The European Commission has published three new studies on technical methods to label and detect AI-generated content. The research focuses on text, audio, images, and video under Article 50 of the EU AI Act. The studies were released on May 8, 2026 via the Commission’s platform.
The studies are meant to support a European code of conduct for labeling AI content—aimed at boosting transparency around media produced by generative AI systems.
What exactly is the Commission investigating?
The Commission is examining how AI-generated media can remain technically traceable. Techniques include:
- Watermarks in images or audio
- Metadata labels
- Cryptographic signatures
- AI-text detection systems
- Authenticity checks for video
The studies assess both existing technologies and new experimental methods, weighing effectiveness, reliability, limitations, and real-world viability.
According to the Commission, future rules must be grounded in “the latest technical developments” and account for differences between text, audio, and visual content.
Three separate lines of research
The work is split into three distinct domains.
Audio
The audio study, led in part by Xavier Serra and Martín Rocamora, explores how AI-generated voices, music, and sounds can be made identifiable.
The topic is accelerating with the rise of voice cloning and AI music generators. Deepfake audio, in particular, poses mounting risks for disinformation, identity theft, and manipulation.
Images and video
Researcher Mario Joachim Fritz focused on AI-generated images and video—specifically how platforms and users can tell whether visuals were synthetically produced.
The study examines:
- Invisible watermarks
- Provenance verification
- Manipulation detection
- Techniques to prevent label removal
The stakes are rising as AI video becomes more convincing. Modern video generators can now create lifelike humans that are hard to distinguish from real footage.
AI text remains hard to spot
The third study, by Giovanni Puccetti, targets AI-generated text—the most technically challenging area.
Many current AI detectors have high error rates. Human-written text is sometimes mislabeled as AI, while advanced systems can evade detection.
The study therefore explores alternatives such as:
- Token-level watermarks
- Statistical pattern analysis
- Model-specific identification
- Verification during generation
The issue spans education, journalism, and search. Organizations increasingly need reliable ways to separate human and AI-generated content.
Why Article 50 of the AI Act matters
Article 50 of the EU AI Act requires certain AI providers to clearly disclose when content is artificially generated.
The EU aims to:
- Curb disinformation
- Increase transparency
- Protect consumers
- Preserve trust in digital content
Deepfakes and synthetic media are a key concern. Content that could mislead people must be explicitly identifiable under the law.
These new studies lay the technical groundwork for future European guidelines and standards.
A tough challenge for the AI industry
The research also makes one thing clear: there’s no silver bullet. Many detection methods are vulnerable to tampering or lose effectiveness when content is edited, compressed, or re-saved.
Interoperability is another flashpoint. AI content moves across platforms, tools, and file formats—making consistent labeling technically complex.
For AI companies, this could mean extra obligations around:
- Transparency
- Content authenticity
- Watermark implementation
- Compliance with EU rules
For users and media platforms, expect new verification systems for online content.
Europe doubles down on AI governance
With these studies, the European Commission reinforces its role as a global AI regulator—not just setting legal guardrails, but actively shaping technical standards.
The upcoming code of conduct on AI labeling could have major implications for developers of generative systems like chatbots, image generators, and video models.
As AI content becomes harder to tell from the real thing, pressure mounts on the industry to make transparency technically enforceable.
Popular news
Grok Says SpaceX Could Hit $5 Trillion After Starfall Mission
Dutch startup Nearfield Instruments rides the AI chip wave to $380 million
Nearfield Instruments: What It Is—and Why It’s Critical to AI Chips
GPT‐NL, Explained: Everything You Need to Know About the Dutch AI Model
Lalaland aims to take the fashion world by storm with AI
Latest comments
Loading