Veo is Google's video generator, and by 2026 its short clips β usually around eight seconds, with sound baked in β are convincing enough to slip past a quick glance in most feeds. The useful part is that Google, like OpenAI with Sora, ships every Veo video with two markers on purpose: a watermark you can see in the corner and an invisible SynthID signature you can't. The catch is that the visible one can be cropped and the invisible one, while sturdier, only tells you whether Google's tools were involved.
So "is this a Veo video?" has a layered answer. When the markers are present they get you most of the way there. When they're gone or inconclusive, you're back to reading the footage and, more importantly, checking the story attached to it. Here's the order I'd work through.
The 6-step check
Check the lower-right corner for the Veo mark
Videos generated with Veo 3 carry a visible watermark β a small Veo logo tucked into the lower-right corner of the frame. It's the fastest tell, so play the clip full-screen and watch that corner for the first second or two. One catch worth knowing: clips made by Ultra-tier subscribers inside Google's Flow filmmaking tool ship without the visible mark, so a corner with no logo doesn't clear a video. Treat a logo you can see as a strong yes and a logo you can't see as unfinished business.
Run it through SynthID in the Gemini app
Beyond the logo you can see, Google embeds an invisible SynthID watermark into every Veo clip β woven into both the picture and the audio track. You don't need special software to read it. Open the Gemini app, upload the video (up to 100 MB and 90 seconds), and ask something like "Was this made with Google AI?" Gemini scans for the SynthID signature and tells you whether it found one, often pointing to which segments carry it. Google said this in-app check had already been used more than 50 million times by May 2026, and it's the most reliable single move you can make on a suspected Veo clip.
Know what SynthID can and can't confirm
SynthID is designed to survive the rough handling that strips other metadata β compression, resizing, cropping, re-encoding, screenshots of a frame, the usual re-upload churn. That durability is what makes it useful. But it only recognizes content made by Google's own models: Veo for video, Imagen for images, Lyria for audio. A clip from a different generator won't carry it, and a genuine camera recording won't either. So a positive SynthID result is close to confirmation that Google AI was involved; a negative result just means "not made by Google's tools" β not "real."
For high-stakes clips, note the SynthID Detector portal
Google also runs a dedicated SynthID Detector β a standalone portal that scans an upload and highlights exactly which parts carry the watermark. As of I/O 2026 it opened to early testers first: journalists, researchers, and media professionals, through a waitlist on the SynthID page at DeepMind, with no published date for general access yet. If you do verification work for a living it's worth requesting; for everyone else the Gemini app check covers the same signature and is available now.
Read the footage for Veo's own tells
When the watermark's been cropped away and you can't run a SynthID check, fall back on what the clip shows. Veo is good, but it still leaves seams: base clips tend to run only about eight seconds, so unusually long single takes lean real. Its synchronized audio can sound a touch too clean β no room tone, no incidental background noise a real mic would pick up. Watch for objects that pop in or vanish at the edges, hands or faces that shift oddly between frames, on-screen text that scrambles, and stumbles with local languages or pronunciation. None of these is proof by itself; a cluster of them on a suspiciously flawless clip is a reason to keep checking.
Verify the claim, not just the pixels
Most of the damage from a synthetic clip comes from the story riding on it β a supposed event, a "leaked" moment, words put in someone's mouth. Even when you can't pin down which model made a file, you can test the claim around it. Has any credible outlet reported the event? Does a reverse image search on a keyframe surface the original, earlier footage? Are fact-checkers already on it? That last one is quick: FAXTR searches 100+ fact-checking organizations in a single query, so you can see whether the clip in your feed already has a published verdict before you pass it on.
Signals worth training your eye on
You won't run a SynthID check on every clip that scrolls past. These are the tells worth catching at a glance β any one of them means a video is worth a closer look before you believe or forward it.
- βA small Veo logo in the lower-right corner β the giveaway that a clip was exported from Veo 3.
- βA roughly eight-second clip with polished, synchronized audio but no room tone or ambient noise.
- βObjects appearing or disappearing at the frame's edges, or hands and faces morphing between frames.
- βOn-screen text or subtitles that dissolve into nonsense when you pause, or garbled local-language speech.
- βA blurred, cropped, or overlaid corner where the watermark could have been removed.
- βA dramatic "breaking" moment that no news outlet has reported and that exists only as this one video.
Why "no watermark" keeps fooling people
The natural inference is that a real Veo clip would show the logo, so a clean-looking video must be genuine. That's backwards. The corner mark is easy to crop, and some Flow exports never carry a visible one to begin with. Meanwhile the invisible SynthID signature β the part that actually survives resharing β is silent by design; you only learn it's there by running the check. The most deceptive clips are precisely the ones stripped and cropped until they look native to your feed. A missing watermark is a prompt to verify harder, never a stamp of authenticity.
What this guide is not
This isn't a claim that Veo footage is inherently deceptive β most of it is people making short films and clearly-labeled fun. And these checks won't catch everything: a clip from a rival generator won't carry Google's SynthID at all, and no single test settles every case. The point is a repeatable habit for the moment a "real" video makes a big claim β read the markers if they're there, read the footage if they're not, and check the story either way before it travels any further on your say-so.
Check the claim in seconds
A viral clip making a big claim? FAXTR searches 100+ fact-checking organizations in one query β free, no login required.