Mandatory AI labelling from August 2026

Please note: it is currently often claimed that ALL AI-generated content must be labelled. HOWEVER, it is not mandatoryto label all AI-generated content; rather, labelling is required only when

• the content appears realistic
• deception is possible
• users might believe the content to be ‘real’

e.g.:

• deepfakes
• AI-generated characters appearing as real people
• scenes that appear real

Labelling is not usually required if content:

• is obviously artificial or artistic
• does not give rise to any expectation of deception (e.g. fantasy characters, illustrations, 3D characters, abstract visuals)

Yes, that is a crucial difference.

IMPORTANT:
If a text is generated entirely by AI and published without human oversight, there may be a duty to label it – particularly in the case of content on topics of public interest.

If the text is subject to strict editorial oversight, is revised and for which a human takes responsibility, this obligation generally does not apply.

For texts, therefore, what matters less is whether they are generated by AI, and more whether a human takes responsibility for them. I would always recommend the latter anyway. Do not publish AI-generated content without strict human oversight.

“[…] must disclose that the text has been artificially generated or manipulated. This obligation does not apply […] if the AI-generated content has been subject to human review or editorial control and a natural or legal person bears editorial responsibility for the publication of the content.”
(Art. 50(4) AI Act, abridged)

In short: No. Although the terms are closely related, they have different meanings.

The distinction:

Labelling requirement – the specific, legal obligation to label content.
Example:

• This video was generated using AI
• This image was generated using AI
• This logo was generated using AI
• This slogan was generated using AI
• This text was generated using AI

IMPORTANT: Not all AI-generated content needs to be automatically labelled by us. Labelling is particularly necessary where content appears realistic and users might mistake it for the real thing (e.g. in the case of deepfakes or AI-generated people).

‘The duty of transparency is the overarching concept, or rather the idea behind it. Users should be able to tell what is real and what is not’

This term encompasses:

• labelling requirements
• additional information (e.g. how AI works, risks, data sources, etc.)

Examples of this include:

• I’m Healthee – your AI chatbot for healthy eating
• You are chatting with an AI here
• AI-based prediction – for guidance only, no guarantee.
• I generated this image using AI.

Important: A single statement can fulfil both requirements at once – labelling is part of transparency. In practice, the distinction is not even that relevant; rather, the key question is: Is it clear, understandable and immediately recognisable to everyone that they are interacting with AI or that content has been generated using AI?

In short: nothing. However, many people believe that if they are the ‘author’ of AI-generated content – due to the high degree of creativity involved – then the obligation to attribute the work no longer applies. But the one has little to do with the other. This is because:

• Copyright law governs who created a work and whether it is protected.
• The labelling requirement (AI Act) governs whether users can recognise that content is AI-generated.
• Degree of creativity ≠ Degree of deception

Both issues run in parallel, but pursue different objectives.

Remember:
Copyright law asks: Who created it?
The AI Act asks: Is anyone being misled?

The AI Act is therefore not concerned with how much human creativity is involved in the result , but rather with how the result appears to the user . And if it appears deceptive, then we must clearly label it as such.

Further information on copyright for AI-generated content will follow shortly.

Of particular relevance is Article 50 of the AI Act (transparency obligations).

The AI Act makes a clear distinction between providers and operators of AI systems.

Specifically, it stipulates that:

• Providers must ensure that AI-generated content is technically identifiable as such (e.g. through labelling or metadata). This includes companies that develop AI systems, such as OpenAI, Google, etc. 
• Operators must disclose when they use AI to generate or manipulate content – particularly in the case of so-called deepfakes. This means that anyone using AI-generated content, for example for marketing purposes, must label the content in certain cases, though not as a general rule. As already explained in the question “Do I have to label all AI-generated content...”.

In short:
Technical labelling is the responsibility of the provider.
For operators, the labelling obligation applies only in exceptional cases regulated by law.

IMPORTANT: The AI Act distinguishes between two levels of labelling:

1. Machine-readable (mandatory for providers)

• Technical or machine-readable labelling in the background, e.g. metadata or watermarks
• detectable by platforms and systems

2. Human-readable (relevant during use)

• visible information for users
• e.g. ‘This image was generated using AI’

Important to note: Machine-readable labelling does not automatically replace visible labelling.

When do the transparency requirements come into force?

From 2 August 2026, the transparency requirements (Art. 50) will become binding.

Are fines to be expected for breaches?

Yes, substantial fines are to be expected for breaches:

• up to €15 million or 3 per cent of global annual turnover for breaches of obligations such as transparency
• in more serious cases, even higher (depending on the category of breach)

At present, there are virtually no relevant court rulings on this matter.However, there are already clear ‘indications of the direction things are heading’

1 Competition law & misleading advertising (very important)

It is already the case that if content is misleading, this is definitely a legal issue (even without the AI Act). Here is a brief example:

• AI-generated people are portrayed as real customers
• Fake scenes are passed off as real

This can therefore already be considered misleading advertising.

Real-life case studies:

PR Council / Industry rulings

An intriguing case (German Public Relations Council, DRPR, AfD Göppingen case, 2024).

• AI-generated individuals were portrayed as real members
• without any indication
• Reprimand for misleading the public and a lack of transparency or labelling

The criticism was that one could have been led to believe the individuals were real. There was no labelling whatsoever, leading to accusations of misleading the public. The DRPR regards this as a clear breach of transparency, truthfulness and the obligation to label. Whilst this is not a court of law, the case clearly illustrates how such matters are assessed.

Another case:

The Tasy AI case (2025)

• Platform offers AI influencers (i.e. fake people)
• It was deliberately advertised without any labelling

A warning was issued regarding misleading content and the lack of labelling

It was also emphasised: the risk of deliberate misleading through AI avatars that appear realistic

These cases demonstrate quite clearly:

The problem is not AI itself, but when AI appears real yet is not labelled as such, this becomes a critical issue.

IMPORTANT! These are not court rulings, but industry decisions. Nevertheless, they are important because they show how such cases are assessed and the direction in which case law is likely to develop. Legal experts state quite unequivocally:

• The labelling requirement applies to content that appears realistic
• Aim: to prevent deception

A brief assessment of how this might develop:

Highly likely: Initial phase (now – 2026)

• Hardly any judgements
• a great deal of uncertainty
• Best practices emerge

Second phase (from 2026–2028)

• first warning letters
• First court cases
• Clarification through case law

That’s when things get really exciting (and stricter).

1. Never use AI-generated content without checking it first
AI should not be seen as a finished end product, but rather as a starting point. Content should always be checked, contextualised and – where necessary – further developed in terms of design or editing.
 

2. Always treat copyright and attribution as separate issues
Whether content is protected by copyright is a different question from whether it must be attributed. Whilst the two issues are linked in practice, they serve different legal purposes.
 

3. The degree of originality is particularly relevant for copyright
The more an AI-generated piece of content is creatively edited, combined or further developed, the more likely it is to constitute an independent human creative work. However, this alone is not decisive for the labelling requirement.
 

4. When it comes to labelling, the impact of the content is what matters most
What matters is not only how content was created, but also how it affects users. The more realistic content appears, the more carefully it should be scrutinised.
 

5. Classify images, videos and audio content with particular care
With visual or audiovisual content in particular, there is a higher risk that something will be perceived as genuine even though it has been artificially generated. This applies especially to realistic-looking people, voices or situations.
 

6. Content that is obviously artificial or creative is usually viewed less critically
Not all AI-generated content is automatically problematic. A stylised illustration, an animated dinosaur or a clearly fictional 3D character should generally be assessed differently from a deceptively realistic depiction of a person.
 

7. With text, human review and accountability are crucial
If a text is created using AI but is editorially checked, revised and endorsed, it is legally classified differently from content published purely by automated means. The human review process is particularly crucial when it comes to texts.
 

8. It is best not to remove existing AI labels
If content is already marked as AI-generated – for example, by platforms, metadata or content credentials – these labels should not be deleted lightly. Transparency is often the better approach here.

9. Always bear in mind platform rules as well as the AI Act
On YouTube, Instagram, TikTok and other platforms, additional guidelines often apply to AI-generated content. Anyone publishing on these platforms should bear in mind not only the legal situation but also the respective platform guidelines.
 

10. When in doubt, opt for transparent communication
Even if there isn’t an explicit labelling requirement in every case, a voluntary disclosure can be sensible. This comes across as professional, builds trust and demonstrates a responsible approach to AI.

Is the transparency requirement a nuisance or a sensible measure?

To be honest: I think it’s sensible.

In my view, clear rules are particularly needed when it comes to deepfakes, manipulated content or deceptively realistic AI-generated material. Not only to protect users, but also to raise awareness of what is real — and what isn’t. If a machine-readable label were also made mandatory, this could, in the long term, help to track content more effectively. And that is anything but a bad thing, particularly when it comes to problematic or deliberately misleading content.

Fundamentally, I do not see labelling as a disadvantage either. On the contrary: a transparent disclosure can actually build trust. Particularly in the case of text, a voluntary transparency notice comes across as professional rather than problematic these days. Along the lines of: ‘We use modern tools – but we’re open about it.’

Will everything be monitored now?

Probably not.

I can hardly imagine that every leaflet, every social media post or every website will be manually checked in future. Realistically speaking, the focus will initially be on large platforms, well-known providers and major market players. In other words, the usual suspects who actually have real reach and impact.

Nevertheless, I wouldn’t take the issue lightly. After all, things can always take a critical turn if someone reports a specific breach or takes issue with a particular case. And, unfortunately, we know how it goes: whenever new rules emerge, it usually doesn’t take long before some law firms spot a new business model in them. In other words: warning letters never sleep.

Just how specific is the AI Act, actually?

In my view, the AI Act primarily sets out the framework, but it does not specify down to the last detail how everything is to be implemented in practice.

Legally speaking, quite a few things have been deliberately left open. Terms such as ‘clear’, ‘recognisable’ or ‘reasonable’ leave room for interpretation. On the one hand, this makes sense because technology and platforms change rapidly. On the other hand, of course, this very fact also creates uncertainty in practice. In short: the AI Act provides the playing field rather than the complete set of rules.

Why transparency is still important

Even setting aside legal obligations, I believe transparency is important. Not just out of legal prudence, but also as a matter of principle.

In my view, anyone working with AI shouldn’t have to hide — but neither should they pretend that everything was created purely by humans. Particularly in communication, marketing and creative processes, an open approach to AI can be a genuine sign of trust. And quite honestly: a brief mention is usually far more pleasant than a major row later on.

Deepfakes? That’s where the fun stops.

When it comes to deepfakes, my personal stance is quite clear: I very much hope that there will be consistent monitoring and enforcement.

As soon as content is deliberately misleading, misrepresents people or manipulates reality, a line has been crossed for me. It’s no longer about creative experimentation or clever tools, but about responsibility.

Looking ahead

I also find it fascinating to consider how this will develop technically. I can well imagine that, in future, users will want to have greater control over whether they wish to see AI-generated content at all. Perhaps at some point there really will be a sort of ‘AI blocker’ – in other words, what the ad blocker once was for adverts, but for synthetic content. That’s still a long way off, but the idea isn’t entirely far-fetched.

For me, all of this underscores one thing above all else:
AI is a tool — people remain the decisive factor.

The technology is capable of a great deal. But how we use it, how transparently we handle it and what standards we set for ourselves – that is, and always will be, a human decision. And perhaps that is the most important rule of all: not everything that is technically possible is necessarily a good idea from a communication perspective.

The content is based on the AI Act, a specialist presentation by solicitor Notash Taheri and an article by E-Recht24.
This is not intended to be exhaustive – and is definitely not legal advice.