YouTube's AI labels catch the deepfake. They miss the script.

YouTube's AI disclosure rule asks creators to label videos where AI meaningfully alters or generates photorealistic content. Everything else slips through: the premise, the research, the script, the voice clone. A 30-minute video where AI does all four needs no label.

YouTuber and author Hank Green walked the platform's policy end-to-end and surfaced the gap in a recent post. Green is not arguing that AI-made video is bad. He is pointing out that the disclosure regime is keyed to visual realism, and visual realism is the wrong axis for what actually shapes what a video says.

The policy language is specific. Creators must disclose AI use "when they use AI to meaningfully alter or generate photorealistic content." The line below that list exempts everything from "idea generation" through "production assistance, like using generative AI tools to create or improve a video outline, script, thumbnail, title, or infographic." Creators may clone their own voices for voiceovers without disclosure. A missile AI-generated or altered inside a fully animated video does not require disclosure. Green's two test cases, a 10-second AI-generated video of someone riding a unicorn through the Soylentius IV swamps and a photorealistic AI lute ballad about the same swamps, get opposite treatment under the same rule: the first needs no label, the second does.

The structure is consistent with itself and unhelpful for viewers. The test is what the camera shows, not what the production pipeline did. A viewer who sees a polished explainer, hears a familiar cloned voice, and follows a tight script has no on-screen signal that the research, premise, and writing were AI-driven. The label the platform does ask for, on the photorealistic case, is the case where the visual cue is already doing most of the disclosure work. The deepfake scenario tends to flag itself by looking wrong. The scripted, AI-voiced argument does not.

Green's concrete thought experiment: a 30-minute geopolitics video. AI generates the premise, gathers the research, drafts the outline, writes the script, and supplies a cloned voiceover. Animations are stylized, not photorealistic. By YouTube's own rule, none of that triggers a label. The video is, in production terms, almost entirely AI-mediated. The platform's disclosure regime sees none of it.

Ars Technica's Ashley Belanger walked the same policy and framed the gap as structural: disclosure regimes that key on visual realism will systematically miss the production pipeline that does the actual persuasion. The lute-ballad deepfake flags itself; the fully scripted, AI-voiced argument does not.

The portable frame is pipeline over photorealism. A useful disclosure rule asks what generated the meaning, not what generated the pixels. YouTube's current rule does the second. A viewer trying to read the trust signal on any future AI-mediated video should look past the label and ask who wrote, who voiced, and who assembled what they are watching. That is not a label the platform can stamp; it is a habit the platform's policy does not currently support.

YouTube has not signaled a policy change. The policy page remains the authority; the loophole is on the page, not a moderation failure. A policy that catches a lute ballad but not a 30-minute scripted argument is sorting the wrong pile.

The next time a video labeled "made with AI" turns out to mean "the thumbnail was touched up," viewers will have a better question to ask.