Google Veo video detector

Video that talks

Most early video generators produced silent clips, and the missing or mismatched sound was itself a giveaway. Google DeepMind’s Veo 3, released in 2025, generates audio together with the picture: characters speak with matching lip movement, footsteps land on cue, and crowds murmur in the background. That made it the engine behind a wave of fake vox-pop interviews, fake news segments, “found footage” and short comedy formats, many of which travelled far from their original captions.

What GPTTrace checks

Container metadata and credentials

GPTTrace reads the MP4’s boxes for encoder names, creation dates, Content Credentials and IPTC source-type metadata that Google attaches in some products. A Google-signed manifest that records AI generation is reported as proof.

Frames

Eight frames sampled across the clip go through the neural classifier and spectral checks. Veo’s output is sharp and cinematic, which helps the classifier on original downloads; re-encoded copies from social platforms lose much of that detail.

What it can’t read

SynthID in Veo frames is invisible to everyone except Google. A clip can carry SynthID and show nothing in GPTTrace’s metadata checks if it was re-encoded. For a definitive answer on Google content, use Google’s verifier.

Visual and audio cues

• A visible Veo watermark, or a cropped or blurred corner where one might have been.
• Clips of about eight seconds, or longer videos made of eight-second segments.
• Interview subjects who all speak in the same rhythm, or background signs and text that shift between frames.
• Audio that is too clean for the scene — a busy street with studio-quality speech and no wind noise.

Combine these with the automatic check and with the source: who posted the clip first, and do they show any other material from the same “event”?

Fake news segments made with Veo

One of the most common Veo formats is the fake broadcast: a presenter at a desk or a reporter on location announcing an event, complete with a lower-third caption and a channel-style logo. The logos are invented or imitate real broadcasters. If a clip claims to come from a news organisation, look for it on that organisation’s own website and channels; real broadcasts are published there. The absence of the segment, a logo that doesn’t quite match, or a presenter who doesn’t appear on the channel’s staff pages is usually decisive.

Street-interview formats — a series of passers-by answering a question — are another favourite, because each speaker only needs a few seconds. Notice whether every interviewee speaks with the same pace and the same framing, and whether the background crowd moves naturally.