YouTube's AI labels: what to disclose, and what you don't

Many video creators still read YouTube's AI rules as a question about whether they used AI. YouTube's own documentation asks a different question: does any part of the video claim to be an unaltered recording of a real person, place or event when it is not.

That is the whole test, and it decides whether you touch the disclosure setting at all. Since May 2026 it also decides something else: YouTube now runs the check itself, and applies the label when a creator does not.

What you are required to disclose

YouTube's help centre page on disclosing generative AI content lists three triggers. You must disclose when AI content makes a real person appear to say or do something they did not, when it alters footage of a real event or place, or when it generates a realistic scene that did not actually occur. Each of the three is about a claim on reality rather than about the tool that made the file.

The worked examples on the same page are more useful than the rule. Required disclosures include AI-generated music, an AI-generated shot of a real place added to a travel video, realistic AI video of a match between two real tennis players, footage making it appear that a public figure stole something or admitted to it, and a realistic depiction of a tornado moving toward a real city that did not happen. Also on the list: making it look as though hospital workers turned away patients, or that a real person was arrested or imprisoned.

The label itself has two placements. YouTube announced in May that the disclosure for photorealistic and meaningfully AI-altered content moves to directly below the player on long-form video, and to an on-screen overlay on Shorts. Unrealistic, animated or lightly altered content is disclosed in the expanded description instead. A creator sets the flag during upload in YouTube Studio, under Attributes, in the "AI use" field.

What you do not have to disclose

The exemption list is longer, and it is where most AI video work actually sits. Production assistance does not require disclosure, which YouTube defines to include using generative AI to create or improve a video outline, script, thumbnail, title or infographic, plus idea generation and caption creation. Cloning your own voice for a voiceover or a dub is exempt. So are beauty and colour filters, special-effects filters such as background blur, video sharpening, upscaling and repair, and audio repair.

Non-realistic work is exempt as a category. The help centre gives a unicorn ride through a fantastical world, a green screen placing someone in space, and an AI-generated or altered animation of a missile inside a fully animated video. Gameplay footage is exempt. So is AI generating or extending a backdrop to simulate a moving car.

Read together, the two lists put a stylised AI short, a cartoon explainer, and an AI-assisted edit of your own footage in the clear. A photorealistic synthetic presenter delivering commentary in a video presented as documentary is not.

How the automatic detection works

Since May 2026 YouTube has used internal signals to identify AI content, and states plainly that if a creator does not specify whether they used AI but the systems detect significant photorealistic AI use, the platform applies a label. Creators get a notification in YouTube Studio when a label is applied automatically.

Some labels cannot be removed. Content made with YouTube's own AI tools, such as Veo and Dream Screen, and content carrying C2PA provenance metadata that marks it as fully AI-generated, keep the label. In other cases a creator who believes the system got it wrong can change the disclosure survey under Attributes, and the notification stays until they do. Coverage of the change, including The Next Web's report, has described Google's SynthID watermarking among the signals, which YouTube's own posts do not confirm.

What happens if you skip it

YouTube's position on the label itself is unusually clear: a disclosure "does not change how a video is recommended or whether it's eligible to earn money." The penalty attaches to the pattern, not the label. The help centre states that creators who consistently choose not to disclose may be subject to a manually applied label, or penalties including removal of content or suspension from the YouTube Partner Program.

That enforcement is not theoretical. In September, YouTube told Semafor that it had terminated 20 channels, 13 named in the outlet's investigation and seven related ones, for violating its spam policies. The network Semafor described used human actors reading AI-generated scripts while posing as independent commentators, with the outlet reporting that the majority of the videos carried heated and often inaccurate political attacks. The disclosure rule was not the stated basis for that action. Spam and deceptive practices were.

What this means for AI video creators

Disclosure carries no stated reach or revenue cost, which makes it the cheap side of the bet. The expensive side is a pattern of not disclosing content that plainly should have been labelled.

For anyone building a workflow on Sora, Kling, Veo, Seedance or a local model, the practical move is to decide per video whether any part of it presents itself as an unaltered record of a real person or event. If nothing does, the AI use field can stay where it is. If something does, set the flag at upload and leave it on.

The harder question the policy leaves open is the definition of realistic. YouTube has not published criteria for how photorealistic a scene has to be before the system calls it, which leaves creators estimating the line. Its stated safeguard is the correction path in Studio for cases the system gets wrong.

The disclosure setting is in YouTube Studio under Attributes, and the full example lists are at support.google.com/youtube/answer/14328491.