GrowthCamel generates talking-head video from a digital twin. That makes this topic one we have an interest in, so take the framing with that in mind — and hold us to the sources rather than the summary.
How a label gets applied
Two routes. The first is automatic: Meta reads provenance metadata embedded by the tool that made the file. The industry standards here are C2PA and IPTC, and they are supported by a long list of generation tools. If that metadata says a file was AI generated, Meta can detect it on upload without anyone declaring anything.
The second is manual. Where content is photorealistic video or realistic-sounding audio that was digitally created or altered, the expectation is that the person posting discloses it using the platform's own tools. That obligation does not disappear because a tool stripped the metadata.
There is also a third, newer thing: an optional account-level label that lets someone identify as an AI creator across their profile. Optional is the operative word, and it is a different decision from labelling a single post.
What this means in practice
If you publish AI-generated video, assume it will be identifiable whether or not you say so. The metadata route means the decision is not really “disclose or not”; it is “disclose, or be labelled by someone else”.
The practical reading for a creator using a digital twin: label it. The cost is a small badge. The cost of being caught not labelling is the thing a publication built on trust cannot afford.
If a label is coming either way, the only variable is whether you chose it.
The numbers we will not repeat
A figure circulating widely says AI-generated voices or music can face “up to 80% reach reduction” when judged deceptive. We have not been able to trace it to a primary source, and the phrasing — a round number, an “up to”, no methodology — is the shape of a figure invented to fill a paragraph.
What is documented is directional, not numerical: Instagram has said plainly that it reduces the visibility of certain content, and the published demotion list is specific about format problems. Deceptive synthetic media being treated unfavourably is consistent with that. A precise percentage is not.
This is the same test we applied in the attention span myth and to the “sends are worth three to five likes” claim in what the platforms actually say. It keeps returning the same answer.
The line we draw
There is a difference between synthetic and deceptive. A creator using their own cloned voice and face, labelled, to say something they actually think is synthetic and not deceptive. A video of a real person saying something they never said is deceptive regardless of labelling.
The rules are mostly aimed at the second. The first is a production method, and the honest position is to say so on the post.