Provider watermark vs deployer label

SynthID and C2PA are provider marks in the file. The visible label is the deployer's job at publish. Provider marking for systems already on the market before 2 August 2026 has a grace period to 2 December 2026. Creators still label on publish. A watermark is not a caption. A caption is not a detector mark.

This page is the two artefacts. Who you are on a hosted studio is Article 50: deployer vs provider. Article 50 has applied since 2 August 2026. Guidelines 20 July 2026. Fines can reach €15 million or 3%.

The mark in the file

Article 50(2) asks providers of generative systems to mark synthetic audio, image, video, and text so it is machine-readable as generated or manipulated. The reader is a detector, a platform ingest, or a downstream verifier.

Two common answers, with opposite failure modes:

SynthID is Google DeepMind's invisible watermark — a signal in the pixels or samples, not a logo in the corner. It is built to remain readable after compression, crop, and ordinary social re-encoding. A detector hit is a claim about a generator family, not a claim about who published the file, whether it was disclosed, or whether it is a deepfake of a real person. Cropping a corner does not "remove the AI mark." The mark is not in the corner.

Content Credentials (C2PA) are a cryptographically signed manifest bound to the exact bytes of one copy. Re-encode, screenshot, or platform ingest that changes the bytes invalidates the hash. That is the mechanism working. Social platforms re-encode on upload as a matter of course. A green tick in a supporting viewer is a claim about one exact file at signing time, not a property that follows the clip through every re-upload.

Neither artefact is the sentence a scrolling viewer reads. YouTube may auto-label from a surviving C2PA fully-AI signal; that is still a platform badge, and some of those labels you cannot peel off in Studio. TikTok and Meta also read provenance when it survives. Cooperative signals are useful. They are not 50(4).

The 2 December 2026 grace is for that provider marking duty on systems already on the market before 2 August 2026. It is not a postponement of deployer labelling. It is not a reason to wait until winter to put a line on a talking head you are shipping tomorrow.

The label a person can perceive

Article 50(4) asks deployers who publish a deepfake to disclose that the content was generated or manipulated, clearly and distinguishably, at first exposure. The reader is a person.

Burned type, a spoken line in the first seconds, an official EU icon plus words: those are deployer objects. YouTube's overlay, TikTok's built-in AI label, Meta's AI Info: those are platform objects that can help a person perceive something, and they are still not the Act by themselves.

First exposure means the viewer does not have to expand a description, wait until the last second, or catch a two-frame flash. Metadata is not first exposure. A SynthID detector result in a lab is not first exposure.

If you sign in, prompt a hosted row, and publish, you are usually the deployer of that file. You cannot add a robust invisible watermark after the fact. You can add a line. The Article 50 disclosure template is the copy. Official icons are optional; labelling is not.

Platform toggles stack. Set them. Do not skip them because the export "already has C2PA." The manifest may die on ingest. Do not skip the burned line because YouTube will overlay "AI" on Shorts. The overlay is YouTube chrome. It does not travel to a paid unit on another surface, and it may not be the wording the Act wants a person to perceive.

What to keep, what to add

Keep the original export. That is the copy whose manifest, if any, still hashes, and whose watermark, if any, was applied at generation.

Do not screenshot a generate and call the PNG a new original. You broke the manifest and you still have a synthetic picture.

Do not strip credentials to dodge a badge. YouTube said on 27 May 2026 that labels alone do not impact recommendations or monetization eligibility. Originality is the penalty document. A stripped file is not more original.

Add the perceivable line when the file is a deepfake for an EU audience, and when the platform photoreal test is met. Photoreal is the disclosure test is the YouTube / TikTok / Meta hinge. Article 50 is resemblance plus authenticity. They overlap on a talking-head UGC take.

Artistic work may label more quietly. Ads and UGC testimonials may not. Provider marking does not get quieter because the work is a sketch. 50(2) is unchanged by the artistic limb.

A UGC video generator may or may not hand you a marked export. Veo 3.1 sits in a family that uses invisible watermarking on Google's side; that still does not print your on-screen sentence. Check the file you actually downloaded. Then label the file you actually publish.

Production order

1. Generate. Archive the provider export.
2. Edit without assuming the manifest survives. Invisible watermarks may; hashes may not.
3. Burn or speak the deployer line. Place it where chrome will not cover it.
4. Set platform toggles.
5. Publish. Treat each destination cut as its own first exposure.

If you want the line on the same timeline as the take, sign up and lock it at picture lock. Do not wait for 2 December 2026 to make the mark "count." The grace is the provider's calendar. Publish is yours.

Two artefacts. Two readers. One file that needs both when you are the deployer of a deepfake.

FAQ

If SynthID is in the pixels, do I still need an on-screen label?

Yes, when 50(4) applies. SynthID is a detector-facing watermark. A person watching a Short does not run a detector. First exposure is a perceivable disclosure.

If C2PA dies on YouTube ingest, did I fail Article 50?

Losing a manifest on re-encode is normal. It is a problem for the provider-style proof of that copy, not an automatic fail of a burned deployer line. Keep the original export for the proof. Put the line on the published cut for the viewer. YouTube may still auto-label from other signals.

Does the 2 December 2026 grace let me ship unlabelled ads until winter?

No. The grace is described as provider marking for systems already on the market before 2 August 2026. Deployer labelling of deepfakes has applied since 2 August 2026. An unlabelled photoreal ad in September is not "waiting out marking."