AI video from prompts: what video models put in a clip, and what our Kling file held

By The NeutronField editorsPublished Updated

Checked on 2026-11-02 with c2pa-rs (c2pa-python), ffprobe and an MP4 box reader

Short answer

Video generators label clips in different places. Our Kling 3.0 clip had no content credential and no visible mark, but carried a metadata tag declaring it AI-generated, the format China's labelling rules require. Google says its Veo videos carry an invisible SynthID watermark. A plain re-encode removed Kling's tag completely.

Evidence order

Strongest first
  1. Signed content credentialNo C2PA manifest anywhere in the file.Absent
  2. Metadata fieldsAn AIGC tag: Label 1, producer kling, a content ID.Present
  3. Visible markNo logo or text on any frame we checked.Absent

Status on the test clip in frame 51, as the API returned it.

Our file test

One ten-second clip from Kling 3.0, read box by box before any re-encoding.

frame 51Frame 51, a still from the test clip: a white cup of tea on a wooden windowsill with rain on the glass behind
▸ 51NF 09 · TEST CLIP

What frame 51 carries

SignalYesNo
Content credentialnot markedmarked
Metadata fieldsmarkednot marked
Prompt saved in filenot markedmarked

Checked on 2026-11-02 with c2pa-rs (c2pa-python), ffprobe and an MP4 box reader. Clip made through Higgsfield's API with Kling 3.0 standard, text to video, 720p, no sound.

Fields read from the test clip, an MP4 of 1280 by 720 at 24 frames a second, 10 seconds, made on 2026-11-02.
FieldPresentValue
C2PA manifestNoNo JUMBF data in the file
AIGC tagYesLabel 1; ContentProducer and ContentPropagator kling; a ProduceID
Encoder tagYesLavf60.16.100 (the FFmpeg muxer)
Prompt textNoNot stored anywhere in the file
Visible markNoNone on the frames we checked
Raw output excerpt · ffprobe
major_brand: isom
AIGC: {"Label":"1","ContentProducer":"kling","ProduceID":"KLingMuse_9906...","ContentPropagator":"kling", ...}
encoder: Lavf60.16.100
c2pa: ManifestNotFound

Three ways a clip can be labelled

Video makers have the same three tools as image makers, used in different mixes. A visible mark is a logo or caption burned into the picture. An invisible watermark is a signal hidden in the frames. A metadata label is text in the file's container, which can be a signed C2PA credential or a plain tag.

Our Kling clip used only the third, and not the signed kind. Its MP4 container holds a tag named AIGC with a short record: a label value of 1, Kling named as both the producer and the distributor, and a content ID. China's rules for labelling AI-generated content, in force since 1 September 2025, require providers to add exactly this kind of hidden label to file metadata, naming the content as AI-generated, the provider and a content number (Cyberspace Administration of China, 2025). The rules also encourage, but do not require, a digital watermark.

Google takes the other route for its Veo videos. Google DeepMind says SynthID adds an invisible watermark to AI-generated video segments and is designed to stand up to changes such as cropping, filters, frame-rate changes and compression (Google DeepMind, SynthID). We tried to make a Veo clip for this page on 2026-11-02, but Google had retired the Veo versions our tools call, so that test is pending. OpenAI's help article on provenance covers images; a Sora clip will get its own test here.

What editing and reposting remove

A metadata label lives in the container, so anything that writes a new container can drop it. We re-encoded the Kling clip with FFmpeg's default settings and the AIGC tag was gone. When we only re-packaged it, copying the video stream and telling FFmpeg to keep custom tags, the tag survived. Most editors and upload pipelines re-encode, so assume a reposted clip has lost its tag.

A watermark is in the frames themselves, which is why it is the record meant to last. But it only helps someone with the vendor's detector, and only for that vendor's videos. For Google media, the Gemini app can check uploaded video for SynthID and reports which parts of the clip carry it.

A visible mark is the easiest to remove, by cropping, and the easiest to fake. Treat its presence as a hint about the source and its absence as no information at all.

Read a clip yourself

You need the original file, not a screen recording or a copy saved from a feed.

  1. Run ffprobe -show_format on the file (it ships with FFmpeg) and look at the format tags for an AIGC entry or other labels.

  2. Open the file in a content credential viewer to see whether a signed C2PA manifest is present.

  3. For Google video, upload it to the Gemini app and ask whether it was created or edited by Google AI; Google allows about ten video checks a day (Gemini Apps Help).

  4. Write down the date and what you found next to the file. Labels change as vendors update their tools.

What this can and cannot prove

An unsigned tag like Kling's states what the tool wrote, but anyone with a text editor could write the same tag into any video or delete it. A signed C2PA credential is stronger because tampering breaks the signature, and a detected watermark is stronger again because it sits in the pixels. None of them shows who wrote the prompt or what was edited later.

Realistic people raise the stakes. If a clip shows a real person, the rules on consent and disclosure matter more than any label; see real people in prompts. For still images from the same companies, see Gemini and Nano Banana and ChatGPT images.

A missing label does not mean a video is human-made.

Sources

  1. Cyberspace Administration of China: Measures for Labeling AI-Generated Synthetic Content, 2025-03-14 (accessed 2026-11-02)cac.gov.cn
  2. Google DeepMind: SynthID (accessed 2026-11-02)deepmind.google
  3. Gemini Apps Help: Verify AI-generated images, videos, and audio (accessed 2026-11-02)support.google.com
  4. C2PA Technical Specification 2.2 (accessed 2026-11-02)c2pa.org

Questions

Do AI videos have watermarks?

Some do, in different forms. Google says its Veo videos carry an invisible SynthID watermark. Our Kling 3.0 clip had no visible or signed mark but carried an AIGC metadata tag declaring it AI-generated.

What is the AIGC tag in a video file?

A metadata entry required by China's labelling rules since September 2025. In our Kling clip it held a label value of 1, the producer name kling and a content ID. It is plain text, not signed.

Does re-encoding remove AI labels from video?

It removed Kling's AIGC tag in our test with FFmpeg's default settings. Container labels and credentials rarely survive re-encoding; a watermark in the frames is designed to.

Do Kling videos have a visible watermark?

Our clip, made through an API on 2026-11-02, had none on the frames we checked. Downloads from Kling's own app may differ by plan, so check your own file.

How can I check a video for SynthID?

Upload it to the Gemini app and ask whether it was created or edited by Google AI. Google allows about ten video checks a day and reports which parts of the clip carry the watermark.

Related reading