techcrunch.com web signal

Google Vids Lets Users Star as AI Avatars via Gemini Omni

TL;DR

  • Google Vids will let users generate a personal AI avatar from a selfie and voice recording, powered by the new Gemini Omni multimodal model.
  • Every Omni-generated video carries an invisible SynthID watermark, and avatars are tied to the account holder's Google account.
  • Personal avatars are limited to users 18 and older in select regions, and onboarding requires recording yourself speaking a series of numbers.

Google Vids is quietly turning into a deepfake studio that lives inside your Workspace tab. On Thursday, TechCrunch reported that Vids will let users create a personal AI avatar from a selfie and voice recording, then use that avatar to star in videos generated by Gemini Omni, Google's multimodal model. The same update lets Omni take step-by-step edits (background swaps, lighting fixes, effects added to phone-recorded footage) instead of forcing you to regenerate from scratch.

The onboarding is where the safety story lives. To spin up an avatar, users have to record themselves and speak out a series of numbers, which Google frames as a safeguard against people uploading someone else's face. Every avatar is tied to the account holder's Google account, and everything Omni generates carries an invisible SynthID watermark. Access is limited to users 18 and older in select regions.

The strategic read is that Google is dragging avatar generation from the standalone-startup market (HeyGen, Synthesia, Captions, D-ID) into the default Workspace toolset. If your HR or comms team is already paying one of those vendors to churn out training videos and localized company updates, a Vids-native avatar with the same watermarking story is going to force a procurement conversation you were probably going to have anyway.

The honest caveat is that a SynthID watermark is only as useful as the platforms that check for it, and TechCrunch's write-up does not spell out pricing, exactly which regions are in the initial rollout, or how long Google retains the voice-and-face training data used to build an avatar. The 'speak a series of numbers' step will slow down casual impersonation, but liveness checks have historically been an arms race rather than a fix.

Where this lands over the next quarter is on internal comms and L&D teams: the executive video update and the localized training module are the two use cases that get cheap first, and the AI-video vendors that do not already live inside a productivity suite are the ones with a harder sales conversation next quarter.