TikTok Shop sells on video and images more than any other marketplace, and generative tools have made both cheap. A product can be rendered before it exists, a model can be invented, a result can be shown that the product has never produced. On 2 September 2026 TikTok Shop published an updated set of rules for exactly that, applying to sellers, creators and affiliates in the United States.
This is what the rules say, in TikTok's words, what they mean for the images and videos on a product listing, the listing policy that sits underneath them, and the way to brief a creative team so nothing it makes has to be taken down.
On TikTok Shop the product itself has to be photographed: since its listing policy update of 1 September 2026, placeholders and digital renderings of a product are prohibited. AI may still be used for editing, backgrounds, styling, translation and scripts, as long as it does not change the product's size, colour, material, quantity, contents, features, compatibility, safety or performance. And content that AI fully or significantly created or altered has to be labelled, with on screen text or the platform's AI generated content toggle, under the rule in force in the US from 2 September 2026.
TikTok's governing line: "Content must be truthful and accurate and must not contain exaggerated, false, or misleading claims". Everything else in the policy is an application of it to AI.
The disclosure rule follows: "If your content is fully or significantly created/altered by AI, you should label it", either with on-screen text or with the platform's AI-generated content toggle. The threshold is fully or significantly. A colour correction does not trigger it. A generated scene, a generated person or a generated demonstration does.
The policy lists the attributes that have to survive any generation or edit intact. AI may not be used to alter a product's "size, color, material, quantity, bundle contents, features, compatibility, safety attributes, or performance". That list is the whole of what a listing image is for, which is the point: an image exists to tell a buyer what will arrive, and TikTok's earlier guidance on the same subject says plainly that "AI making your product look different from what buyers actually receive" is a product-not-as-described problem, not a creative one.
Results are treated the same way. TikTok's example is a "shampoo that instantly grows hair", and the rule extends to showing a product performing faster than it realistically does. A before-and-after that the product cannot produce is a false claim whether a camera or a model made it.
People and property are covered too. Content may not use the "likeness, voice, name, trademarks, or other intellectual property, of any third party without proper permission", and a generated doctor, professor or expert endorsing a product is prohibited outright. TikTok's shorter guide gives the example of AI avatars acting as experts to push efficacy claims.
The label is required when content is fully or significantly created or altered by AI. TikTok does not publish a list of exactly what counts, so the practical line is drawn from what the rule is for: telling a viewer when what they see is not what a camera recorded.
Usually below the line. Correcting colour so the product matches the real thing, removing dust or a stray mark, evening out a background, cropping, sharpening, adding captions or subtitles, and translating the words. These make the footage clearer without changing what happened in front of the camera.
Usually above the line. A generated background or setting, a generated hand or person, a generated demonstration, a voice that is not a real person's, a scene that never happened, and anything that changes how the product looks, moves or performs. These are the cases the label exists for, and the last of them is not allowed at all.
Style changes. TikTok permits converting live footage into a cartoon, anime or illustrated style. A viewer can usually see that a cartoon is not a recording, but a style change made by AI is a significant alteration, so labelling it is the safer reading.
When in doubt, label. A label on a borderline video costs very little. A missing label on a video the moderators judge significant costs content restrictions and, for the seller, health rating points.
Write the brand's own version of this line into the creative brief, with examples from its own products, so every editor and creator applies the same judgement.
The policy is not a ban on AI. It permits converting live-action footage into styles such as cartoons, anime or illustrations; using AI for editing and quality improvement; translation, dubbing and subtitles; and copywriting and script support, provided the content stays accurate. For a brand that means the production work (cutting, cleaning, captioning, versioning for markets) can be automated freely. The representation of the product cannot.
The AI policy sits on top of TikTok Shop's Product Listing Policy, updated on 1 September 2026, and the two have to be read together. The listing policy says "Placeholders and digital renderings of a product are not allowed" and that images must show what customers will receive. So a fully generated product image is not a disclosure question at all; it is prohibited as a rendering, label or no label. The main image must show the front physical view of the product on a pure white background, which is the subject of its own rules, and the whole image set may not include "any added logos, text, borders, watermarks, or graphics".
The technical limits are modest: up to nine square images of at least 600 by 600 pixels, in colour, and one video of up to 5MB. The modesty is deceptive. A nine-image set with no text on any image has to do its explaining through the photographs themselves, which is harder than an infographic and the reason
product photography that sells treats each shot as the answer to one question.
Watching What Affiliates Post
A brand controls its own listing images and its own videos. Affiliate creators make far more content about its products than the brand does, and every one of their videos links to the brand's listing. The rules apply to them as creators, and the consequences reach the seller through the listing.
Brief before they post. Every affiliate who receives a sample or joins a campaign gets the same short brief: what the product does and does not do, the claims that may not be made, and the AI rules in plain words, including the label and the ban on generated experts.
Look at what is linked to the products. Read the videos that link to the brand's listings regularly, the best selling ones first. A video selling well on an overstated result is a risk growing with every view.
Ask for changes quickly. A creator asked politely to remove or correct a video usually does. Make the request with the specific claim or scene and the rule it breaks.
Stop working with repeat offenders. A creator who keeps overstating, or keeps posting generated demonstrations, is not worth the commission. Remove them from campaigns and from the sample list.
Keep a record. The brief sent, the videos reviewed, the requests made and the outcomes. If a listing is penalised over an affiliate's content, the record shows the brand did what it could.
The work takes an hour or two a week for an active affiliate programme, and it is the only control the brand has over most of the content selling its products.
The AI policy names both account-level and content-level consequences. Accounts receive warnings for minor violations, then posting restrictions, commission suspension or permanent bans for repeated or severe ones. Content can have its visibility restricted, its product links removed, and its Creator Health Rating points deducted. The listing policy adds the seller side: deductions from the Account Health Rating, listing removal, revocation of selling privileges and refunds to customers.
That last list is why this is a seller's problem even when a creator made the video. A creator's generated demonstration that overstates a product is a false claim attached to the seller's listing, and
how the Account Health Rating works explains how points accumulate into restrictions. The brief a brand gives its creators is the control it has, and
who should make your TikTok Shop content is where that brief belongs.
Photograph the product. Generate the context. The product in every image is a photograph of the real item: its colour, its size against something known, what is in the box. Backgrounds, settings and styling can be generated, and the image is labelled if the generation is significant.
Demonstrate only what the product does, at the speed it does it. A demonstration video shows a real use in real time. If a result takes a week, the video says a week.
No invented people, no invented experts. A real person on camera, or a plainly illustrated figure. Never a photorealistic generated presenter, and never anybody who looks like a professional endorsing a claim.
Label what is generated, and keep the files. The toggle or the on-screen text, and the source files kept so the label can be shown to have been applied.
Keep text off the images. The listing policy forbids added text and graphics on product images. The explaining goes into the title, the description and the video, and
the video formats that sell on TikTok Shop covers how a video carries the information an Amazon infographic would.
When a listing image or video is questioned, the brand that can show what was photographed, what was generated and what was labelled answers in minutes. The brand that cannot is arguing from memory. A simple record kept as the work is done is the difference.
One line per asset. The file name, the product it shows, and where it is used.
The source. The original photograph or footage the asset was made from, kept in full resolution. The product in it is the real item, and the original proves it.
What AI changed. In plain words: the background replaced, captions added, colour corrected, a scene generated. Enough for someone else to judge whether the change was significant.
Whether it was labelled. For videos, whether the AI generated content toggle or on-screen text was used, with a screenshot of the post showing it.
Who approved it. The person who checked the asset against the product and the rules before it went live, and the date.
Keep it current. When an asset is replaced, mark the old line retired rather than deleting it, so the history of what ran on a listing can still be shown.
The record also keeps the work honest. An editor who has to write down what was generated thinks twice about generating something the product could not live up to.
We build listing image sets that pass these rules by design: the product photographed, the context generated where it helps and labelled where it must be, nothing on the image that the policy forbids, and the claims checked against what the product can be shown doing. That is the work of our
listing images service, on TikTok Shop and on the marketplaces whose rules differ from it. If your TikTok listings were built before September 2026, a review against the two policies is a morning's work and worth doing before a creator's video brings the listing to the moderators' attention.