In Part 1 of this series, we examined commentary by the Commission to assist in navigating concepts of Article 50 that were initially unclear.
We also discussed the Commission’s interpretation of Articles 50(1) and 50(2), including disclosure obligations for AI-human interactions, the narrow scope of the “obviousness” exception, and the marking requirements for AI-generated or manipulated content. We also considered the treatment of GPAI, open-source and agentic AI systems, and the limited exclusions identified in the Guidelines.
Part 2 of this series examines the remaining transparency obligations under Article 50, focusing on emotion recognition and biometric categorisation systems under Article 50(3), alongside the disclosure requirements for deepfakes and AI-generated content informing matters of public interest under Article 50(4). It also considers the practical operation of the artistic and satirical carve-outs, retrospective labelling expectations, and the emerging Code of Practice in demonstrating compliance.
Transparency in Emotion Recognition and Biometric Systems
Article 50(3) focuses on transparency for systems that detect emotions (defined in Article 3(39)) or categorise people using biometric data (defined in Article 3(40)). Emotion recognition systems are always high-risk under Annex III or prohibited under Article (5)(1)(f) in workplace and education settings, and the Guidelines interpret Article 50(3) applying in conjunction with the rules under Annex III.
Article 50(3) covers all forms of biometric data processing AI systems whether high-risk or not.
While the Act does not prescribe exactly how to inform users of interaction with the above-mentioned systems, Article 50(5) requires that disclosures be clear, visible, and accessible to all users interacting with the system. For example, a video game using emotion tracking could display a prominent pop‑up notice before play begins, and where children are involved, the message should be particularly evident. The disclosure should be addressed at the design stage.
Deep fakes and Disclosure Obligations
Article 50(4) Deployers must inform users of AI systems that generate or manipulate deep fake or text published to inform the public on matters of public interest that the content was AI generated or manipulated. Disclosure methods here should be understood by natural persons without the need for detection tools, as required under Article 50(2).
Under Article 3(60) of the Act, a deep fake is AI‑generated or manipulated image, audio or video content that closely resembles realistic persons, objects, places, entities or events and would falsely appear authentic or truthful to those exposed to it. The Guidelines examine the following factors:
- The resemblance must be appreciable but need not be identical to the subject. This Guidelines note that this would be determined on a case-by case basis.
- The AI generated content should resemble subjects that could realistically exist. An example provided is deep fakes portraying mythical creatures in real-world scenarios. This would not fall under Article 3(60). By contrast, simulated persons resembling other persons that could exist may constitute a deepfake.
- The content must be capable of misleading people about its authenticity or truthfulness, regardless of the deployer’s intent. The Guidelines state that this criterion is judged against the actual composition of the audience and not a hypothetical person. Deployers should note that children, elderly people and individuals with lower AI literacy will be viewing these deep fakes.
The Guidelines provide that small and technical AI edits to existing content (e.g. adjusting lighting, fixing colours, reducing noise, or improving sound) normally do not make the content a deepfake, because they don’t significantly affect how truthful it appears. However, this will be judged on a case-by-case analysis. More substantial edits, especially those that change meaning or context (such as altering images in journalism), may affect authenticity.
Article 50(4) of the Act provides for a reduced transparency obligation where deep fakes form part of works or programmes that are clearly artistic, creative, satirical, fictional, or similar in nature. Article 50(4) is satisfied if the deployer discloses the use of deep‑fake content in an appropriate manner that does not interfere with the presentation or enjoyment of the work. Importantly the Guidelines note that the work must ”evidently” fall within one of the five listed works above, implying that any ambiguity will exclude it from the exception. This carve out is without prejudice to the rights and freedoms (personal or intellectual property rights, for example) of third parties.
The Guidelines confirm that AI-generated or manipulated content falling under Article 50(4) already circulating prior to 2 August 2026 does not need to be retrospectively labelled. However, deployers are encouraged to voluntarily apply labelling or marking to such content where appropriate. The same applies to content falling under Article 50(2) as discussed in Part 2.
Approved Code of Practice
The Guidelines explain that providers and deployers can demonstrate compliance with Article 50 by adhering to an approved Code of Practice, which simplifies oversight and supports compliance with the AI Act. This Code is being drafted by eligible stakeholders who answered a public call made by the AI Office and is expected to be finalised and published in June. The Guidelines state that those who do not follow this Code must prove compliance through other means and may face greater scrutiny, while compliance may mitigate potential fines.
Takeaway
Transparency obligations under Article 50 of the AI Act are likely to be interpreted broadly, with narrow exceptions and an expansive approach to disclosure and content-marking obligations for providers and deployers. The Guidelines signal increased regulatory focus on deepfakes, agentic AI, biometric systems and synthetic content ahead of the compliance deadline on 2 August 2026.
Contributed by Isaac Kelly
