AI lip-synced, dubbed documentary scenes: disclosure labels on YouTube and Meta, and what they cost in reach

Date: 2026-10-30 (skills harvest 02:58Z). Why now: on 30 Oct 02:39 the owner asked for each documentary scene in cy/en/es with the mouths re-synced to each language. Every scene is an AI recreation of real dated events with realistic people. That is squarely the content both platforms require us to label. This file covers the posting side only. The render side is the movie worker's.

What the sources say (read 30 Oct 2026 unless marked)

• YouTube Help 14328491, "Disclosing the use of GenAI content" (read in full):
• You must disclose realistic AI content that makes a real person appear to say or do something they didn't; alters footage of a real event or place; or "generates a realistic scene that didn't actually occur".
• Exempt: non-realistic or animated content, minor edits, captions, upscaling, and "cloning one's own voice to create voiceovers or dubs". A third-party lip-synced performance by AI characters is not on the exempt list.
• How: in Studio, Attributes > "AI use" > Yes. Photorealistic content may get a label on the player itself. Non-photorealistic or animated content gets the label in the expanded description.
• Cost: "Disclosing AI content alone won't limit a video's audience or impact its eligibility to earn money."
• Auto-labels: YouTube may apply a label itself for C2PA metadata or detected AI, and a C2PA label cannot be removed.
• Not disclosing: creators who consistently skip it risk a forced label, content removal or suspension from the YouTube Partner Programme.
• YouTube Help 15569972, automatic dubbing (read): "Lip sync … alters the speaker's lip movements to align with the dubbed audio".
• It's an experimental early-access setting (Settings > Channel > Advanced > Allow experimental lip sync), for selected channels and languages only.
• Welsh is not an auto-dub language, so for cy/en/es we upload our own tracks (see youtube-multi-language-audio-tracks-cy-en-es.md).
• YouTube Blog, 4 Feb 2026 (Chandralekha Motati, PM; read):
• Auto dubbing is open to all channels in 27 languages.
• More than 6 million daily viewers watched 10+ minutes of auto-dubbed content in Dec 2025.
• Lip Sync is in pilot.
• Auto dubs have "no negative impact on your original video's discovery algorithm".
• These are YouTube's own figures, not independently checked.
• Meta Newsroom, 6 Feb 2024 "Labeling AI-generated images…" (read):
• Meta requires use of its disclosure tool for organic posts with "a photorealistic video or realistic-sounding audio that was digitally created or altered", and "may apply penalties" if you don't.
• High-risk deceptive content may get a more prominent label.
• For the engagement cost of labels on images, see ai-image-labels-disclosure-engagement-meta-youtube.md.

How to do it

1. Decide per clip. Disclose (Yes) if the clip has realistic AI people, voices or places. Leave it off only for clearly animated or illustrative clips. When in doubt, disclose: the cost is a label, and not disclosing risks the Partner Programme.
2. YouTube upload checklist: - AI use = Yes. - Original language = Welsh. - Add the en and es tracks as our own dubs (delete any auto-dub first). - Leave automatic dubbing and experimental lip sync off. YouTube's machine lip sync would re-move mouths we already synced, and there is no Welsh.
3. Facebook/Instagram/Threads: toggle "AI info" / "Add AI label" on every documentary Reel or video. Do it in the composer, or set the label flag when posting by code.
4. Say it on screen and in text, once and neutrally: - Opening card or end card: "Dramatised reconstruction. Scenes and voices made with AI from the dated records listed at ." - First line of the description: the same sentence plus the evidence link. This turns the label into a provenance point ("reconstructed from records") rather than a suspicion.
5. Keep C2PA if the generator adds it. Platforms read it and will label anyway. Stripping it to dodge a label is the "consistently not disclosing" pattern.
6. Measure: compare views and retention by audio language (YouTube Analytics > Audio track). For Reels, compare labelled reach against our earlier unlabelled archive clips before deciding whether labels hurt. Don't guess.

How we apply it to Cedwyn

• Every door from scene 3 on is a recreation with realistic AI characters speaking, which is the "realistic scene that didn't actually occur" case. Every upload of those scenes (YouTube, the FB page Great House Farm II, IG @cedwynestate) gets AI use = Yes plus the reconstruction card. Doors 1–2 use the same rule.
• The site's /documentary/ player isn't governed by platform labels. Put the same one-line reconstruction note under the player and in the download file's metadata (title/comment fields). A downloaded clip then still carries its provenance when someone re-shares it.
• Editorial rules hold:
• "dramatised reconstruction" and "made with AI", never "real footage";
• no likeness of real family members (the movie runner already bans this, and it also keeps us out of the "real person appear to say" case);
• the banned words never appear in labels or descriptions.
• Language tracks: the owner's cy/en/es lip-sync request suits YouTube's own-dub route. One video with three tracks keeps views, comments and watch time on one URL instead of splitting them across three uploads.
• Not checked: whether YouTube places a player label (vs description) on our stylised-realistic scenes; we'll see it on the first upload. There is also no independent figure for how a label changes reach on video.

Sources

• support.google.com (read 30 Oct 2026)
• support.google.com (read 30 Oct 2026)
• blog.youtube (read; 4 Feb 2026)
• about.fb.com (read)

Raw Markdown: ai-lip-sync-dubbed-documentary-disclosure-youtube-meta.md · version and changes: changelog · machine index: agents.json