AI watermarking. Mandatory by August 2026.

Yes — Art. 50(2) of the EU AI Act requires providers of generative AI systems to mark their outputs in a machine-readable format and as detectable as artificially generated or manipulated. This applies from 2 August 2026. The marking must be effective, interoperable, robust, and reliable as far as technically feasible — the Commission encourages adoption of established standards (C2PA, IPTC, ISO 22144). Art. 50(4) additionally requires deployers who generate or manipulate deepfakes to disclose that the content has been artificially generated. The Commission and AI Office actively support international watermarking standardization.

The Coalition for Content Provenance and Authenticity (C2PA) is an open technical standard for content credentials — cryptographically signed metadata embedded into images, videos, audio, and PDFs. Members include Adobe, Microsoft, Sony, OpenAI, BBC, Nikon, Leica, and the AP. C2PA records: who created the content, what software/AI was used, what edits were applied, and a cryptographic chain of trust. The metadata is tamper-evident — if removed or modified, the credential becomes invalid. C2PA is the leading candidate for EU AI Act Art. 50(2) compliance for image and video generation.

SynthID is Google DeepMind's watermarking technology for AI-generated images, audio, text, and video. Unlike C2PA (which adds metadata), SynthID embeds the watermark directly into the pixel/sample data, surviving common transformations like cropping, compression, and screenshotting. Released to selected partners in 2024, with broader availability through 2025-2026. SynthID is more robust against active attacks than metadata-based approaches but is currently Google-specific — limiting interoperability. Multi-standard implementation (C2PA + SynthID) is increasingly common.

The International Press Telecommunications Council (IPTC) standardizes Digital Source Type metadata for news and media content. IPTC's 'Digital Source Type' field can be set to values like trainedAlgorithmicMedia, compositeSyntheticMedia, or algorithmicMedia to indicate AI involvement. IPTC interoperates with C2PA — many publishing pipelines use both. For news organizations, IPTC alignment is the path of least friction since it integrates with existing photo and video workflows.

Yes — and this is a known limitation. (1) Metadata-based watermarks (C2PA, IPTC) can be stripped by screenshotting, format conversion, or explicit metadata removal — though doing so to evade AI Act obligations is itself a violation. (2) Steganographic watermarks (SynthID, deep learning-based) survive more transformations but can be defeated by aggressive resampling, adversarial preprocessing, or generative regeneration. The AI Act language acknowledges this with the phrase 'as far as technically feasible.' Best practice: use multiple watermarking standards in parallel and pair with visible labels.

Both, at different stages. Provider (Art. 50(2)): the GPAI or image generator must embed the watermark into its output (machine-readable mark). Deployer (Art. 50(4)): when generating or manipulating content that constitutes a deepfake, must disclose the artificial origin to recipients — typically a visible label or disclaimer in addition to the embedded mark. For internal use without external publication, the deployer disclosure obligation generally does not apply, but the provider watermark obligation does. Matproof tracks both layers automatically.

Self-hosted open-source models (Llama, Mistral, FLUX, Stable Diffusion) present a unique challenge: the user is both provider and deployer. If you fine-tune and operate a generative model yourself, you take on Art. 50(2) provider obligations. This means: implementing watermarking in your inference pipeline (post-generation embedding), maintaining the watermark integrity through your processing chain, and documenting your approach in your Annex IV technical documentation. Most open-source ecosystems are converging on C2PA support — check the model's licensing and release notes.