Insights
Platforms are under pressure to handle AI deepfakes after the EU published a draft Code of Practice and technical provenance standards gained traction. The debate now focuses on machine-readable labels, watermarks, and how platforms detect or disclose manipulated media to keep users safe.
Key Facts
- The European Commission published a draft Code of Practice for marking AI-generated content in December 2025.
- Technical provenance standards such as C2PA and watermarking tools like Google’s SynthID are being adopted by platforms.
- Detection tools help but do not fully stop realistic deepfakes; platforms must combine marks, forensics and human checks.
Introduction
Who: regulators, platforms and standards groups. What: new rules and technical tools aimed at AI deepfakes. When: a draft Code of Practice was published in December 2025 and consultations are ongoing. Why it matters: more realistic synthetic media makes clear signals and platform safeguards essential for public trust and online safety.
What is new
In December 2025 the European Commission released a draft Code of Practice that asks platforms and creators to mark AI-generated content in machine-readable ways and to explain when synthetic material may affect public information. At the same time, technical standards advanced: the C2PA provenance specification (v2.2) defines signed content credentials that record origin and edits, while vendors such as Google have added invisible watermarking tools like SynthID to their toolkits. Platforms are updating labeling policies and testing combined approaches — marks in metadata, invisible watermarks and detection signals — but implementation details vary across services.
What it means
For users, clearer labeling and provenance data should make it easier to spot manipulated media and decide what to trust. For platforms, the change means technical work: embedding C2PA content credentials, integrating watermark checks, and improving automated detection. For creators and businesses, new disclosure duties may require documenting sources and training data. Risks remain: detection can lag behind generation quality, watermarks can be stripped in some workflows, and competing standards can fragment the market. Overall, the shift is toward layered defenses rather than a single fix.
What comes next
The Code of Practice is currently in consultation and may be finalised in the first half of 2026. Regulators expect machine-readable labels and provenance to be part of compliance roadmaps ahead of broader AI Act transparency obligations. Practically, platforms will need to operationalise provenance validators, choose whether to adopt open standards (C2PA) or proprietary watermarks, or both. Independent audits and shared test suites are likely to follow so detection and marking methods can be compared and improved over time.
Conclusion
AI deepfakes are prompting a practical safety test for platforms: combine technical provenance, watermarking and detection with clear user labels and human review. No single measure is sufficient, but interoperable marks and regular audits can reduce harm and improve trust.
Join the conversation: share this article and tell us how you think platforms should label synthetic media.
Leave a Reply