Key facts
- The EU AI Act requires clear labeling of deepfakes and AI-generated content.
- Generative AI system providers must ensure content is detectable.
- Deepfakes imitating real persons face stricter labeling rules.
- Content for artistic, creative, satirical, or fictional purposes is exempt.
- Technical limitations may impact the effectiveness of enforcement.
The European Union's AI Act, effective this week, mandates that deepfakes and other AI-generated content must be clearly identifiable as artificial. This legislation aims to enhance transparency around synthetic media by requiring companies developing or using generative AI systems to implement detection mechanisms and labeling.
The rules are particularly stringent for deepfakes—realistic AI-generated or manipulated images, audio, or videos that could mislead individuals into believing fabricated statements or actions were real. While the EU acknowledges that not all synthetic content is deceptive, with exceptions for artistic, satirical, or fictional purposes, the potential for harm through manipulation and disinformation is a significant concern.
Examples of deepfake misuse include a viral AI-generated image of Pope Francis in a puffer jacket, which was largely understood as a joke, and the non-consensual spread of explicit AI-generated images of pop star Taylor Swift. During the Russia-Ukraine war, a fake video of President Volodymyr Zelenskyy urging surrender demonstrated the potential for weaponizing synthetic media.
To address these challenges, the European Commission has established a voluntary code of practice. This framework encourages companies to use machine-readable markings and visible labels for deepfakes. The initiative also emphasizes the need for collaboration between AI developers, social media platforms, civil society organizations, and fact-checkers to develop effective detection mechanisms.
However, the effectiveness of these regulations faces several hurdles. The global nature of online content means deepfakes created outside the EU can still reach European users. Furthermore, current detection and marking technologies are still evolving, with no single industry-wide solution. Approaches like watermarking and metadata are being experimented with, but they can be fragile and easily altered or lost during content editing, compression, or sharing across platforms. Experts suggest a multi-layered approach combining various methods is more powerful than any single technology.
