AI labelling obligation: 18 questions answered

Article 50 has applied since 2 August 2026. The Digital Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and did not move that date. It added one narrow transition in a new Article 111(4): providers of generative AI systems placed on the market before 2 August 2026 have until 2 December 2026 to add the machine-readable marking required by Article 50(2).

That grace period does not help deployers. Labels for deep fakes and for AI text on matters of public interest, and notices for chatbots, have been required since 2 August 2026. The Commission's guidelines on Article 50 confirm that a system which is partly interactive and partly generative benefits from the transition only for the marking duty, while its disclosure to users had to be in place from 2 August 2026.

Article 50 splits the duties. Providers, meaning those who develop an AI system and offer it under their own name or trademark, must build chatbots that disclose themselves (Article 50(1)) and mark generated content in a machine-readable way (Article 50(2)). Deployers, meaning businesses and other organisations that use an AI system under their authority, must label deep fakes and certain AI texts (Article 50(4)) and inform people exposed to emotion recognition or biometric categorisation (Article 50(3)).

Most businesses that create images or texts with a generative AI tool are deployers. Freelancers, influencers and agencies act professionally too: the Commission's guidelines treat any activity with a regular economic benefit, and any business, trade or freelance activity, as professional. Employees are not separate deployers; their employer is. According to the guidelines, a company that merely commissions an agency, without deciding whether and how the agency uses AI, is not a deployer.

Not every AI output. For images, audio and video, the deployer duty covers deep fakes: content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful (Article 3(60)). A photorealistic fake of a real person or event needs a label. The guidelines name an AI image of a sphinx flying over the Eiffel Tower as an example that is not a deep fake.

For text, the duty covers AI-generated or manipulated text published to inform the public on matters of public interest, such as politics, public health, public security, the environment, consumer safety or economic and scientific developments. Fantasy novels are outside, and so are ordinary advertising texts and product descriptions without claims on topics such as health, consumer safety or sustainability. Deep fakes that are part of an evidently artistic, satirical or fictional work need only a disclosure that does not spoil the work (Article 50(4)).

Separately, providers must mark generated audio, images, video and text in a machine-readable way, with exceptions for standard editing such as grammar correction or minor colour adjustments (Article 50(2)). Labelling more than the law requires is allowed and can build trust. Our labelling guide includes a short questionnaire that tells you whether a label is required for a given piece of content.

Only if the text is published to inform the public on matters of public interest. A company blog post about a new product feature usually is not; a post on health effects, energy policy or a new law can be. The guidelines count a text as published when an indeterminate, fairly large number of people can access it, even behind a paywall, while internal communications and private correspondence are not published.

Even then, no label is needed if the text has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication (Article 50(4)). The guidelines set a real bar: a knowledgeable person must examine the substance, including fact-checking. Spell-checks, automated reviews, a mere editorial policy or cursory approval are not enough, and substantive AI changes after sign-off cancel the exception. The person or function responsible should be publicly findable, for example on a legal information page.

It depends on whether the image could mislead. The guidelines list an AI-generated product image in advertising or packaging that makes the product look different, more appealing or better than in real life as a deep fake that needs a label. By contrast, a real product shown against an AI-generated background is not a deep fake, as long as the ad does not mislead about how the product looks, what it does or how it is used.

Colour correction, background extensions, re-scaling and arrangements of existing products are, in the guidelines' words, likely to have only a minor impact on authenticity. Realistic AI people presenting your product are a different matter: a realistic synthetic influencer testing a real product is one of the guidelines' deep fake examples. A label also does not make misleading advertising lawful, so images still have to show the product fairly.

Breaches of Article 50 can be fined up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher (Article 99(4)(g)). For SMEs and start-ups the lower of the two applies (Article 99(6)), and since the Digital Omnibus the same cap applies to small mid-caps. National market surveillance authorities enforce the rules, on their own initiative or after a complaint, and weigh all circumstances, including size, intent and cooperation (Article 99(7)).

In Germany, a second question is whether competitors or associations can send a cease-and-desist letter (Abmahnung) under the Unfair Competition Act (UWG) over a missing AI label. Some German law firms argue that Article 50 may be a market conduct rule under section 3a UWG, or that leaving out the AI origin may be misleading under sections 5 or 5a UWG. As far as we know, the Federal Court of Justice has not ruled on this, so the question is open and depends on the case.