The short version, before the detail: generated content is allowed on every major short-form platform. What is not allowed is undisclosed realistic synthetic media, and the enforcement mechanism is detection rather than an honesty system. Anyone treating the label as optional is betting against classifiers that have already labelled over a billion videos.
Everything below is current as of August 2026 and carries dates, because this is the fastest-moving corner of the rulebook and undated advice about it rots within months. Re-check before relying on any of it for a campaign.
TikTok
TikTok permits AI-generated content and asks you to label anything realistic. Since integrating C2PA Content Credentials in January 2025 it also reads provenance metadata embedded by the tool that produced a file, and applies the label itself when it finds it, whether or not you switched the toggle on.
The practical read is that disclosure has become a courtesy and detection is the enforcement. What gets actioned is overwhelmingly the material that was not labelled rather than the material that was AI, which makes the correct behaviour easy: label it, and move on.
Instagram and Facebook
Meta has required disclosure of AI-generated content across Instagram and Facebook since May 2026, and reads the same class of provenance signals alongside its own classifiers. The ad side is stricter than the organic side, which is the detail that catches people out: a video that ran organically without incident can be rejected the moment it is put behind spend.
Organic policy and ad policy are different documents. Passing one is not passing the other, and the rejection tends to arrive after the budget has been allocated.
YouTube Shorts
YouTube requires creators to disclose realistic synthetic content at upload and surfaces a label in the description, with a more prominent one for sensitive topics. The threshold is the same one every platform has converged on: could a reasonable viewer mistake this for a recording of something that happened.
The category that is exempt everywhere
AI-assisted text does not need disclosing. Scripts, captions, titles, hashtags and ideation all sit outside these rules on every platform, which is worth knowing because a great deal of anxious over-labelling comes from assuming otherwise. The requirements are aimed at synthetic people, synthetic voices and synthetic scenes.
- A generated presenter saying words: label it.
- A real recording edited so someone appears to do something they did not: label it.
- A clearly stylised animation nobody would mistake for footage: generally not required, though check the specific platform.
- An AI-written script read by a real person: not required anywhere.
- AI-generated hashtags, titles and captions: not required anywhere.
The two mistakes worth avoiding
The first is stripping metadata to avoid a label. It does not reliably work, because classifiers infer origin from the content itself rather than only from the file, and an undisclosed post that gets detected is treated considerably worse than a disclosed one would have been.
The second is generating a testimonial rather than dramatising one. Using a generated presenter to deliver a genuine customer review is a recognised format that needs a label. Generating the review itself is a fabricated endorsement, which is an advertising-standards problem in most markets and not something a disclosure label cures.
Where we think the line sits, and why, is set out on the responsible-use page. The short version is that none of this makes generated content a problem to hide, and the rule that matters is simply that the label should never be a surprise to you.