What it’s about
Art. 50 of the AI Act sets out four separate transparency obligations: labelling of AI interaction (para. 1), machine-readable marking of synthetic content (para. 2), information duties for emotion recognition and biometric categorisation (para. 3), and disclosure obligations for deepfakes and AI-generated text on matters of public interest (para. 4). These duties apply regardless of the system’s risk class – even a simple customer-service chatbot falls within scope.
Chatbots: duty to disclose interaction
Under Art. 50(1), providers must design systems intended for direct interaction with natural persons so that users are aware they are communicating with an AI. The exception applies only where this is “obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use” – a high bar that should rarely be relied on in practice. A typical gap: chatbots presenting themselves as a “team” or under a person’s name, without making the AI nature clear at the first point of interaction. Art. 50(5) requires this information to be given “at the latest at the time of the first interaction”, in a clear and accessible manner – a note buried in the footer or in the terms and conditions is not sufficient.
Synthetic content: machine-readable labelling
Paragraph 2 concerns providers of systems that generate synthetic audio, image, video or text content – expressly including general-purpose AI systems. It requires marking “in a machine-readable format” that makes the output identifiable as artificially generated or manipulated. The legislator requires the solutions to be effective, interoperable, robust and reliable “as far as technically feasible”, taking the state of the art into account – leaving some room for manoeuvre, but also creating a need to justify why no solution has been implemented at all. Exempted are assistive editing functions that do not substantially alter input data (standard spell-checking, for instance), as well as applications used for law enforcement purposes. Typical gap: companies use third-party GPAI models and assume the labelling duty rests with the model provider – but this does not automatically discharge them once they make their own adaptations or further processing.
Deepfakes and text of public interest
Paragraph 4 is addressed to deployers, not providers. Anyone generating or manipulating image, audio or video content that constitutes a deepfake must disclose that the content is artificially generated or manipulated. For artistic, satirical or fictional works, disclosure that does not impair the enjoyment of the work is sufficient. For text published to inform the public on matters of public interest, a comparable disclosure duty applies – with one important exception: it does not apply where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility. For editorial teams and content departments producing AI-assisted articles, this exception is central – but it must actually be applied in practice, not merely exist on paper.
Timeline and what to do now
Under the “Digital Omnibus on AI” (adopted, not yet published in the Official Journal), the transparency duties under Art. 50 apply from 2 August 2026. For systems already placed on the market before that date, a grace period runs until 2 December 2026. This is not a blank cheque to wait: anyone currently operating chatbots, text or image generators should now check where in the user flow labelling is missing, which exceptions genuinely apply, and whether markings are technically machine-readable rather than merely visual. The most common gap in practice: the duty is treated as a pure UX matter and left undocumented – so if a supervisory authority asks, there is no evidence of why an exception applies or how the labelling works technically.
Whether your systems fall under Art. 50 and which of the four duties are relevant to you can be clarified in a few minutes using the free risk check at /einstufung.