Research · 7 min read
Will your AI product video get an AI-generated label?
Straight from the generator, it can: all 45 clips we checked from five AI video generators carried a machine-readable AI marker, 27 of them the C2PA kind TikTok reads. After one ordinary edit, usually not: every edit we tried removed it. The label depends on your workflow, and disclosure is on you.
Key takeaways
- In our November 2026 check, 45 of 45 clips from five AI video generators carried a machine-readable AI marker: 27 a C2PA Content Credentials manifest and 18 an AIGC label in the format China's labeling rules require.
- Every edit we tried removed the C2PA manifest from all 27 clips that had one, even a 3-second trim that changed no pixels. The AIGC label survived only a re-save that was told to keep metadata.
- TikTok labels content that carries C2PA Content Credentials automatically, and LinkedIn shows a Content Credentials icon on it. The same clip can get a label straight from the generator and none after a trim.
- The rules don't depend on the marker: TikTok, Meta and YouTube ask you to disclose realistic AI content yourself, and Amazon asks sellers to tag images with photorealistic AI-generated people.
- Product photos from an AI image generator carried C2PA too, 4 of 4, and lost it in every conversion we tried, including a plain resize.
An AI product video can carry a label nobody sees: a few kilobytes of metadata saying the clip was generated. TikTok reads that metadata and shows an AI-generated label on its own. A seller who wants to know whether that will happen to their video has to know two things: whether the file has the marker, and whether it still has it after editing.
We checked both on 45 clips.
Do AI video generators label their videos?
All of them did. We took 45 clips from five widely used image-to-video generators, exactly as each generator’s API delivered them, and read their metadata. Every clip carried a machine-readable marker that says it was made by AI.
The markers come in two formats. Three generators embed a C2PA manifest, also called Content Credentials: a signed record inside the file whose “created” action names the source type trainedAlgorithmicMedia, the IPTC term for AI output. Two generators write an AIGC tag into the video metadata instead, the implicit label China’s rules on AI-generated content have required since September 1, 2025.
C2PA manifest: action c2pa.created
digitalSourceType trainedAlgorithmicMedia
AIGC label: {"Label":"1","ContentProducer":"…","ProduceID":"…"}
The signatures differ more than the markers. Only one generator’s manifests chained to the official C2PA trust list, and that generator’s manifest also records an invisible watermark applied to the pixels. Another generator’s signature was intact, but its certificate was not on the list. The third signed every manifest with a self-signed certificate marked “LOCAL USE ONLY”. Whether a platform accepts an unlisted certificate is the platform’s decision; we did not upload test files to any platform.
| Generator | Marker in the file | Clips with it | Signature |
|---|---|---|---|
| A | C2PA manifest | 9 of 9 | intact, not on the C2PA trust list |
| B | C2PA manifest | 9 of 9 | self-signed, marked “LOCAL USE ONLY” |
| C | AIGC metadata label | 9 of 9 | not signed |
| D | AIGC metadata label | 9 of 9 | not signed |
| E | C2PA manifest | 9 of 9 | on the C2PA trust list; notes a pixel watermark |
Photos are marked the same way. All 4 product photos from the image generator carried a C2PA manifest that names AI as the source and notes a watermark.
What happens to the AI label when you edit the video?
It disappears. After each of the seven edits we tried, the C2PA manifest was gone from all 27 clips that had one. That includes a re-save that kept every frame bit for bit and a trim to 3 seconds without re-encoding. The AIGC label behaved the same way, with one exception: it survived a re-save in which we told the tool to keep the metadata, 18 of 18 clips.
Data table, including product photos
| Video edit | C2PA kept (of 27) | AIGC label kept (of 18) |
|---|---|---|
| Straight from the generator | 27 | 18 |
| Re-saved without re-encoding | 0 | 0 |
| Trimmed to 3 seconds, no re-encoding | 0 | 0 |
| Music track added, picture untouched | 0 | 0 |
| Re-saved with "keep metadata" on | 0 | 18 |
| Re-encoded, as an editor exports | 0 | 0 |
| Upscaled to 1080×1920 | 0 | 0 |
| Exported with the macOS media framework | 0 | 0 |
| Photo edit | C2PA kept (of 4) |
|---|---|
| Straight from the generator | 4 |
| Converted to JPG | 0 |
| Resized to half | 0 |
| Resized with the macOS tool | 0 |
| Metadata stripped on purpose | 0 |
The reason is in how the markers are stored. A C2PA manifest lives in its own box inside the MP4 container, bound to the file with a hash. A tool that writes a new file doesn’t carry over a box it doesn’t know. The AIGC label is an ordinary metadata tag, so it survives only when a tool copies such tags on purpose.
Photos lost their marker just as easily. A conversion to JPG, a resize to half, a resize with the macOS image tool and a metadata strip each removed C2PA from 4 of 4 photos. A catalog that resizes images before upload loses the marker at that step.
Which platforms show an AI-generated label, and what do they read?
TikTok and LinkedIn read C2PA directly. TikTok said in May 2024 that Content Credentials “attach metadata to content, which we can use to instantly recognize and label AIGC”, including content uploaded from certain other platforms. LinkedIn shows a Content Credentials icon on image and video content with C2PA information. Meta, YouTube and Pinterest also use metadata, with their own detection on top.
| Platform | What it reads | What it shows | What it asks of you |
|---|---|---|---|
| TikTok | C2PA Content Credentials | AI-generated label, added automatically | label realistic AI content yourself |
| Instagram, Facebook | C2PA and IPTC “AI generated” metadata | AI info | disclose photorealistic AI video and audio |
| YouTube | C2PA, in some cases | altered or synthetic content label | disclose realistic altered or synthetic content |
| C2PA | Content Credentials icon | – | |
| IPTC metadata, plus its own classifiers | AI modified | you can declare AI use yourself |
In practice, the same clip can go either way. Upload the file straight from the generator, and a platform that reads C2PA can label it. Trim it, add music or export it from an editor first, and in our test the file arrived with no marker at all.
Do you have to disclose AI in product videos and listing photos?
Often, yes, and the duty doesn’t depend on the marker. The rules attach to what the content shows and where it is published:
- TikTok, Meta and YouTube ask you to label realistic AI content yourself. YouTube asks for disclosure of realistic altered or synthetic content, such as a scene that looks real but never happened. Clearly unrealistic content and AI help with scripts don’t need it.
- Amazon tells sellers to tag product images that contain photorealistic AI-generated people with the XMP keyword
contains-synthetic-performer, and it adds an indicator for customers. - Google Merchant Center says that “all images created using generative AI must contain meta data indicating that the image was AI-generated”, and asks merchants not to remove the IPTC DigitalSourceType tag.
- Etsy announced on October 22, 2026 that each listing needs at least one unaltered photo of the actual item, that AI-assisted images must accurately represent what the buyer receives, and that sellers must disclose when AI is used to create an item.
- The EU AI Act requires generators to mark their output in a machine-readable format, and requires anyone who publishes a deep fake to disclose it. Its definition of a deep fake covers content that resembles existing “persons, objects, places, entities or events” and would falsely appear authentic. These duties apply from August 2, 2026; generators already on the market before that date have until December 2, 2026 to comply with the marking duty.
Product videos sit close to that line. A video of your real product in a scene that was never filmed resembles an existing object, and a shopper can take it for footage. That is our reading of the text, not legal advice: check the rules for each market you sell in.
Can you avoid the label by removing the marker?
No, because the label was never the rule. A stripped file changes what a platform can detect automatically, not what you have to disclose. Pinterest runs classifiers that find AI content “even if the content doesn’t have obvious markers”. The C2PA specification itself says provenance metadata can be removed, and points to invisible watermarks and fingerprint lookup as the way to find it again. One of the five generators we checked records such a watermark in the pixels. It isn’t metadata, and we have no way to check whether it survived our edits.
The practical risk runs the other way. A clip you trim, caption or set to music before upload is likely to reach the platform with no marker. Then the AI setting at upload is the disclosure you control.
How to check your video for an AI label before you upload
Check the final file, not the one the generator gave you. Three commands cover the formats we found:
# C2PA manifest, if any
c2patool final.mp4
# AIGC label and other metadata tags
ffprobe -v quiet -show_entries format_tags -of json final.mp4
# AI metadata in an image
exiftool -XMP:all -IPTC:all photo.jpg
If you prefer a browser, a C2PA viewer such as contentcredentials.org/verify reads the same manifest.
The file format, length and resolution for each marketplace are a separate check: our free video requirements checker runs it in the browser. What the generator does to the product itself, its label and its logo, is in our frame-by-frame label test.
We make product videos with our engine and check every frame against the product photos before delivery. If you want to see how that works on your catalog, we start with a pilot on 3–5 products. More videos made this way are on our samples page.
Sources
- Quelta AI label test, November 8, 2026: 45 clips, 4 photos, every edit and what survived (CSV)
- TikTok Newsroom, Partnering with our industry to advance AI transparency and literacy (May 9, 2024)
- Meta, Labeling AI-Generated Images on Facebook, Instagram and Threads (February 6, 2024, updated)
- Meta, Our Approach to Labeling AI-Generated Content and Manipulated Media (April 5, 2024, updated: Made with AI becomes AI info)
- YouTube Help, Disclosing use of altered or synthetic content
- LinkedIn Help, Content Credentials
- Pinterest Help, Gen AI labels
- Amazon Seller Forums, News: You must now tag AI-generated people in listings and A+ content
- Google Merchant Center Help, images created with generative AI
- Etsy News, Creativity and Trust in the Age of AI (October 22, 2026)
- Regulation (EU) 2024/1689 (AI Act), Article 50, transparency obligations
- Regulation (EU) 2026/1744 amending the AI Act (Digital Omnibus on AI)
- C2PA Technical Specification 2.4, Explainer