YouTube built a checkbox, not a detector

Updated 11 October 2026: YouTube’s help page lists a third non-adjustable class, labels applied after manual review, which is not an origin fact.

Since May 2026, YouTube applies an AI disclosure label on its own when its systems detect significant photorealistic AI use in a video the creator never flagged. A creator who thinks the content was incorrectly identified can update the disclosure status in YouTube Studio. The exceptions tell you which signals YouTube actually trusts.

YouTube also moved the label somewhere people will see it. On long-form video it now sits directly below the player and above the description; on Shorts it overlays the video itself. This is the single label format for photorealistic or meaningfully AI-altered content. Content that is unrealistic, animated, or only slightly altered keeps its disclosure in the expanded description, where almost nobody looks.

The labels fall into two classes. Most can be updated. If the automated signal fires, a creator who thinks the content was incorrectly identified can change the disclosure status in Studio. The ones that stay are backed by provenance: content made with YouTube’s own tools, Veo and Dream Screen, and content carrying C2PA metadata that marks it as fully generative AI. Those disclosures are permanent. There is no toggle. YouTube’s help page adds a third permanent class: labels applied after manual review.

The split almost follows a clean line. Two of the three permanent classes are origin facts; the manual-review label is an enforcement decision, and a consumer of the label cannot tell it from provenance. Veo and Dream Screen produced the pixels, so YouTube knows. C2PA is a signed manifest that travels with the file and can be verified. The overridable labels are inference. The “internal signals” that guess at photorealistic AI get a soft label YouTube won’t defend against a creator who pushes back. Manual disclosure is still required for realistic AI, and the detector is framed as help, not enforcement.

If you consume these labels, in a trust-and-safety pipeline, a platform reposting YouTube content, or a study measuring how much AI is out there, the badge is two different things wearing the same colours. A permanent label is either provenance you can rely on or a label YouTube’s reviewers chose to defend. A removable label is either a self-report or an inferred classification a creator can update if they think it was wrong. Do not read “no AI label” as “not AI”: the system still depends on creators disclosing realistic AI manually, and a removable label can be changed.

The incentives point the wrong way for honesty. YouTube says the label changes neither recommendations nor monetization, so a disclosure earns nothing and the label itself costs nothing. The labels that stick are the provenance ones, from YouTube’s own tools and C2PA fully-generative metadata. Outside those, YouTube leans on manual disclosure plus automatic detection of undisclosed photorealistic AI, and a creator who thinks a label was wrongly applied can update it in Studio.

So treat a permanent label as high-confidence provenance or a reviewed enforcement call, its absence as no information, and a removable label as a weak prior. The signal worth relying on is C2PA and known-tool origin. The detection YouTube announced for May 2026 is described only as “internal signals” to “help identify” AI content, backed by a Studio override, which is the shape of a feature meant to nudge disclosure rather than enforce it.