EU AI Act: What applies to AI content from August 2026

The obligation does not apply retrospectively. It applies to content generated and published from the cut-off date onwards. You do not have to label content published before then retrospectively.

This article sets out the rules so that you know what needs to be done for each type of content. The focus is on images and video. Not legal advice, but practical guidance.

The two levels of labelling

The AI Act requires two separate things. Anyone who confuses them labels content incorrectly.

Machine-readable (Art. 50/2). An invisible marker in every AI output. The tool provider is responsible (OpenAI, Google, Meta, Midjourney …). Technically via C2PA metadata or an embedded watermark, for example SynthID.

Visible (Art. 50/4). A recognisable label on the published content. You, as the publisher (deployer), are responsible. Implemented via the official EU “AI” icon, clearly perceptible.

Key takeaway: A visible label does not fulfil the machine-readable obligation, and vice versa. For compliant content, you need both levels. The tool provides the machine-readable one. The visible one is your responsibility.

AI images & video: When does the obligation apply?

The trigger for visible labelling is called deepfake. The term is defined more broadly in the law than most people think.

Deepfake (Art. 3/60): AI-generated or AI-manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic to a person.

Two criteria, both must be met: it resembles something real, and it appears deceptively authentic.

The most common misconception: “A deepfake is the face of a real celebrity.” Wrong. The definition also covers realistic objects, places and scenes, entirely without a real person.

Must a photorealistic image without a real person be labelled?

As a rule, yes. The trigger is not “real person”, but Photorealism plus the appearance of authenticity.

• Photorealistic but fictional person, appears to be a real photo: Yes, deepfake.
• Realistic fictional place or scene, passes as a real photo: Yes, deepfake.
• Obviously artificial (CGI, cartoon, illustration, surreal): No, not a deepfake.

Rule of thumb: “Could someone mistake this for a real photo?” If yes: label it.

The artistic exception

This is the most important point for creative work, and the one most frequently misunderstood.

Art, satire and fiction do not exempt you from labelling. But they change the form: for clearly artistic, creative, satirical or fictional work, the obligation is limited to a disclosure “in an appropriate manner” that does not interfere with the enjoyment of the work (Art. 50/4).

In plain English: You do not need an intrusive label in the middle of the image. A label in the caption, in the credits or next to the work is sufficient. The notice must exist, but it may be discreet.

That is the difference between an advertising deepfake that imitates a real photo (clearly visible label) and an artistic AI series (notice in the caption).

Why surreal content has an advantage

The deepfake obligation hinges on one word: Authenticity. Content only becomes a deepfake if it could appear genuine. If this criterion does not apply, the visible labelling requirement does not apply either.

This is precisely where surreal and stylised visual language becomes strategically interesting:

• Obviously artificial aesthetics do not meet the criterion of ‘appears genuine’. This means they are not a deepfake and do not require a visible label.
• The content still conveys the brand, idea and impact, just without the suspicion of being a real photo.
• This reduces labelling effort and authenticity risk in one go.

The forecast is obvious: brands and agencies will increasingly replace photorealistic AI images with deliberately artificial looks. Not out of fear of the rule, but because surrealism restores the creative freedom that photorealism restricts from a regulatory perspective.

The tool’s machine-readable marking remains in place in both cases. It is the provider’s responsibility, not yours.

AI text

Text is also covered, but more narrowly. Published AI-generated text on topics of public interest must be disclosed (Art. 50/4).

The exception is crucial for agencies: the requirement does not apply if the text has undergone a human review or editorial control and a natural or legal person bears editorial responsibility.

In practical terms: AI writes the draft text, a person reviews it and takes responsibility for it, then you do not need a label. Pure automatic publication without review, on the other hand, does require one. Editorial responsibility here is not bureaucracy, but a mark of quality.

AI audio

Audio is also subject to the requirement, but is often overlooked. Synthetic speech, music and sound design require machine-readable marking (Art. 50/2). A cloned voice that resembles a real person and appears genuine is a deepfake just like any video and must be visibly disclosed (Art. 50/4).

Relevant for AI voice-overs, dubbing, voice cloning and AI music. The rule of thumb is the same as for images: could someone mistake the voice for real? Then label it. Clearly artificial or generic voices with no real-world reference do not require a visible label. The tool’s machine-readable marking remains.

How to label audio in practice

The notice must be clearly recognisable, at the latest upon first exposure.

• AI voice in a video: EU ‘AI’ icon plus a short text label in the image, for example ‘Voice AI-generated’.
• Audio only (podcast, advert, voice message): spoken notice at the beginning or a text label in the player, title, description or credits.
• Machine-readable: retain the tool’s watermark or C2PA metadata when exporting. You do not add these yourself; you simply must not discard them.

Chatbots & AI interaction

If a person interacts directly with an AI, they must be made aware of this (Art. 50/1). The chatbot must identify itself as AI, unless this is already obvious to an average user.

Relevant for every customer service bot, AI assistant or voice agent you build for clients. The information must be clear and recognisable no later than at the first interaction.

Additionally: if you use emotion recognition or biometric categorisation, a separate duty to inform applies (Art. 50/3).

The grey areas

Not everything is clearly defined. If in doubt, the rule is: document it and, if in doubt, label it.

• Photorealism threshold: not clearly delineated by law. Retouching, compositing and full generation lie on a spectrum, not in separate boxes.
• “Existing objects and places”: very broadly defined. The limitation comes through the criterion “appears real”.
• Standard editing: Purely assistive functions that do not substantially alter the input data do not trigger the machine-readable obligation (Art. 50/2). Where “editing” ends and “generation” begins remains a matter of interpretation.

Enforcement: What will actually happen from 2 August

Monitoring is carried out nationally, not from Brussels. Each Member State appoints a market surveillance authority to pursue infringements. In Austria, RTR’s AI Service Centre is the central point of contact.
Anyone who is careless with labelling falls under Article 99/4. The framework: up to 15 million euros or 3 % of worldwide annual turnover, whichever is higher. SMEs and start-ups are subject to whichever amount is lower. Penalties must be effective and dissuasive, but also proportionate.

Important in practice: no one patrols Instagram feeds. Authorities act on a case-by-case basis, usually following complaints. Affected persons have an explicit right to lodge a complaint (Art. 85). The amount takes into account the severity and duration of the infringement, company size, intent or negligence, and whether you cooperate and make improvements. A fine is not the first step; warnings and non-monetary measures are expressly provided for.

Practical takeaways for creators & agencies

• Photorealistic AI image for clients: if in doubt, label it visibly. EU “AI” icon, recognisable, upon first viewing.
• Export & further editing: preserve metadata, do not discard C2PA/SynthID.
• Every project: document, including the tool, prompt and labelling method. Creates evidence.
• Off-platform (website, newsletter, print): no platform labels automatically; here, it is your responsibility.
• Despite labelling: The rights of depicted persons and brands still apply. The label is not a free pass.

Clearly labelled work is not a disadvantage. It is a mark of quality and an argument when dealing with clients who expect legal certainty.

Where to get the official EU labels

The European Commission provides an official icon set. It is free to use, requires no attribution, and is part of the Code of Practice for labelling AI-generated content.

• Basic: AI was involved in deepfake content or published text.
• Fully AI-Generated: Content generated entirely by AI, without human elements (except for the prompt).
• Partially AI-Modified: existing human-created content was altered into a deepfake using AI.

There are four variants (black, white, each also with 50 % transparency), available to download as SVG and PNG.

Important: The icons are optional, but the labelling obligation under Article 50 is not. The icon alone does not ensure legal compliance. Responsibility for correct disclosure remains with you.

Download & details: EU Icons for labelling AI-generated content

Official tools & help

If you want to know what applies in your specific case, official bodies can help:

• Compliance Checker: Self-check to find out which obligations apply to you. artificialintelligenceact.eu
• AI Act Explorer: the full text of the legislation, searchable in all EU languages. artificialintelligenceact.eu/ai-act-explorer
• AI Act Service Desk: the Commission's official point of contact for detailed questions. ai-act-service-desk.ec.europa.eu
• AI Service Centre of RTR (Austria): national point of contact and information hub. digitalaustria.gv.at

Sources: EU AI Act, Article 50 (artificialintelligenceact.eu) · EU Icons for labelling AI-generated content, European Commission · Digital Austria / RTR · This article provides practical guidance, not legal advice.