On August 2, a rule most AI filmmakers had been ignoring became the thing their distributors care about. The EU AI Act's transparency obligations under Article 50 are now in force, and they say the quiet part out loud: if a machine made your video, the file has to admit it.
This isn't a far-off regulatory threat or another content-moderation debate. It's a labeling requirement with a compliance date, and the major labs saw it coming. Sora, Veo, and Kling already stamp their output with signals built to survive the trip from render to upload to re-encode. For anyone shipping AI video into the EU — or onto the platforms that operate there — provenance has quietly moved from a nice-to-have into part of the deliverable.
What Article 50 actually requires
The rule is narrower and more practical than the headlines suggest. Providers of generative AI have to mark their outputs in a machine-readable format so the content is detectable as artificially generated. That's it: not a ban, not a quality bar, just a requirement that the work carry a signal saying a model made it.
The obligation took effect on 2 August 2026. Systems already on the market before that date get a grace period until 2 December 2026 to meet the machine-readable marking requirement, and there's a separate disclosure duty for deepfakes that lands on whoever deploys them, not just the lab that built the model. The through-line is detectability. The regulators don't want to stop synthetic video. They want it to be legible as synthetic.
Two signals doing two different jobs
The catch is that "machine-readable marking" isn't one technology. It's two, and they fail in opposite ways.
The first is C2PA Content Credentials: cryptographically signed metadata attached to the file that records what made the content, when, and with which model. When the manifest is intact, it's the cleanest provenance you can get. The problem is how easily it disappears. Screenshot the frame, re-export through an editor that drops metadata, or upload to a platform that strips it, and the credential is gone without a trace.
The second is SynthID, Google DeepMind's approach: an invisible watermark trained directly into the pixels rather than bolted onto the file. Two models are trained together, one to embed the signal and one to detect it, so the mark survives compression, resizing, and format conversion. The tradeoff is that it's invisible by design and needs a detector to read, and coverage only extends as far as the models that emit it.
This is why the labs run both. OpenAI signs Sora output with a Content Credentials manifest and an invisible watermark. Google embeds SynthID across Veo, Imagen, and Lyria. Neither layer is sufficient alone: metadata is fragile, watermarks are silent. Together they cover for each other.
What this changes for creators
Here's where it stops being abstract. Most serious AI film work mixes models, and every model marks its output differently. Assemble a final cut from Kling, Veo, and a stock clip, and you've built a patchwork: some frames watermarked in the pixels, some carrying credentials in metadata, some stripped clean by an export you didn't think twice about. Editing and re-encoding routinely weaken or drop C2PA data. The burden of shipping a clean, disclosed file lands on whoever assembles the final render, which is you.
Platforms are reading these signals too. TikTok, Meta, and YouTube already apply "AI-generated" labels partly off C2PA and SynthID. So provenance isn't only a legal question about EU exposure. It shapes how your work gets presented and how far it travels once it's live.
What to do about it
Treat provenance as a line on your export checklist, not an afterthought you handle if someone complains. Know which of your tools emit Content Credentials and which watermark the actual pixels, so you're not guessing what survived the edit. Don't strip metadata when you don't have to. And when you deliver, disclose plainly, because a label is cheaper than a takedown.
The cleanest place to get this right is the final render, where a stack of clips from different models becomes one file. If your pipeline builds a whole movie from a single prompt, that last step is where the provenance either holds together or falls apart. It's worth designing for on purpose rather than discovering after the fact — the kind of thing a prompt-to-movie pipeline should handle so you don't have to babysit it clip by clip.
The watermark you can't see is now part of the picture. The creators who treat it as infrastructure instead of a nuisance are the ones who'll still be shipping cleanly in December.