AI Labelling Obligation: 21 Questions, 21 Short Answers

Since 2 August 2026, the labelling obligation for AI content under Article 50 of the EU AI Act has applied. Contrary to what many headlines suggest, not every use of AI has to be labelled. Advertising copy, product descriptions, translations and normally edited photos are generally not covered at all. Three things essentially do have to be labelled: chatbots that talk to people, deepfakes – convincingly realistic depictions of real people, places or events – and certain AI texts on matters of public interest that nobody checked before publication. Anyone who reads their AI texts themselves and takes responsibility for the content is exempt from the text obligation anyway.

Below are the 21 questions most frequently asked about this – each answered in a few sentences.

Contents

1. Do I have to label my AI texts and AI images at all?
2. What happens if I do not label?
3. Is a wave of warning letters coming?
4. Do I have to label my old content retroactively?
5. Do I have to label AI images – including edited product photos?
6. Does this apply to me as a sole trader or small business?
7. How and where do I label? Which sentence is enough?
8. Does it still count as AI content if I revise the text myself?
9. Do I have to label product descriptions in my shop?
10. Do I have to label AI in invoices, quotes or emails?
11. Do I have to label my chatbot?
12. Do I have to label social media posts?
13. What actually is a deepfake?
14. Do I have to label supplier images if I do not even know whether they come from an AI?
15. How do I label on Amazon or eBay, where text in images is prohibited?
16. Does the obligation also apply in pure B2B?
17. Is the AI checkbox on YouTube enough, or does a notice have to go into the video?
18. Are AI translations and spelling corrections affected?
19. Is a real model in front of an AI background a problem?
20. Who actually monitors this?
21. Wasn't the date postponed after all?

1. Do I have to label my AI texts and AI images at all?

In most cases, no. The obligation is not triggered by the fact that you used an AI, but by the possibility that someone could be deceived. What is mainly covered are chatbots, deepfakes and unchecked texts on matters of public interest. An advertising text written with ChatGPT, or an AI image that is clearly decorative, is normally not covered.

2. What happens if I do not label?

Breaches of the transparency obligations are covered by Article 99(4)(g) of the AI Act: fines of up to 15 million euros or 3 per cent of worldwide annual turnover are provided for, whichever is higher. For small and medium-sized enterprises the regulation expressly reverses this – under Article 99(6), the lower of the two amounts applies to them. How high a fine turns out in an individual case depends, under Article 99(7), among other things on the nature, gravity and duration of the breach, on intent or negligence, and on the size of the company. How the authorities will handle this in practice cannot yet be said – the rules have only been in force for a few days.

3. Is a wave of warning letters coming?

We cannot answer that. The AI Act itself does not govern warning letters, but fines imposed by authorities. Whether missing labelling can additionally be challenged under competition law depends on how the courts classify the provision – we are not aware of any judgments on this. Several law firms have commented on it since the beginning of 2026 and reach differing assessments; we cannot judge which of them will prevail.

4. Do I have to label my old content retroactively?

This is the question that makes most people break into a sweat – and we cannot answer it with certainty. The AI Act contains no express transitional rule for content that has already been published; the transitional provisions in Article 111 concern high-risk systems and providers, not published texts or images. We are not aware of any official clarification, and law firms and chambers of commerce have published differing views on it, which we do not evaluate.

Regardless of this, it remains the case that advertising copy and normal product photos are not covered by the labelling obligation at all – so nothing changes for them, however old they are. The open question only concerns content that falls under Article 50 in the first place, in practice mainly deepfakes.

There is a genuine grace period in only one place, and it concerns manufacturers rather than you: providers of AI systems that generate synthetic image, audio or text content must mark their outputs in a machine-readable format. For systems already on the market before 2 August 2026, the Digital Omnibus Regulation added a deadline of 2 December 2026 for this (new Article 111(4) of the AI Act). As a user of such a system, this obligation does not apply to you.

5. Do I have to label AI images – including edited product photos?

For images, visible labelling is only required for deepfakes. A purely artificially generated mood image that does not feign anything real therefore generally does not have to be labelled. Ordinary image editing – cutting out, brightening, retouching, including with AI functions in Photoshop – also remains outside the scope, as long as the picture still shows what actually exists. This does not affect the fact that misleading product images may already be open to challenge under competition law – but that is a different legal basis from the AI Act.

6. Does this apply to me as a sole trader or small business?

Yes, there is no turnover or headcount threshold. The obligation attaches to your role: anyone who supplies an AI system is a provider, anyone who uses it professionally is a deployer – and you are already a deployer if you use ChatGPT for your website. The reassurance lies not in your size, but in the fact that the typical content of small businesses almost never falls under the obligation.

7. How and where do I label? Which sentence is enough?

Article 50(5) requires the information to be provided clearly and distinguishably "at the latest at the time of the first interaction" and to meet accessibility requirements. Hidden away in the legal notice or the privacy policy does not satisfy that. A brief note is enough, for example "This image was created using artificial intelligence" directly on or below the picture, or "You are chatting with an AI assistant" at the start of the chat. There is no legally prescribed wording and no official logo.

8. Does it still count as AI content if I revise the text myself?

This is the most important relief in the whole law. For texts, the labelling obligation does not apply if the text was reviewed and approved by a human before publication and a natural or legal person holds editorial responsibility for it. So anyone who reads through their AI drafts, corrects them and publishes them deliberately does not have to label anything. What is meant is genuine review, not a fleeting glance before clicking "publish".

9. Do I have to label product descriptions in my shop?

No, classic advertising copy and product descriptions are not covered by the text obligation. It only applies to texts published in order to inform the public on matters of public interest – meaning news items and political contributions, not the description of a garden hose. And for borderline texts there is additionally the exemption for editorially reviewed content from question 8.

10. Do I have to label AI in invoices, quotes or emails?

No. Invoices, quotes, order confirmations and reminders are addressed to a specific recipient and do not inform the public – they fall under neither the deepfake rule nor the text rule. Templates and text modules generated for you by software make no difference either. So if you have document texts or cover letters generated with assistance in office1.cloud, no labelling requirement arises on the invoice.

11. Do I have to label my chatbot?

This deserves a closer look, because under Article 50(1) the obligation expressly falls on the provider of the system: they must build the chatbot so that people are informed that they are interacting with an AI system. Anyone who buys a ready-made chatbot and uses it unchanged is initially only a deployer – in which case the obligation lies with the manufacturer. You do slip into the provider role, however, as soon as you offer the system under your own name or substantially modify it.

The obligation does not apply in any case where it is obvious from the circumstances to a reasonably attentive person that they are talking to a machine. Because the dividing line is hard to draw in everyday practice and a single sentence costs nothing, a clear notice at the start of the chat is the practical standard route.

12. Do I have to label social media posts?

For the post itself, the same rules apply as everywhere: no deepfake, no unchecked text on a matter of public interest, no obligation. An AI-generated image for an Instagram post is therefore usually unproblematic. Independently of that, the platforms themselves often require labelling via their own toggles. Those are house rules, not a legal obligation; what happens if you breach them is up to the platform.

13. What actually is a deepfake?

The term is defined in the regulation itself (Article 3(60)): AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic. What matters, therefore, is the risk of confusion with reality. An obviously drawn cartoon robot is not a deepfake; a convincingly realistic video of your mayor is. For artistic or satirical works, a notice that does not spoil the work is enough.

14. Do I have to label supplier images if I do not even know whether they come from an AI?

A great many retailers ask this, and we do not have a clear answer. The regulation does not say what applies if you do not know the origin of an image at all; nor are we aware of any official statement on it.

What can be said from the text: the disclosure obligation for deployers under Article 50(4) only bites for deepfakes. Normal catalogue and product images are not deepfakes, so the question does not arise there. It remains open for material showing real people or situations that never existed in that form – such as invented application scenes with apparently real people. Whether and to what extent you can rely on your supplier's statements in such cases is not regulated.

15. How do I label on Amazon or eBay, where text in images is prohibited?

Many people ask this too, and we are not aware of any official model solution. Two things can at least be established: for normal product images the question does not arise, because they are not deepfakes. And the regulation does not prescribe any particular form – Article 50(5) only requires the information to be clear and distinguishable and available at the latest on first contact with the content.

Where the notice has to appear when text in the image is prohibited by the platform is therefore not answered by the regulation. Whether a statement at the start of the item description is sufficient has, to our knowledge, been settled neither by an authority nor by a court.

16. Does the obligation also apply in pure B2B?

Yes. Article 50 does not distinguish between consumers and business customers; it speaks of natural persons. The same therefore applies to a chatbot in a portal for business customers as in a consumer shop.

17. Is the AI checkbox on YouTube enough, or does a notice have to go into the video?

That is open. The checkbox on YouTube first of all satisfies a platform rule. Whether the notice YouTube generates from it also meets the requirement in Article 50(5) – clear, distinguishable and at the latest on first contact with the content – has, to our knowledge, been decided neither by an authority nor by a court.

18. Are AI translations and spelling corrections affected?

No. For you as a user, the text obligation only bites for texts on matters of public interest – a translation or a spelling correction is not that. The regulation itself also excludes systems that merely "perform an assistive function for standard editing" or do not substantially alter the input (Article 50(2)). Internal drafts that are never published trigger nothing anyway.

19. Is a real model in front of an AI background a problem?

The benchmark is the deepfake definition: does the image resemble existing persons, places or events so closely that it would falsely appear to be authentic? With an obviously artificial backdrop that is remote; with a specific, existing place where the person has never been, it is closer. Exactly where the line runs has not yet been decided.

20. Who actually monitors this?

In Germany this is governed by the AI Market Surveillance and Innovation Promotion Act, promulgated on 28 July 2026 and in force since 29 July 2026. A dedicated coordination and competence centre has been set up for it at the Federal Network Agency; the agency is the competent market surveillance authority, unless the federal states designate their own bodies. The focus of supervision is on high-risk applications, for example in critical infrastructure or law enforcement – not on product photos in an online shop.

21. Wasn't the date postponed after all?

Something else was postponed. The Digital Omnibus Regulation of 8 July 2026 pushed back the obligations for high-risk AI systems, partly to 2 December 2027 and partly to 2 August 2028. It changed only a marginal paragraph on codes of practice in Article 50 – the transparency obligations themselves have applied unchanged since 2 August 2026, as provided for by Article 113 of the AI Act. Anyone relying on an extension is relying on the wrong report.

What to watch next

Several points remain unresolved: how far "matters of public interest" reaches, how already published content is to be treated, how labelling should work on marketplaces, and whether missing labelling can be the subject of a warning letter. So far there are neither judgments nor official determinations on these that we could cite.

If you want to follow developments, the official information points are the German Federal Network Agency and the AI Act Service Desk of the European Commission. Both answer questions on the AI Act and update their information continuously.

Status of this article: 3 August 2026. It is based on the wording of the AI Act (EU) 2024/1689, the Digital Omnibus Regulation (EU) 2026/1744 and the German AI Market Surveillance and Innovation Promotion Act. This article provides general information and does not replace individual legal advice; in case of doubt, a lawyer can clarify your specific situation.