AI labelling requirements: What you’ll need to do from 2026 onwards

From 2 August 2026, the AI labelling requirement will come into force in the EU. Many people think this is an issue for large tech companies. In fact, it affects almost every brand that uses AI for images, text or chats. This guide shows you what you need to label and how.

Please note: This article provides an overview of the legal situation and is not a substitute for legal advice. If in doubt, consult a solicitor about your specific circumstances.

What is the AI labelling requirement?

The AI labelling requirement stems from Article 50 of the European AI Regulation (AI Act) and requires disclosure when content is generated by AI or when people interact with an AI system. It is therefore not a standalone law, but rather part of the broader EU regulation governing artificial intelligence.

The Regulation itself came into force in 2024, but will be implemented in stages. Bans on particularly high-risk AI applications have been in force since February 2025, whilst obligations relating to very large AI models have applied since August 2025. The transparency obligations set out in Article 50, which are the subject of this article, constitute the next major phase.

The essence of Article 50 can be summarised in a single sentence: anyone who uses AI in customer interactions or uses AI to generate content that could be mistaken for the real thing must make this clear. You can find the full, binding version in the original text of the Regulation (EU) 2024/1689 on EUR-Lex, a more practical classification is provided by the European Commission FAQs on Article 50.

For many companies, this initially feels like yet another regulatory issue that primarily affects large tech firms. In fact, the focus lies elsewhere in many respects: it is about what every company must disclose when it uses these tools in its own marketing, on its own website or in its own customer service.

This shift is a deliberate one. The European legislator’s primary aim in Article 50 was not to regulate the technology itself – that is covered by other parts of the AI Act – but rather to address the impact on people who come into contact with AI-generated content without realising it.

When does it come into force, and who does it affect?

The labelling requirement will apply from 2 August 2026 to virtually every organisation that uses generative AI, regardless of the size of the organisation. Although there is a transition period until 2 December 2026, this only applies to a specific aspect of the requirement, not to the visible labelling requirement as a whole.

It is therefore worth clearing up a common misunderstanding. According to the final version of the Omnibus Amendment, the deadline of 2 December 2026 applies exclusively to the technical, machine-readable labelling required under Article 50(2) for AI systems that were already on the market before 2 August 2026. This obligation primarily affects large AI providers anyway, not companies that merely use AI outputs. No comparable grace period applies to visible labelling, which is actually relevant for most agencies and brands – for example, in the case of chatbots, deepfakes or AI-generated texts of public interest. It will apply as standard from 2 August 2026, regardless of whether the underlying content was created before or after that date.

In practical terms, this means that whether a company is a large corporation or a two-person business makes no difference to this fundamental obligation. An overview of the deadlines is provided by the EU AI Act Explorer: transparency rules under Article 50. So anyone currently working on campaigns or website relaunches for late summer 2026 should factor in visible branding from the outset, rather than mistakenly putting it off until December.

The allocation of roles also warrants closer examination. The AI Act distinguishes between providers – that is, companies that develop an AI system and make it available under their own name, such as OpenAI or Google – and operators, that is, anyone who uses an AI system in a professional capacity. Most companies reading this article are, under this framework, operators rather than providers. This does not alter the fundamental obligation to provide visible labelling, but it does exempt them from the technical requirement for machine-readable labelling at system level, which falls to the major providers.

What exactly do you need to label?

AI-generated or modified images, videos and audio files, deepfakes, certain AI-generated texts on topics of public interest, and chatbot interactions must be labelled. These four categories can be clearly distinguished from one another.

Firstly: Direct interaction with an AI, for example via a chatbot or an AI voice on the telephone. In such cases, it must be clear that a real person is not responding. Secondly: deepfakes, i.e. image, video or audio content that bears a deceptive resemblance to real people, places or events and could be mistaken for the genuine article. Article 3(60) of the AI Regulation defines a deepfake literally as AI-generated or AI-manipulated content that resembles real persons, objects, places or events and would falsely appear to be genuine. It is important to note that the law is expressly not limited to faces or people; objects and places are also explicitly covered. Thirdly: AI-generated texts on topics of public interest, such as politics, health or the judiciary, provided they are published without genuine editorial review. Fourthly, on a different level: providers of AI systems themselves must additionally label their output in a machine-readable format, for example via watermarks or metadata.

What matters most is always the effect, not the tool. Whether an image was created using traditional CGI or AI is, in itself, irrelevant. The decisive factor is whether an average viewer might mistake the content for a real photograph. A realistic-looking person who has never actually been photographed in that way falls into this category. A clearly recognisable fantasy creature in an illustration generally does not, because nobody assumes that something that actually exists has been depicted there.

This difference can be clearly illustrated by two similar examples. An AI-generated, photorealistic dragon winding its way around a product does not need to be labelled, because nobody would believe a real dragon to be possible, regardless of how realistic the rendering appears. An AI-generated human, on the other hand, holding the same product and looking deceptively real, must be labelled because a human could, in principle, be real. The difference therefore lies not in the image quality, but in whether the depicted object could actually exist at all.

An important, official clarification on this point was only issued shortly before the deadline: a real product, such as a car, shown in an advert against a background generated entirely by AI, does not have to be labelled as a deepfake, provided that the product itself is depicted accurately and is not misleading. A car driving along a fictional, generic road is therefore not automatically subject to the labelling requirement, provided that only the background has been replaced and the product itself remains unchanged and accurate.

For everyday image editing, a simple rule of thumb applies: if you are simply removing a distracting element in the traditional way or adjusting the exposure, this counts as standard image editing and does not require labelling. If, on the other hand, generative functions are used to create new, realistic-looking image content – such as a new environment, a replaced product or a scene expanded using Generative Expand – the labelling requirement generally applies. The same applies if a real, recognisable person is placed into a new scene using AI, even without a traditional face swap. This also counts as a deepfake, because the content may appear real to viewers.

What is exempt from labelling?

Exceptions to this are purely assistive uses of AI, such as spell-checking, and texts that have been editorially reviewed and approved by a human. Both exceptions require that a human bears responsibility for the content, rather than merely signing off on it as a formality.

Purely spelling or grammatical corrections carried out by AI do not substantially alter the content of a text and therefore do not need to be flagged. The same applies to minor adjustments to the contrast of a genuine photograph or the removal of image noise. The situation is different, however, if a text is generated entirely by AI and published on a topic of public interest without any genuine review of its content. If, on the other hand, an expert reviews the text, corrects errors and assumes editorial responsibility, there is no obligation to label it, even if the AI played a significant role in its creation. The decisive factor here is having genuine authority to make changes to the content, not merely a cursory glance over the text. A purely formal ‘approve’ click without any content review is not sufficient; a documented approval process with a named person in charge is more likely to be acceptable.

There is a similar grey area in customer service. An internal, AI-assisted pre-formulated email, which an employee reads and takes responsibility for before sending, does not need to be labelled as AI-generated. A fully automated chat system, however, which communicates directly without human review, falls under the labelling requirement for direct AI interaction, regardless of how well the responses are formulated.

Content teams would therefore do well to follow a clear internal rule of thumb: as long as a person reads the final content, understands it and puts their name to it before it goes live, you’re generally on the safe side of the exception.

How do you label things correctly?

You ensure correct labelling by placing the notice in a visible position on the content itself, so that it is recognisable from the very first access, in addition to a machine-readable mark. A notice in the legal notice section alone is expressly not sufficient.

The regulation does not specify in detail exactly how the labelling should be worded or designed. The European Commission’s Code of Practice proposes two options for an icon: “AI GENERATED” for content generated entirely by AI, even if text or layout has been added subsequently, and “AI MODIFIED” for real-world material that has been altered using AI at a later stage. The use of these icons is voluntary; a clear, self-explanatory phrase works just as well.

In practice, major platforms have already established their own solutions: YouTube labels content that has been altered in this way as ‘altered or synthetic content’, Meta displays an ‘AI Info’ notice directly on the post, and TikTok labels content as ‘AI-generated’. At the machine-readable level, the Content Credentials from the Coalition for Content Provenance and Authenticity (C2PA) a technical standard has been established that embeds provenance information directly into a file’s metadata. It is important to note that, in the case of deepfakes, metadata alone does not replace visible labelling, and platform-specific labels are not necessarily sufficient on their own. The safest approach remains labelling that is attached directly to the content itself.

In practical terms, this means that a short, clearly legible text label directly on the image, video or post constitutes the visible part of the labelling. A single, central notice in the footer or in a general banner is risky when dealing with mixed content, as viewers cannot tell which specific image is affected. Content-specific labelling for each individual image is therefore the more legally sound approach. We currently apply exactly this principle when labelling our own AI-generated visuals in our content creation.

A common mistake in practice is to only schedule the labelling at the end of a project, essentially as a box to tick before publication. It makes more sense to treat the labelling as an integral part of the template from the outset, whether it’s a social media template, an image template or a chatbot module.

What brands should be doing now

Brands should now carry out an assessment of their use of AI, establish a uniform labelling standard and assign clear responsibility for it. These three steps can be implemented regardless of the size of the company. Identify where generative AI is currently being used within the company

1. Specifying how a label should appear visually and in machine-readable form

2. Clearly define review processes for texts on sensitive topics, including a documented approval stage

3. Produce new content published from 2 August 2026 onwards in a way that complies directly with labelling requirements

4. Document the creation dates of existing AI content in full; if in doubt, it is better to flag it

5. For existing technical AI systems, check whether the separate transition period until 2 December 2026 applies to machine-readable labelling

The assessment clarifies where generative AI is used within the organisation. The labelling standard specifies what a notice should look like and where it should be placed. The accountability mechanism ensures that someone actually checks the content before it is published, rather than only after a complaint has been received.

One particularly relevant practical issue concerns existing material. According to the current EU guidelines, the decisive factor in determining whether content is covered by the new requirement at all is the date of creation, not the date of publication. An AI-generated image that can be proven to have been created before 2 August 2026 therefore does not need to be labelled retrospectively, even if it is only published after that date. It is important to note that this interpretation has changed: in an earlier version of the guidelines, the Commission still based its assessment on the publication date; it was only later that this was corrected to the creation date. This shows that these guidelines are still evolving and do not constitute a definitive legal interpretation. Anyone relying on a creation date prior to the cut-off date should therefore document this comprehensively, for example via time stamps in the tool or in project files. In the absence of such evidence, the principle in case of doubt is: it is better to label the document than to risk being unable to rely on a date prior to 2 August 2026 in the event of a dispute.

To summarise as a brief checklist of actions:

1. Identify where generative AI is currently being used within the organisation

2. Specifying how a label should appear, both visually and in machine-readable form

3. Clearly define review processes for texts on sensitive topics, including a documented approval stage

4. Produce new content published from 2 August 2026 onwards in a way that complies with labelling requirements straight away

5. For existing technical AI systems, check whether the separate transition period until 2 December 2026 applies to machine-readable labelling

Anyone who takes this checklist seriously will, in most cases, not need to purchase any new tools; it is sufficient to add a clear labelling step to existing approval processes. The severity of the penalties is also important: breaches of transparency obligations can be punished with fines of up to 15 million euros or 3 per cent of global annual turnover, whichever is higher.

How this diligence pays off is directly linked to trust, which is at the very heart of this regulation. You can read more about why trust is the hardest currency in our article on Brand Integrity, and we show how generative AI is changing the visibility of content as a whole in our article on Generative Engine Optimisation. If you’re looking for support in the responsible use of AI in your content creation, get in touch with us via our Content creation with pechschwarz to.