TikTok's AIGC label fires on realism, not on captions

TikTok's AI-generated content label is a realism rule. It is not a "you used a tool" rule. Creator Academy and the Integrity and Authenticity guidelines both point at the same trigger: AI-generated or significantly edited content that shows realistic-looking scenes or people.

Captions, scripts, colour, crop, and the rest of a normal edit are not that trigger. The clips on this page are caption presets burned into a talking head. Word-timed type is an edit. It is not what the AIGC toggle is for.

The trigger is a reasonable viewer

TikTok requires creators to label AIGC that would lead someone to believe the person on screen is real, or that the event took place in the real world, in the style or quality of a photograph or a video. Unlabelled content in that bucket can be removed, restricted, or labelled by TikTok depending on the harm.

Write the test as a sentence you can actually use at export:

Could a reasonable viewer mistake this picture or this voice for unedited reality?

If yes, apply the label. If no — animation, an obvious illustration, a clearly invented world — the Community Guidelines' AIGC labelling duty does not fire on this rule. Other rules still might.

The cases that always pass the test:

• A photorealistic generated or face-swapped human who looks like a real person.
• AI speech that impersonates a real person.
• A fully generated photorealistic scene, background, or event presented as if a camera captured it.
• Real footage substantially altered so a person appears to say or do something they did not.

The cases that generally fail the test, and so do not need the AIGC label for this reason:

• Burned-in captions, karaoke or otherwise.
• A script, a voiceover you recorded, a colour grade, a crop, noise reduction.
• Stylised or animated work nobody could file as documentation.
• Product graphics that read as graphics.

Do not false-label a real clip with captions on it. The label is supposed to tell viewers the picture is synthetic. Using it as a generic "I edited this" sticker trains people to ignore it.

Captions, scripts and the rest of the edit

Creator Academy's posting flow is a toggle: Post → More options → AI-generated content. The helper text is "Add this label to tell viewers your content was generated or edited with AI." "Edited with AI" is the phrase that makes teams nervous, because every modern timeline uses a model somewhere.

Read it against the Community Guidelines, not against the marketing line. The duty is realistic-looking scenes or people. A caption engine that times words to a mouth is not generating a person. Adding captions in Versely does not turn a camera original into AIGC, and it does not take a generated talking head out of AIGC. The picture decides. The type does not.

Same split for scripts. An LLM draft of the voiceover is production assistance. It is not a realistic synthetic scene. If the face on screen is generated, label the face. If the face is you and the words were drafted by a model, that is not the AIGC trigger.

Minor image work — brightness, resize, a crop to 9:16 — is called out in every neighbouring platform's guidance as not significant. TikTok's own wording is "significantly edited." A caption burn-in is not a significant alteration of a scene. A face swap is.

Auto-label via C2PA is a second path

You do not have to be the one who flips the toggle. TikTok reads C2PA Content Credentials on upload and can apply an "AI-generated" label itself when those credentials are present, or when you used a TikTok AI effect. Once an auto-label is on a post, you cannot take it off.

That is a second path, not a substitute for the duty. A file whose manifest died on a re-encode still needs the toggle if the picture is realistic and synthetic. A file whose manifest survived will get the label whether you wanted it or not. Plan for both: disclose realistic synthetic media yourself, and expect C2PA to disclose it again when the bytes still carry the claim.

The label's job is to let a viewer distinguish real from generated. It is not a reach setting, and it is not a Shop originality test. Those are different documents.

How to apply it without guessing

1. Look at the picture, not the toolchain. If a stranger would think a real person did a real thing, label it.
2. Apply the toggle in the posting flow. Do not bury the fact in a hashtag and call that the label.
3. Keep captions as captions. Burn them in so the silent watch still works; do not treat the caption pass as the disclosure pass.
4. If TikTok auto-labels from C2PA, leave it. Fighting the auto-label is a waste of a review.

If you are unsure, label. The cost of an extra badge on a photoreal clip is a badge. The cost of an unlabelled realistic deepfake is removal.

FAQ

Do burned-in captions require TikTok's AIGC label?

No. Captions are type on top of a picture. The label is for realistic-appearing people or scenes. A camera-original talking head with karaoke captions is not AIGC on this rule. A generated talking head with the same captions still is, because of the face, not the type.

Does using an AI script or colour grade trigger the label?

Generally not. Production assistance is not a realistic synthetic scene. The Community Guidelines key on whether the picture or the voice could pass as real, unedited reality.

If C2PA already labelled the post, do I still need the toggle?

If the file arrived with Content Credentials, TikTok may auto-label it and you cannot remove that label. For realistic synthetic media, apply the creator toggle anyway when you still control it. Auto-label is a backup, not a licence to skip disclosure on a stripped file.

Will the AIGC label hurt distribution?

TikTok has said the label is a transparency tool, not a ranking penalty. What actually hurts distribution is unoriginal, spammy, or deceptive content — a different policy. Do not skip the label hoping for reach.