Machine-readable marking of synthetic content — Article 50(2)
Article 50(2) requires providers of generative AI to ensure outputs — images, audio, video and text — are marked as artificially generated in a machine-readable format, and detectable as such. Marking must be effective, interoperable, robust and reliable so far as technically feasible. The grace period for systems placed on the market earlier ends December 2, 2026.
Who must comply
Providers of generative systems whose outputs reach the EU: image/video/audio generators, LLM-powered writing products, synthetic-voice APIs. If you build on a major API (OpenAI, Google, AWS, Azure), part of the marking may already be handled upstream — but you must verify it and not strip it.
How to comply
- Embed an invisible watermark at generation time (SynthID-class systems, or open-source approaches like Stable Signature for self-hosted models).
- Attach C2PA Content Credentials so provenance survives as verifiable metadata — pairing both is emerging best practice.
- Verify marking survives your delivery pipeline (compression, resizing, CDN re-encoding).
- Record your technical-feasibility reasoning; 'state of the art' is a moving target regulators will ask you to evidence.
Tools for Art. 50(2) (25)
Most teams start in Watermarking & Content Provenance; everything below maps to this obligation.