Mandatory AI Labeling Takes Effect – Traceability Remains an Open Question

The next phase of implementation of the AI Regulation has been in effect in the EU since Sunday: Under Article 50, companies must label AI chatbots as well as text, images, and videos that have been created or edited using artificial intelligence—provided they provide information on matters of public interest for commercial purposes. What sounds like a clear requirement raises numerous questions of interpretation in practice: Exceptions include, among other things, obviously unrealistic content and texts generated entirely by AI that are subsequently reviewed by a human—conversely, if a text written by a human is edited using AI, the requirement applies immediately. According to competition law experts, these exceptions alone are likely to lead to disputes. Violations are subject to fines of up to 15 million euros or three percent of global annual revenue; the operators of the systems within the company itself are held responsible.

The Handelsblatt has analyzed the new labeling requirement and its practical pitfalls for companies in a separate article, in which it also features Heiko Beier, who develops AI solutions at moresophy that make it possible to understand how the systems work. In his view, Article 50 merely answers the question of whether AI was used—but not how a result was arrived at: which models, which training data, and which decision-making logic are behind it. That, in particular, could make all the difference in disputes. Instead of a mere labeling requirement, the focus should be more on traceability and accountability for the result. In Beier’s view, a label confirming that content is “100 percent human-made” would be more valuable in the future.

The immediate need for action, therefore, does not lie in even more labeling, but in processes: Companies should establish traceability based on their own data—for example, when making discretionary decisions in customer service—and in doing so, hold both employees and AI systems to the same standard of care in training and oversight. With generative models, the limit to explainability lies in the lack of access to the sources of the model’s knowledge—but transparency can be achieved within one’s own data sources if AI systems are consistently controlled based on their own logic rather than through pre-modeled workflows applied to isolated raw data. According to Beier, the music industry already illustrates just how real this problem is: there, a simple label does not resolve the blurred lines between human and machine authorship, with direct consequences for licensing and compensation issues.