AI Video for TikTok: Specs, the AIGC Label, and How Detection Actually Works
TikTok labels AI content automatically through C2PA and frame classifiers — over a billion videos so far. The export specs, what triggers a label, and what unlabelled content risks.
TikTok is the platform with the most developed automatic AI detection, and the most explicit about it. If you publish generated video here, the label is not a decision you make — it is a decision you can participate in.
The specs
1080 × 1920, 9:16. Other ratios upload but do not fill the screen. Export MP4 or MOV, H.264 video, AAC audio at 128 kbps or higher, around 8 Mbps.
Length: up to 60 minutes for pre-recorded uploads, 10 minutes when recording in-app. Engagement peaks at 15–30 seconds.
File size: the mobile app caps uploads around 72 MB on Android and 288 MB on iOS; the web uploader at tiktok.com accepts up to 4 GB. If a longer generated piece is bouncing off the app, upload from the web.
Safe zone: a username at the top, action buttons down the right, caption block across the bottom. Keep the subject central — which for generated video means composing that way at generation time, not cropping later.
The generation-side note: models frequently default to 16:9. Set the aspect ratio to 9:16 before you run, not after.
The label, and how it gets applied
TikTok requires a visible label on any video using AI to generate or significantly alter realistic depictions of people, places or events. The policy is disclosure, not prohibition.
Detection runs two ways:
C2PA Content Credentials — provenance metadata carried in the file. If those markers are present, TikTok applies the AI-generated label automatically, whether or not you disclosed. Over 1.3 billion videos have been labelled through automated detection.
Frame classifiers — analysis for the artefacts generated video produces: inconsistent lighting between subject and background, unnatural hand and finger geometry, temporal inconsistency in facial movement, and diffusion-characteristic noise patterns.
That second list is worth reading as a quality checklist. The things TikTok's classifier looks for are the same things a viewer registers as "this looks AI".
Unlabelled AI content risks being auto-labelled, having distribution reduced, or being removed. AI-assisted text work — scripts, hashtags — is exempt; the policy is about realistic depiction.
What this means in practice
Disclose. The label is applied either way, and self-disclosure costs you nothing while non-disclosure risks distribution.
Stylised content is lower-stakes. A cartoon or an obviously unreal transformation effect is not a realistic depiction of a person, place or event. The policy is aimed at content that could be mistaken for real.
Photoreal content with a real face is the highest-stakes case — and if the face is not yours, it is the case where the platform, the models and likeness rights all point the same way.
Format notes for generated video specifically
Generation models are good at exactly the length TikTok rewards. Five to eight seconds per generation holds identity; three of those cut together is a 20-second video in the peak engagement band.
Hooks matter more here than on any other platform because the algorithm weights watch time heavily. A generated clip that takes three seconds to establish what it is has already lost — put the reveal first.
The checklist
- Generate at
9:16, 5–8 seconds per shot. - Export 1080 × 1920 MP4, H.264, ~8 Mbps, AAC.
- Compose for the central safe zone.
- Self-disclose with the AIGC label.
- Use in-platform audio.
- Upload from web if the file exceeds the app's cap.