming part of an evidently artistic, creative, satirical or fictional work, a lighter obligation applies. The disclosure must be made in an appropriate manner that does not hamper the display of the work. A science fiction film need not carry a distracting label across the image, though the audience must be able to see somewhere that it contains artificial content, which can be done clearly in the credits.

For AI-generated text, the disclosure obligation falls away where there is human review or editorial responsibility. The guidelines take that review seriously. It must involve a deliberate substantive assessment by someone with relevant knowledge and professional judgment, with a person or an editorial team bearing final responsibility for publication. A spellcheck or a purely procedural check expressly does not count as human review.

The guidelines and the Code of Practice

The obligations in Article 50 are generally worded. The guidelines of 20 July give them substance. Alongside them sits the Code of Practice on the transparency of AI-generated content, a voluntary and practical instrument. Those who follow the code can be confident that they meet the marking and labelling obligations of Article 50. Those who do not will have to demonstrate compliance in another way at the regulator’s request.

Enforcement and transitional rules

Enforcement lies primarily with national market surveillance authorities. In the Netherlands it is divided among existing sectoral regulators, with the Dutch Data Protection Authority (Autoriteit Persoonsgegevens) and the Netherlands Authority for Digital Infrastructure (Rijksinspectie Digitale Infrastructuur) taking the leading roles. Breach of the transparency obligations can attract a fine of up to 15 million euros or 3% of worldwide annual turnover, with proportionality taken into account for smaller companies.

Transitional rules do apply. For AI systems placed on the market before 2 August 2026, the machine-readable marking obligation has a run-off period until 2 December 2026. Content generated before 2 August 2026 need not be labelled retrospectively, although the Commission encourages it where possible.

What does this mean for the media and creative sector?

If you generate visual material with AI, be aware that a transparency obligation rests on you. Where you use AI in post-production or finishing, in an assistive role in standard editing, the position will probably be different. Know which side of that line each production falls on, and check it.

Where a disclosure obligation does apply, the disclosure must be distinguishable and perceptible to people. It must also be available when the content is first made public. The Commission gives the following practical examples of disclosures that meet the requirement:

• a visible label in the advertisement;
• clear wording such as “This advertisement contains AI-generated images”, “The people and surroundings shown were generated with AI”, or “AI-generated advertising content”;
• a comparable notice directly alongside the video, image or social media post.

The disclosure obligation does not apply to editorial teams using AI to produce articles, as long as genuine human editing takes place, with substantive assessment and editorial responsibility.

Do you use a chatbot on your website? Then the bot must identify itself as AI from the first contact, unless that is already evident.

Conclusion

The transparency obligation in the AI Act opens a new phase for everyone using AI to make and distribute content. The norm itself is general, but the guidelines and the Code of Practice now provide a concrete and workable framework, including practicable exceptions for creative work and for content with human editorial oversight. That is just as well, because enforcement is now possible and the potential fines are high.

Do you use AI in your productions, publications or communications, and are you wondering whether the transparency obligation applies to you and how to comply with it in practice? Please do get in touch. We can look together at where disclosures need to appear in your case, how they should be worded and where, at a minimum, you need to show them.