YouTube will label your video as AI whether you say it is or not

YouTube announced on Tuesday that it will start automatically detecting and labelling videos that contain "significant photorealistic AI use", overriding creators who do not self-disclose. The policy is framed as viewer protection. It is also a transfer of authority: from the creator, who knows what tools they used, to the platform, which is now the arbiter of what counts as AI enough to warn the audience about.

Both of those things can be true, and the interesting questions live in the gap between them.

What actually changes. Until now, YouTube ran a self-disclosure regime. Creators ticked a box if their video contained realistic synthetic media; the label, when applied, sat in the description, where almost nobody looked. Enforcement was inconsistent and creator-dependent.¹ The new policy does two things at once. It moves the label out of the description and onto the video itself, including directly on Shorts. And it adds an automated detection layer that applies the label whether the creator agrees or not.²

Content produced with YouTube's own Veo tool, or carrying C2PA (Coalition for Content Provenance and Authenticity, an industry-standard provenance metadata format) credentials, gets a permanent, non-disputable label. Everything else flagged by the detector can be disputed, on terms YouTube has not yet published.¹

Who actually loses. The press-release framing treats creators as a single block, and viewers as the protected class. That is not how this will land. The creators who lose under auto-labelling are not the deepfake operators the policy is nominally aimed at; those accounts get taken down, not labelled. The creators who lose are the ones doing partial AI work, background fills, voice cleanup, upscaling, b-roll generation, colour grading, whose videos will sit on the wrong side of an undefined photorealism threshold and get labelled as AI when they do not feel like AI videos to the people who made them.

A label is not a takedown. But on YouTube, a label is a signal to the recommendation system, to advertisers, and to viewers who have been trained over the last two years to read "AI" as a warning. The downside risk of a false positive is not symmetrical with the upside of a true one.

And the dispute process, the mechanism that is supposed to make this fair, is exactly the kind of thing that works for creators with managers, lawyers, and policy contacts at Google, and fails for everyone else. We have years of evidence on this from copyright strikes and demonetisation appeals. The platform's contestation queues are not a neutral instrument. They reward platform literacy, free time, and existing relationships.³

The threshold is the policy. YouTube says the trigger is "significant photorealistic AI use". That phrase is doing enormous work. It has no published technical specification. It will be operationalised by an internal classifier whose behaviour is the only place the standard actually exists.

This is a familiar pattern. When a platform writes a rule it cannot quite define, the algorithm becomes the rule, and the rule becomes whatever the algorithm happens to do this quarter. Appeals get adjudicated against the algorithm's output, not against the written policy, because the written policy is too vague to adjudicate against. The creators most affected will learn the threshold by hitting it.

Alphabet on both sides of the trade. YouTube is owned by Alphabet. Veo, the AI video generator whose outputs get permanent labels, is also Alphabet. The labelling regime arrived in the same week as Gemini Omni, Alphabet's new flagship multimodal model. So the company shipping the tools to make photorealistic AI video at scale is also the company deciding how that video gets marked when it lands on the largest video platform in the world.

There is a generous reading of this and a less generous one. The generous reading: Alphabet cannot exempt its own product, which gives the labelling regime credibility it would otherwise lack. The less generous reading: a permanent, non-disputable label on Veo content is not a penalty. It is a permission structure. Veo video that carries the label is, by definition, compliant. The label becomes the thing that lets the content move through the system, not the thing that holds it back. Alphabet gets to sell the generator, host the output, and certify the disclosure, all at once.

I do not think this is sinister. I think it is what vertical integration looks like when a platform decides to regulate a category it also produces in.

The EU is the audience. Article 50 of the EU AI Act takes effect in August. It requires that AI-generated content be disclosed to users. YouTube's move arrives months ahead of that deadline and looks a lot like pre-emption — voluntary action that demonstrates good faith, shapes the implementing guidance, and sets the de facto industry standard before regulators write theirs.

This is the move that platforms have learned to make. Voluntary frameworks adopted ahead of binding rules tend to become the binding rules, because regulators ratify what already exists rather than build from scratch. The threshold YouTube picks now, whatever "significant photorealistic AI use" turns out to mean in practice, is likely to be the floor everyone else builds from. It is worth paying attention to floors that get set quietly, by one company, six months before a regulation that was supposed to set them democratically.

What I would watch. Three things. First, the dispute outcomes — not whether YouTube publishes a process, but whether independent creators can actually win false-positive appeals at rates comparable to studios. Second, the recommendation-system effect — whether labelled videos suffer reach penalties the platform does not acknowledge. Third, what Brussels does with this in August. If the Commission's Article 50 guidance reads like a description of YouTube's policy, we will know which direction the standard-setting went.

The creator-disclosure model did fail. That is the strongest argument for what YouTube is doing, and it is a real one. But "the old system was broken" does not settle who should run the new one, on what standard, with what appeal, in whose interest. Those are the questions, and the press release does not answer them.

Footnotes

¶