The European Commission (EC) is facing a paradox as it develops its Code of Practice on Content Labeling for content generated by artificial intelligence (AI), central to its broader AI regulation efforts. This initiative, championed by the EU’s top leadership, creates a double standard: it mandates strict transparency and labeling requirements for AI-generated content from the private sector, while the EC itself actively uses these technologies for internal purposes without applying the same rigorous standards to its own output.
Key Aspects of EU AI Regulation and Content Labeling
The political impetus behind mandatory labeling for synthetic content came from Věra Jourová, the European Commission Vice-President for Values and Transparency, whose mandate expired in 2024. In June 2023, she urged tech giants to swiftly implement measures to combat disinformation generated by neural networks.
The newly established European AI Office has taken on the task of overseeing compliance with these regulations. The primary goal of regulation under the labeling code is to protect citizens from disinformation and fraud. Responsibility falls on AI system providers and operators, who must implement machine-readable labels that are resistant to removal. This is a technically complex task, and its inherent imperfections make achieving these requirements difficult.
To ensure transparency, the European Commission encourages developers to use technologies that allow for reliable embedding of content origin information:
☛ Digital Watermarking: A hidden, human-imperceptible code embedded directly into the file (audio, video, or image). It is difficult to remove without damaging the content.
☛ Cryptographic Attestation: A more robust method where a digital signature (certificate) is attached to the file. This signature tracks the content’s entire ‘journey,’ verifying when and where it was created with AI.
☛ Embedded Metadata: Standardized information added to the technical data of a file, but easily lost during editing.
Key challenges for the ‘European Fortress’
The labeling requirements highlight several critical issues that the European Commission must address for successful implementation. First, there is the challenge of international compatibility of standards. Creating EU-specific labeling standards could force global platforms to develop separate, costly versions of their AI models. Furthermore, if EU standards (e.g., for digital watermarks) are incompatible with approaches from the US (e.g., C2PA) or China, European companies will face significant operational overheads.
Second, there is the issue of protecting Open Source. Regulators need to clearly define responsibility for content created with AI. Today, most breakthrough AI models are released as open source. If a developer releases a model that can create deepfakes but lacks built-in labeling mechanisms, who is accountable?
Excessive pressure on developers could inadvertently stifle innovation in this critical sector.
Third, there is the illusion of ‘irremovable’ labeling. Technically, it simply does not exist. For example, a label could be removed using another AI model specifically trained to erase watermarks. This merely reduces the risk, rather than eliminating it.
The Logical Discrepancy
The main logical discrepancy lies in the fact that the European Commission, which itself uses AI for internal purposes (document analysis, translations) and external publications, imposes such stringent requirements on the external market. The EC justifies this by invoking the principle of risk differentiation.
According to the Commission, internal AI use to enhance the efficiency of the state apparatus, where final documents are human-verified and published under official responsibility, presents minimal risk. In contrast, widespread market use of AI carries a high risk of disinformation and fraud, necessitating immediate regulation.
However, this approach creates bureaucratic pressure and risks slowing down the development of the European technology sector.