C2PA vs watermark: which one survives a screenshot or re-upload
The short answer is that C2PA manifests may not survive a screenshot, re-upload or heavy compression, while an invisible watermark embedded in the pixels, audio or video stream often does. That is why the EU’s Code of Practice on Transparency of AI-Generated Content calls for a multi-layered approach that uses both. C2PA (Content Credentials) binds a signed, machine-readable manifest to the file, but that manifest can be removed if the file is re-exported, cropped, or posted to a platform that strips metadata. An invisible watermark, by contrast, is part of the content itself and can persist through many common transformations.
What C2PA manifests do and where they break
C2PA is an open standard that creates a cryptographically signed manifest attached to a file. The manifest records who created the content, what edits were made, and whether AI was used. It is the technical backbone of Content Credentials, the label you see on some images and videos. Because the manifest is stored as metadata, it can be lost when the file is resaved without the manifest, when a screenshot is taken, or when a platform removes metadata on upload. This makes C2PA excellent for provenance in controlled workflows but vulnerable once content leaves a secure pipeline.
What invisible watermarks do and where they break
Invisible watermarks embed a signal directly into the content—into the noise of an image, the frequency of audio, or the frames of a video. Because the watermark is part of the data, it can survive cropping, resizing, format conversion and moderate compression. However, aggressive compression, heavy filtering, or certain types of manipulation can degrade or remove it. Watermarks also do not, by themselves, carry rich provenance data; they typically store a small identifier that must be looked up in a database to retrieve the full history.
Why the EU Code of Practice asks for layers
The EU Code of Practice on Transparency of AI-Generated Content, finalised in June 2026, recommends a multi-layered approach for marking AI-generated content. Section 1 for providers asks for machine-readable marking that is effective, interoperable, robust and reliable as far as technically feasible. It explicitly recognises that no single method covers every use case. Combining C2PA manifests with invisible watermarks addresses the weaknesses of each: the manifest provides detailed, verifiable provenance when present, while the watermark provides a fallback that can persist even when metadata is stripped. The Code also calls for metadata that is digitally signed and time-stamped in a secure, tamper-evident way, which C2PA already supports.
How C2PA and invisible watermarks work together in practice
In practice, the most resilient workflow pairs C2PA manifests with invisible watermarks. A creator or AI system first generates content and attaches a C2PA manifest that records the AI model, prompt, and any edits. The same process embeds an invisible watermark that carries a unique identifier. When the file is shared, platforms and tools can check the manifest for full provenance. If the manifest is missing, the invisible watermark can still be detected and used to look up the original manifest or certificate. This dual approach is what the Code of Practice envisions and what many providers are now implementing to meet the EU AI Act’s Article 50 marking duty.
What the EU AI Act Article 50 marking duty requires
Article 50 of the EU AI Act requires providers of generative AI systems to mark synthetic audio, image, video and text output in a machine-readable format so it can be detected as artificially generated or manipulated. The marking must be effective, interoperable, robust and reliable as far as technically feasible. This duty has applied since 2 August 2026, with a grace period until 2 December 2026 for systems already on the market before that date. The Code of Practice offers a practical way to meet this duty by using a combination of C2PA and invisible watermarks, and signatories can rely on it to demonstrate compliance.
Real-world examples: what survives and what does not
Consider an AI-generated image posted to a social platform. If the platform strips metadata, the C2PA manifest is lost, but an invisible watermark may still be detectable. If that image is then screenshotted and posted elsewhere, the watermark may still persist, while the manifest is long gone. For audio, a podcast clip with a C2PA manifest may lose that manifest when converted to a lower-bitrate format, but an invisible audio watermark could remain. For video, a re-encoded clip might drop the manifest but keep the watermark. In each case, the invisible watermark acts as a fallback, while the C2PA manifest provides the richest provenance when available.
Checklist: choosing the right marking for EU AI Act compliance
To comply with Article 50(2), start by ensuring your generative AI system can produce machine-readable marking. Use C2PA manifests for full provenance in controlled environments, and add invisible watermarks for resilience. Test your marking against common transformations: screenshots, re-uploads, format conversions, and compression. Verify that your watermark can be detected even when metadata is stripped. If you adopt the Code of Practice, follow its multi-layered approach and ensure metadata is digitally signed and time-stamped. For creators and studios, tools like L’AItelier can certify files with a tamper-evident provenance certificate, combining C2PA manifests for images and video with invisible watermarks for audio and documents. Finally, remember that content generated before 2 August 2026 does not need retroactive labelling, but any new output must be marked.
Frequently asked questions
Does C2PA work without a watermark?
Yes, C2PA can work without a watermark, but the manifest may be lost if the file is re-exported, screenshotted, or uploaded to a platform that strips metadata. For resilience, the EU Code of Practice recommends pairing C2PA with an invisible watermark.
Can an invisible watermark survive heavy compression?
Invisible watermarks can survive moderate compression, but aggressive compression or certain types of filtering may degrade or remove them. Their robustness depends on the watermarking algorithm and the type of content.
What does the EU AI Act require for marking AI-generated content?
Article 50(2) requires providers of generative AI systems to mark synthetic audio, image, video and text output in a machine-readable format so it can be detected as artificially generated or manipulated. The marking must be effective, interoperable, robust and reliable as far as technically feasible.
Does the EU AI Act require retroactive labelling of old content?
No. Content generated and published before 2 August 2026 does not need to be labelled retroactively. Only new output must be marked.
How can I check if a file has C2PA or a watermark?
You can use public validators like the one at laitelier.com/validator to check for C2PA manifests and watermarks. Many platforms and tools also support Content Credentials verification.