Where the Obligations Arise

The most relevant obligations under Article 50 of the AI Act fall into three categories: (i) interactive AI systems, (ii) synthetic content, and (iii) deep fakes.

One distinction matters upfront: the obligations for interactive AI systems and synthetic content apply to providers, whereas the deep fake obligations apply to deployers. The same company may be a provider in one context and a deployer in another.

Interactive AI Systems (Article 50 para. 1 AI Act)

Providers of AI systems intended to interact with natural persons must ensure that users are informed that they are interacting with AI, unless this is obvious. Typical examples include chatbots and voicebots. As these systems become increasingly sophisticated, the line between human and AI interaction is no longer always clear. This is precisely why the obligation exists.

The practical challenge lies in building disclosures into the system design without creating friction at the point of user interaction. Solutions may be text or audio notices. The disclosure must be provided no later than the first interaction. There is no prescribed technical form, but what matters is that the information reaches the user effectively, including vulnerable groups.

Synthetic Content (Article 50 para. 2 AI Act)

Providers of AI systems that generate synthetic audio, image, video, or text content must ensure that the output is marked in a machine-readable format and remains detectable as AI-generated or manipulated. This is not a labelling exercise that can be addressed after deployment. It requires technical solutions embedded in the system itself, ensuring that marking is effective, interoperable, robust, and reliable.

Which solution meets all these requirements?

Implementation is where this becomes difficult. The European Commission has commissioned three studies on technical solutions to support its Code of Practice on marking and labelling AI-generated content. Possible approaches include cryptographic signatures, metadata, content credentials, watermarks, logging, and fingerprinting methods.

The conclusion is consistent: no single technical approach currently meets all the AI Act’s requirements. Organizations are likely to need a layered, risk-based approach combining multiple methods. Those that defer the question until a definitive best practice standard emerges risk running out of time.

Deep Fakes (Article 50 para. 4 AI Act)

Deployers of AI systems that generate or manipulate audio, image, or video content that qualifies as deep fake must disclose that the content is artificially generated or manipulated. A deep fake is AI-generated or AI-manipulated content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.

Adopting a broad interpretation in light of the protective purpose of Article 50 para. 4 of the AI Act, a deep fake may already exist where a viewer could reasonably believe that the content is realistic and truthful. For content designed to achieve a particularly realistic effect, transparency obligations can, therefore, arise quickly. This applies regardless of intent to deceive.

The point is simple: users must be able to understand whether what they see or hear is authentic.

Disclosure of a deep fake can be achieved through a symbol or clear notice. The EU Commission has recently published a set of icons consisting of a basic icon “AI” and the supplements “AI GENERATED” and “AI MODIFIED”. The use of these icons is optional. That optionality, however, does not extend to the disclosure obligation itself. Organizations remain responsible for meeting the requirements under Article 50 of the AI Act.

A common misconception is to treat creative or artistic use as an automatic exception. Creativity is not a blank check. Article 50 para. 4 of the AI Act provides for reduced transparency obligations for artistic and creative content. However, this exception must be interpreted narrowly. It applies only where the artistic or creative character is clearly recognizable and the context is not primarily functional or commercial in nature. Not every creative idea is, therefore, a legally creative exception.

Transparency as a Governance Requirement, Not a Compliance Checkbox

Transparency is not a mere labelling exercise. It is a governance issue requiring clear ownership across technology, legal, compliance, marketing, and risk management. In many organizations, the honest answer is that no one has yet taken ownership of this matter. This is where the risk begins.

The risk of non-compliance is real: fines may reach up to EUR 15 million or 3% of global annual turnover, whichever is higher. That makes transparency a board-level concern, not a back-office compliance task.

Why Implementation Cannot Wait

The obligations apply from 2 August 2026. For AI systems generating synthetic content already placed on the market before that date, the timeline has been extended to 2 December 2026 following the European Parliament’s vote of 16 June 2026, pending formal adoption. However, that extension applies to a specific category of systems only.

Thus, the urgency to act now is not reduced. Building embedded transparency mechanisms and implementing the related processes cannot be completed in a few weeks. In addition, ongoing monitoring of best practices for technical solutions is required.

Next steps

Companies should, therefore, clarify their role as provider, deployer, or both, map their AI use cases against the categories under Article 50 of the AI Act and embed the requirements into system design from the start. The cost of retrofitting is higher than the cost of getting it right.