Article 50 of the EU AI Act, the part covering transparency of AI-generated content, applies from 2 August 2026. Headlines along the lines of "you must now label everything AI touched" appeared within days. They are wrong, and the error is expensive in both directions: some businesses now label every icon just in case, while others wave away the material that genuinely does need disclosing. Below we break the rule down from the point of view of a company that uses AI for graphics, photos and text.
The AI Act splits the roles: provider versus your business
Article 50 of the AI Act splits duties between two different roles, and most of the confusion comes from mixing them up.
- A provider is the company that builds an AI system and puts it on the market: OpenAI, Google, Midjourney and the like. They carry the duty to mark outputs in a machine-readable format, meaning signed metadata, imperceptible watermarks or content fingerprints.
- A deployer is your business, using an off-the-shelf tool in a professional capacity. Your duties are narrower and concern disclosure visible to a human being in two situations: deepfakes, and text published in order to inform the public on matters of public interest.
The practical upshot: if you generate an image in Midjourney or text in ChatGPT, you do not have to embed any watermark or metadata yourself. Machine-readable marking is the tool provider's problem. Yours begins where a viewer could mistake generated content for something authentic.
One caveat to that reassurance: the role does change for anyone who builds a generative system or puts their own name on one. An ordinary subscription does not do that.
Images and video: when a graphic becomes a deepfake
This is the heart of the matter for most businesses, and where the common mistake sits. A deepfake in the legal sense is not only a politician's face pasted onto someone else. The definition covers AI-generated or manipulated content that resembles existing persons, objects, places or events and could falsely be taken as authentic. The European Commission's July 2026 guidelines read this broadly: the content need not depict a real person or a specific object, it is enough that it looks realistic and a viewer could believe it is a photograph.
In practice that pulls in material few people had filed under this heading:
- realistic product shots generated instead of a photo session,
- clothing or accessories shown on a model who does not exist,
- "team photos" or company interiors that were never there,
- virtual staging used in property listings,
- food photography for menus and delivery apps,
- genuine photos heavily reworked by AI, for instance with a replaced background.
What does not need labelling is content no reasonable viewer would take for a photograph: illustrations, icons, stylised graphics, drawings, abstract backgrounds. The test is the impression on the viewer, not the technology. A dragon is not a deepfake, because dragons do not exist. A photo of a sofa in a living room that was never photographed is.
Clearly artistic, satirical or fictional work sits in its own category. The rule allows a lighter touch there: disclosure must not spoil the enjoyment of the work, so it can go in the closing credits, for example. Purely commercial material does not get that concession.
Text: far narrower than everyone assumes
Text is where the panic ran highest and where the duty is surprisingly narrow. Disclosure applies to AI-generated text published in order to inform the public on matters of public interest. That means politics and democratic processes, public administration, justice, fundamental rights, security, public health, environmental protection, consumer safety, and significant economic or scientific developments under public debate.
Product descriptions, advertising copy, quotes, customer correspondence and internal instructions are not automatically caught, because they are not published to inform society about public matters. An ordinary services site or online shop largely falls outside this provision.
There is also an exemption that matters most to any business running a blog: no label is required where the text has undergone genuine human review and a specific person or company holds editorial responsibility for the publication. Mind its limits, though. A language check, a spellcheck or a formal sign-off is not enough. It means real control over the content, including responsibility for what it says.
How to label so that it actually counts
The rule requires disclosure that is clear, distinguishable and given no later than the viewer's first exposure to the content. That rules out the convenient options: a note buried in your terms, in the privacy policy or in a page footer does not satisfy it. Disclosure has to be perceivable by a human without any technical tools, so a visible caption or label on the content itself, and an audible message for audio material.
The good news is that you do not have to invent the wording. In June 2026 the European Commission published a voluntary Code of Practice on transparency of AI-generated content, and its annex carries ready-made, user-tested EU icons: one for fully generated content, one for modified content and a basic one for an interactive layer. The Code is voluntary, but anyone not using it has to show they achieve the same result another way. Around 190 organisations had signed it by the end of July 2026.
For a small business this comes down to one organisational decision: agree where the label goes on your site and in your materials, then apply it consistently instead of relitigating every single case.
Dates and penalties, without the scaremongering
Article 50 duties apply from 2 August 2026. One piece was deferred: the simplification package adopted in summer 2026 gave providers of systems already on the market before 2 August until 2 December 2026 to implement machine-readable marking. Systems released after 2 August must comply immediately, and every other duty, including your disclosure of deepfakes, applies with no transition at all.
Penalties for breaching transparency duties reach 15 million euro or 3 percent of worldwide turnover. Before that figure lands too hard: for small and medium businesses and start-ups a mitigating rule applies, under which the lower of the two figures is used rather than the higher. Supervision is also still being set up. Poland's national AI act was signed by the President on 24 July 2026 and creates a Commission for the Development and Security of Artificial Intelligence, which under the statutory deadlines starts work in autumn 2026.
For a small company the real risk today sits somewhere other than an administrative fine. A customer who works out that the "photo" of a product or of your premises never existed draws conclusions about the whole offer. A label costs one sentence; lost trust costs more.
How we handle this in projects
We treat it as part of the standard rather than a separate project. Generated graphics are handled according to what they are: illustrations and decorative elements stay unlabelled, while photorealistic material that could pass for a photograph either gets a visible label or is replaced with a real photograph where the client has one. Offer and blog copy is written and edited by people, so editorial responsibility is unambiguous. The chat on the site introduces itself as an AI assistant from the first message, which we covered in more depth in our post on AI chatbots, GDPR and the AI Act. There is more on producing the material itself on our AI graphics and video page, and broader context on AI for business.
Disclaimer: we are an implementation studio, not a law firm. This post is a practical guide rather than legal advice, and part of the interpretation around the scope of "deepfake" is recent and will be refined in supervisory practice. For unusual cases it is worth having your materials reviewed by a lawyer.
In short
The whole rule fits into three sentences. Machine-readable marking is the tool provider's job, not yours. Visibly disclose anything that looks like a real photo, recording or film but is not. Leave your business copy alone, as long as a person writes it and stands behind it. If you want to check which of the material on your site is affected, get in touch and we will go through it together.