YouTube auto-labels AI from C2PA metadata
YouTube applies AI labels from three sources you don't fully control. Know which ones fire automatically before you decide what the upload toggle is for.
The altered-content toggle on YouTube's upload screen looks like the whole disclosure system. It isn't. It is one of four inputs, and it is the only one you operate by hand. The other three run without asking, and two of them can label a video you never intended to declare.
YouTube's own documentation is explicit about this. It says labels may be applied automatically to content made using YouTube's GenAI tools, to content that contains C2PA metadata, and to content its internal systems detect as AI generated or altered. You get a notification in Studio when that happens. You can change the disclosure in most internal-detection error cases. You cannot take the label off if the file was made with YouTube's own AI tools, if it carries C2PA metadata, or if it was labelled after a manual review.
That changes what the toggle is for. It stops being the mechanism that decides whether your video carries a label and becomes the mechanism that decides whether you were the one who said so.
The three automatic paths
Each fires on a different signal, and they are independent. Satisfying one does not suppress another.
| Path | What triggers it | How much control you have |
|---|---|---|
| YouTube's own GenAI tools | Content produced inside YouTube's generative features | None. Applied at source. |
| C2PA metadata | A Content Credentials manifest present in the uploaded file | Indirect, via your toolchain |
| Internal detection | YouTube's own systems reading the content | First pass is automatic; a mis-detection can be corrected in Studio. YouTube has said this fires on significant photorealistic AI use, with no numeric threshold |
The middle row is the one worth understanding, because it is the only one where a decision made three steps earlier in your pipeline determines the outcome at upload.
A C2PA manifest is a signed record travelling inside the file that describes what produced it and what was done to it afterwards. It is not a guess about the pixels. It is the file stating its own provenance, cryptographically bound to the bytes. When a platform reads one, it is reading a declaration rather than running a classifier, which is exactly why platforms weight it so heavily.
The part that surprises people: nobody in your workflow necessarily chose to attach it. Whether your export carries a manifest depends on whether the generating model signs, whether your editor preserves the manifest on ingest, whether it writes its own edit record, and whether the export path carries it through. Ask most creators whether their files carry Content Credentials and the honest answer is that they have never checked. We have walked the full chain in sign, strip, survive, and the short version is that manifests are fragile in both directions: they attach without ceremony and they vanish without warning.
So "did I disclose?" and "will this be labelled?" are two different questions with two different answers, and only one of them is yours.
Where the label appears depends on the content, not on who applied it
YouTube places the label in one of two positions. Subject matter does not decide this. Photorealism and format do.
- Prominent, for photorealistic AI content: directly below the player on long-form, as an overlay on Shorts. That is the format YouTube set in May 2026.
- In the expanded description, for content that is animated, stylised, unrealistic, or only slightly altered.
A photorealistic clip that YouTube labels automatically from a manifest gets the same prominent treatment as one you disclosed yourself. There is no cosmetic penalty for having been labelled automatically, and no cosmetic reward for volunteering. What differs is the record of intent, which matters at the enforcement layer rather than the display layer.
What the toggle still owns
Automatic labelling covers some cases. It does not cover all of them, which is why the manual disclosure obligation still exists and still has teeth.
YouTube requires disclosure when AI has meaningfully altered or generated realistic content. Its own examples include making it appear as if someone gave advice they did not actually give, showing a realistic depiction of an event such as a tornado that never happened, and AI generated music.
It explicitly does not require disclosure for:
- Beauty filters.
- Colour adjustment and lighting filters.
- Cloning your own voice for voiceovers or dubs.
- Fully animated or fantastical content.
That list is more permissive than most creators assume, and the voice cloning exemption in particular catches people out in the useful direction: narrating your own catalogue in your own cloned voice is not a disclosable alteration under this policy. Applying a look, grading a shot, or building something openly unreal is not either.
The test is realism plus meaning. Would a reasonable viewer take this as a record of something that happened? If yes, and it did not, disclose. If the video is visibly an illustration, the requirement does not attach, and the label, if it appears at all, sits in the expanded description.
The optimisation people reach for, and why it is wrong
The instinct on reading all this is to try to keep the label off. Strip the metadata, skip the toggle, hope the classifier misses.
YouTube removes the incentive for that in one sentence: disclosing AI content will not limit a video's audience or affect its eligibility to earn money. The label is not a demotion. It is an annotation.
Meanwhile the downside of the opposite strategy is concrete. Creators who consistently choose not to disclose may face a manually applied label, or penalties including content removal or suspension from the Partner Program. So the trade is a cost of zero against a tail risk that reaches your monetization. There is no version of that arithmetic where concealment wins.
There is also a mechanical problem with stripping. Metadata removal does not produce a clean file, it produces a file that looks like one that was never signed. If the generating provider applied an invisible watermark in the pixels or audio, that survives the strip. You have removed the cooperative signal and kept the robust one, which is a worse position than either extreme. The same failure mode shows up on Meta, where the AI Info label can fire from C2PA, IPTC metadata, or self-disclosure and stripping the file only disables the metadata paths.
If the goal is durability rather than concealment, the layering approach in durable Content Credentials is the direction to go.
What actually threatens the video
The policy that decides whether AI video earns on YouTube is not the disclosure policy. It is the inauthentic content policy, which targets mass-produced and templated output that is easily replicable at scale. A labelled, well-made, obviously-authored video is fine. An unlabelled one that is the two-hundredth instance of the same template is not, and the label had nothing to do with it.
That is the reframe worth taking away. Disclosure is a compliance task with a known answer and no distribution cost. Originality is the open-ended one. We have written up what the inauthentic content policy actually says separately, because it is the surface where channels actually get hurt.
A five-minute pipeline check
- Export one representative file and inspect its metadata. You are looking for a C2PA manifest and for provenance fields. Whatever you find is what YouTube will find.
- Do the same after your final re-encode. A file that carries a manifest at export and loses it at compression tells you the label will be inconsistent across uploads. Inconsistency is the outcome to fix, in either direction.
- Decide the direction on purpose. Either preserve credentials end to end and let the automatic path handle labelling, or accept that the manifest will not survive and use the toggle as the deliberate mechanism.
- Write the rule down. Whichever direction you picked, it needs to be the same on every upload, or your channel presents an inconsistent disclosure history that nobody can explain later.
- Run the pre-publish pass. Our pre-publish check list covers the rest of the upload-time decisions that are cheaper to fix before the file leaves your machine.
The whole exercise takes one afternoon once. After that the question "is this video going to be labelled?" has an answer you can predict, which is the actual goal. Being labelled costs nothing. Being surprised is what costs.
FAQ
Does an automatic label mean I broke a rule?
No. It means a signal fired. Content produced with YouTube's own GenAI tools is labelled automatically as a matter of course, and a file carrying C2PA metadata is labelled because it declared itself. YouTube notifies you when this happens, and the notification is informational. Consistent non-disclosure of content that required it is the problem, not the presence of a label.
Can I remove a label YouTube applied automatically?
Sometimes. If YouTube's internal systems mis-detected the video, the Help Center says you can change the disclosure in Studio by answering No on the AI-use survey. You cannot adjust the label if the content was made with YouTube's own AI tools, if the file contains C2PA metadata, or if it was labelled after a manual review. If a locked label is on the video, the useful response is to check why the file triggered it, because that tells you what your pipeline is emitting.
If I strip metadata, will YouTube still label the video?
Possibly, because internal detection is a separate path from the metadata path and it does not depend on your file's fields. Stripping removes the signals you could have controlled and leaves the ones you cannot. Given that disclosure carries no stated cost to reach or monetization, there is no upside to buying that trade.
Do I have to disclose an AI voiceover in my own voice?
Not under this policy. Cloning your own voice to create voiceovers or dubs is on YouTube's list of things that do not require disclosure. That is a narrow exemption though. It covers your voice, not someone else's, and it does not extend to putting words in a real person's mouth, which is squarely in the disclose column.