Disclosure & Transparency Rules

The disclosure layer everyone must implement: the AI Act's Article 50 requires that AI-generated or manipulated content — deepfakes, synthetic audio — be labeled as such in machine-readable form, and that AI text published to inform the public be disclosed. Exceptions exist for artistic work and human-edited content, but the default is label.

The enforcement problem is watermarking's weakness: metadata labels strip trivially, and robust watermarking for text doesn't exist. The serious infrastructure is provenance-based — C2PA/Content Credentials signs content at creation so authentic media carries a verifiable history, treating 'real' as the signed case rather than 'fake' as the detected one. The policy consensus is moving to provenance precisely because detection loses.