An AI label says how you made it. It doesn't make it true.

Before a product video goes out, it has to answer three questions, and each one has a different owner.

1. How was it made? If a realistic person, scene or voice was generated or altered with AI, TikTok, Instagram and YouTube each want you to say so. That's the platform AI label.
2. Who paid? If the post promotes a business, or a creator was paid or sent free product, there's a commercial disclosure. That's a separate switch on each platform, and sometimes words in the video as well.
3. Is it true? Every claim about the product needs a basis, and nobody may be presented as a customer with an experience they didn't have. No platform has a switch for this one.

The first two are settings you turn on. The third is work you do on the script and the edit. Teams tend to confuse them because they arrive at the same moment, on the same posting screen, and because ticking "AI-generated" feels like honesty. It is honesty about the production method. It says nothing about whether the person on screen ever used the product.

This guide works through all three for TikTok, Instagram Reels and YouTube Shorts, then applies them to three posts: a synthetic presenter, an altered product scene and an ordinary edit. The platform rules below were checked on 14 October 2026 and change often, so open the linked pages before you rely on a detail. This is operational guidance for a US team, not legal advice; for a specific campaign, especially health, financial or comparative claims, ask a qualified lawyer.

For the general review pass on a draft (facts, visuals, voice and permissions), start with how to review AI-generated social content. This article picks up where that review reaches the posting screen.

The three disclosures at a glance

The rows are independent. One post can need all three, and a post with no AI in it at all can still fail the third row.

AI labels: what each platform asks for

The three platforms agree on the core idea, realistic synthetic content gets labelled, and differ on the edges. Those edges are where ordinary edits get confused with synthetic ones.

TikTok requires creators to label AI-generated or significantly edited content that shows realistic-looking scenes or people, according to its Integrity and Authenticity guidelines. You turn on the AI-generated content setting in the post's options before publishing. TikTok also applies labels itself, including when an upload carries C2PA Content Credentials from the tool that made it. Separately, the same guidelines don't allow AI-generated content that shows a public figure endorsing a product. That's a content rule, not a labelling one.

Instagram follows Meta's rule. Meta's Community Standards require people to use its AI-disclosure tool whenever they post organic content with "photorealistic video or realistic-sounding audio that was digitally created or altered," and say Meta may apply penalties if they don't. The label viewers see is "AI info," and Meta also adds it on its own when it detects industry-standard AI indicators in a file. Note the word "audio": a generated voiceover over real footage is inside Meta's wording.

YouTube has the most detailed list. Its disclosure page asks creators to disclose content that makes a real person appear to say or do something they didn't, alters footage of a real event or place, generates a realistic scene that didn't occur, or creates music that is the main focus of the video. It then lists what doesn't need disclosure: colour and lighting adjustments, beauty filters, background blur, captions, using AI to write a script, title or thumbnail, upscaling, and cloning your own voice for a voiceover. The setting sits in YouTube Studio's upload details. As with the others, files carrying C2PA metadata can be labelled automatically.

Two practical consequences follow. First, YouTube's exemption list is a useful calibration even for the other platforms: tidying real footage is not the same act as fabricating a scene. Second, the platforms don't line up perfectly. A cloned founder voice is listed as exempt on YouTube, but realistic-sounding audio that was digitally created falls within Meta's wording. When the rules split, label for the stricter one on that platform rather than assuming one decision covers all three.

Commercial disclosures are about money, not software

A commercial disclosure answers a different question: is this post selling something, and is someone being paid to say it? It applies to a filmed demo with no AI in it exactly as it applies to a generated one.

On TikTok, the content disclosure setting has two options. "Your brand" is for promoting your own business and shows a "Promotional content" label. "Branded content" is for posting on behalf of another business and shows "Paid partnership." Once the post is live, the label can't be switched; fixing a wrong choice means removing the post and publishing again. A brand account posting its own product is the "Your brand" case.

On Instagram, the paid partnership label belongs to branded content: a creator's or publisher's post influenced by a business partner in exchange for something of value, including free product. Your brand's own Reel about your own product isn't branded content under that definition. A creator you sent product to is.

On YouTube, the paid promotion setting in Studio adds a disclosure label at the start of the video, and the branded content policy says it covers Shorts. Its definition centres on a brand partner, so it's the creator's switch when you sponsor them, not obviously yours on your own channel.

The platform tool may not be the whole job. The FTC's guide for influencers says to put the disclosure in the video itself for video content and not to assume a platform's tool is good enough. If you're paying a creator, write "ad" or "paid partnership with [brand]" into the brief as a spoken or on-screen line, as well as asking them to turn the setting on.

Claims: the part no label covers

US advertising law cares about what a post says and shows, whoever or whatever says it. Four FTC sources do most of the work for a small product team.

• Substantiation. Advertisers need a reasonable basis for objective claims before the ad runs. If the claim names a level of proof, such as "tests show" or "doctors recommend," you need at least that level (FTC policy statement, 1984). "Doesn't leak" is an objective claim. So is a scene that implies it.
• Endorsements. An endorsement must reflect the honest opinion or experience of the endorser, and someone presented as a user must actually have been one (16 CFR 255.1). In the FTC's words, an endorsement "can't be used to make a claim the marketer of the product couldn't legally make" (Endorsement Guides FAQ). Putting a sentence in a presenter's mouth doesn't lower the evidence it needs.
• Fake testimonials. Since 21 October 2024, an FTC rule has barred businesses from writing or creating a testimonial that misrepresents that the person exists, used the product, or had the stated experience (16 CFR 465.2). The FTC's announcement names AI-generated fake reviews as an example, and the rule allows civil penalties against knowing violators.
• AI avatars specifically. The FTC's Q&A on that rule says there's no blanket ban on AI-generated avatars in marketing. What makes an avatar's statement a problem is the same thing that makes a human's statement a problem: a fake or false testimonial underneath it.

Here is why a label can't rescue an invented testimonial. The AI label tells a viewer that the person on screen was generated. It doesn't tell them that the six months of daily use the person describes never happened, and it does nothing about the leak claim inside that sentence. The Endorsement Guides do let advertisers present people who aren't real customers if they clearly disclose it (16 CFR 255.2(c)), but the same section treats a consumer-style endorsement about performance as a claim that the product works for the purpose shown. A disclosure changes who the viewer thinks is speaking. The claim still needs its evidence.

So fix the words first. Then label how the video was made.

Three posts, worked through

The product below is fictional, and so are the posts. They're constructed to show decisions, not taken from a real campaign or test.

Confirmed facts (illustrative): a collapsible silicone lunch box folds to about two inches tall, has a steam vent in the lid and a dishwasher-safe lid. Nothing has been established about leak resistance, heat retention or microwave times. The brand posts from its own accounts and pays no creators.

1. A synthetic presenter

The draft: a realistic AI-generated woman sits at a kitchen counter, holds up the box and says, "I've taken this to work every day for six months and it's never leaked in my bag."

If the script needs a customer's experience, it needs a customer. That's a different production, with its own permission and disclosure questions, covered in AI UGC versus real customer content.

2. An altered product scene

The draft: real footage of the box being opened, with the plain kitchen background replaced by a generated mountain picnic and a wisp of steam added rising from the food.

3. An ordinary edit

The draft: the founder films the box folding on a desk, trims the clip, corrects the colour and adds auto-generated captions. An AI tool helped write the script. The founder records the voiceover in their own voice. The caption reads: "The only lunch box that fits in a jacket pocket."

One change would move this post into a different row. If the founder's voice were cloned instead of recorded, YouTube would still list it as exempt, but realistic-sounding audio that was digitally created falls within Meta's rule. The same file would then need the AI-disclosure tool on Instagram and not on YouTube.

The pattern across the three: the post with the most AI had a testimonial problem, the altered scene had an implied claim, and the post with almost no AI had a superlative. Labels sorted the first two questions every time. The third needed someone with the product facts open.

The pre-publish checklist

Do these in order. Fixing a claim can change the edit, and a changed edit can change the label.

Claims and people

1. List every objective claim in the script, captions, on-screen text and voiceover: sizes, materials, results, comparisons, "only," "best," "never."
2. Match each one to a current source your team owns. Remove or rewrite any you can't match.
3. Look at what each scene implies: steam, a full bag, a smiling "after" shot, a crowd. Treat an implied claim like a written one.
4. Find every first-person statement. If the speaker is generated, an actor or a paid creator, remove experiences they didn't have.
5. Confirm that no real person's face, voice or name appears without permission, and that no public figure appears to endorse the product.

How it was made

6. Note which elements were generated or altered: people, voices, backgrounds, product shots, music.
7. If any is realistic, turn on the AI label on each platform: TikTok's AI-generated content setting, Meta's AI-disclosure tool, YouTube Studio's disclosure.
8. Where the platforms split (a cloned voice, a generated backdrop), decide per platform and record why.

Who paid

9. Brand's own post: choose "Your brand" in TikTok's content disclosure setting.
10. Paid or gifted creator: they use "Branded content" on TikTok, the paid partnership label on Instagram and paid promotion on YouTube, plus a spoken or on-screen disclosure in the video.

Record

11. Save the approved file, caption, the label choices per platform and the source for each claim. If anyone edits the post afterwards, run the list again on the new version.

Before scheduling, the format checks for each destination are in posting one video to TikTok, Reels and Shorts.

Hexwave is in early access. The talking-head format we're building uses ready-made AI characters, never a real person's face or voice, and AI content will be labelled. A label from any tool, ours included, only answers the first question. The claim check stays with someone who knows the product, which is why the workflow we're building puts review before anything is scheduled.

When the checklist isn't enough

The list above handles the everyday cases. Bring in a qualified reviewer when a post makes health, safety, financial or environmental claims, compares you with a named competitor, uses a real person's likeness or voice, runs as a paid ad rather than an organic post, or targets an audience outside the US. Platform ad policies and other countries' rules differ from everything above.

For the rest, the habit is short. Open the product facts beside the draft. Fix what it says, then label how it was made, then mark who paid. If the order ever feels backwards, picture the lunch box presenter with a perfect AI label and a six-month story that never happened. The label was correct. The post still wasn't.