SynthID not detected in Gemini: that does not mean the video is real
Gemini's checker only recognizes Google AI content, so no watermark means no Google mark was found, not that a person filmed it. Know its limits.

If the Gemini verification tool finds no SynthID watermark, that only means Google's own marks were not found. Google's help page says the tool can currently recognize only content created by Google AI tools, so a clip from another generator, or one that has been heavily changed, can pass as unmarked.
What the checker says and does not say
Per Google, a SynthID detection means all or part of the image or video was created or edited by Google's AI models. The absence of a watermark means the content was not created or edited by Google AI, but it could have been made by other AI systems. Google also says Omni videos include an imperceptible SynthID watermark that you can verify in the Gemini app, Gemini in Chrome and Google Search.
| Limit | Value |
|---|---|
| File size | 100 MB or less |
| Video length | Under 90 seconds each |
| Daily video quota | About 10 checks, up to 5 minutes total, per rolling 24 hours |
Pre-flight before you spend a check
Each check uses a slice of a small daily quota, so reject files that cannot pass. The function below applies the documented limits.
import os
MAX_BYTES = 100 * 1024 * 1024
MAX_SECONDS = 90 # help page: videos less than 90 seconds each
def can_check(path, seconds):
size = os.path.getsize(path)
if size > MAX_BYTES:
return False, "file is over 100 MB"
if seconds >= MAX_SECONDS:
return False, "video must be under 90 seconds"
return True, "ok"
with open("sample.bin", "wb") as f:
f.write(b"0" * 1024)
print(can_check("sample.bin", 30))
print(can_check("sample.bin", 120))
What this means for clips you generate
Google's claim is about its own tools. Sume's docs make no claim about whether a clip that you receive through the Video Router keeps a provider watermark, and they do not say how re-encoding by tools such as video trim changes it. Treat that as unknown and test one clip in the Gemini checker before you build a workflow on it.
A better evidence stack
No single check is proof. A claim that a clip is real needs positive evidence, such as the source file, and not just a failed search for a watermark.
- A detector result, with the date, the file hash and the tool.
- The generation record, including the job id and the model.
- The original camera file, if one exists, with its metadata.
- A Content Credentials check, where the file format supports it.
Keep your own record
A checker result is a weak, one-sided signal. Your own record is stronger: store the Sume job id, the model id, the prompt and usage.cost from the poll response next to each file. Google's C2PA Content Credentials check is the other signal, and it is described in the same announcement as rolling out in the Gemini app with Search and Chrome to follow.
Sources
Related posts
More in Use cases
- Talking-head ad: Omni audio, Kling audio on, or Avatar Video?
Pick by length and face. Avatar Video runs 4 to 60 s from a script; Omni makes 3 to 10 s with native audio; Kling 3 adds audio at $0.21 a second.
- Talking photo: H3 Max lip sync, Kling motion control or Recast?
Lip sync follows your audio, motion control follows a reference video, Recast swaps a person. 10-second prices on Sume: $1.00, $1.58 and $3.75 at 768p.
- Tasting menu opener: 40 s, four courses, Omni first/last frames
Four courses, four 10-second Omni clips on Sume: empty plate as the first frame, plated dish as the last. Draft at 360p, finish at 1080p, join on a timeline.
- Tattoo flash sheet reveal clip: Kling 3 for 5 seconds, silent vs audio
A 5-second reveal clip of a tattoo flash sheet with kling-3 on Sume costs $0.70 silent and $1.05 with audio; resolution is 720p or 1080p at the same rate.
Written by Sume