Transparency obligations for machines or people

Article 50(2) of the EU AI Act requires providers of AI systems to ensure that their outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. This provision covers the majority of today’s web, increasingly flooded with AI-generated and AI-manipulated content.

“The beneficiaries of this transparency obligation are therefore all natural persons likely to be exposed to the AI-generated or manipulated content”, states the European Commission in its Draft Guidelines on the implementation of the transparency obligations[1].

There is, however, no express obligation to make that marking human-perceptible. The Draft Guidelines emphasise this point. This raises the central question: how often do you check for AI origin in a detection tool?

The Draft appears to suggest that you, as a beneficiary, would do that often. The same approach is reflected in the parallel Draft Code of Practice on marking and labelling of AI-generated content[2].

But could there be another practical way?

First, the Preamble of the Act does not restrict transparency techniques to invisible technical means. It refers to “watermarks or other techniques, as may be appropriate”. For the avoidance of doubt, the Draft Code chooses the narrower language of “imperceptible watermarking techniques”.

Second, Article 50(5) states that the information referred to in paragraphs 1 to 4 shall be provided to natural persons in a clear and distinguishable manner at the latest at the time of first interaction or exposure. The Draft Guidelines clarify that information is clear where it is noticeable and easy to understand.

Applied to Article 50(2), the Commission suggests that this obligation is fulfilled through detection tools. What it does not explain is how this mechanism satisfies the timing requirement: “at the latest at the time of the first interaction or exposure”.

Compared with other Article 50 obligations, which require human-perceptible disclosures, Article 50(2) is treated differently.

This raises a difficult question: why does the Act leave a broad group of natural persons likely to be exposed to AI-generated or manipulated content without clear and distinguishable information when they are exposed to it? Perhaps the Act does not, and a different understanding of the scope of the obligations should be implemented in practice.

Article 50(2) could focus on the provision of essential transparency information in a machine-readable format, not to weaken human transparency, but to provide provenance and technical information that would be too detailed or intrusive to display directly on the content itself.

Article 50(5), in turn, could serve as a general guarantee of human transparency rights: whatever technical obligations apply under Article 50, the AI character of the content must be made human-perceptible at the moment when it matters. This would give the person a choice to check further information about the content encountered, specifically by using the means provided under Article 50(2).

Under this reading, Articles 50(2) and 50(5) can coexist without diluting each other or weakening the guarantees of human transparency.