Is Gemini Live Avatar safe? Faces, consent and SynthID

Short answer: Gemini Live Avatar is built with more care than most avatar tools, and the risk it creates mostly lands on the person talking to it rather than the company deploying it. Google restricts who can make a custom face and watermarks everything the avatar says and shows. Neither control tells a customer, in the moment, that the friendly face on the screen is software. That part is still up to the business, and to you.

Google introduced Gemini 3.8 Live with Live Avatar on 24 September 2026, and Google Cloud says it is generally available in Gemini Enterprise, with US and EU endpoints. It is a tool for businesses, not a feature in the consumer Gemini app. Everything below is as of 26 September 2026.

What Live Avatar actually is

Gemini 3.8 Live is Google’s model for real-time spoken conversation; the audio-only version arrived about a week earlier, according to TestingCatalog. Live Avatar adds a generated face on top, with lip movements and expressions synced to the speech in near real time. Google says it understands and speaks 97 languages, can watch a camera feed or a shared screen while it listens, and can call tools in the background without pausing the conversation.

The intended use is customer service, product walkthroughs and kiosks. Google names Salesforce and Cox Automotive among the early customers. So the realistic way most people will meet it is on a company’s website or support line, not in an app they chose to open.

The two safeguards Google built in

• —Faces are gated. Businesses pick from a library of preset avatars. Building a custom avatar from a reference image, which is where a real person’s likeness could be copied, is only available after an enterprise allowlisting and verification process.
• —Output is watermarked. Every audio and video stream carries an imperceptible SynthID watermark, so the content can later be identified as AI-generated by tools that check for it.

Both are real and worth having. The allowlist means a random account cannot upload a photo of a colleague and have it talking by lunchtime. The watermark means a recording of the avatar can be traced back as synthetic, which matters if a clip gets passed around out of context.

What those safeguards do not cover

A watermark helps the person who checks for it. It does nothing for a customer on a video call who has no reason to run a check and no tool to run it with. The honest risk here is not a stranger cloning your face. It is being persuaded by a confident, well-lit face that happens to be wrong.

Language models state wrong things fluently. That was true in a chat window and it stays true with a face attached, except that a face makes the wrong answer more believable. A support avatar that misstates a cancellation deadline will do it while nodding. The method in how to check an AI answer when you are not the expert applies just as much when the answer is spoken to you.

The reaction has not been universally warm, either. Engadget called the avatars creepy in its headline, which says something about how far realistic faces still sit from being trusted.

Disclosure is the business’s job

Whether customers are told they are talking to an AI is decided by the company that deploys the avatar, not by Google. In the EU the rule is written down: Article 50 of the AI Act requires, among other things, that people are informed when they interact with an AI system. Outside the EU, it is good practice rather than a clear legal duty everywhere, which is exactly why it is worth asking.

If you run a support team and are weighing this up, start without the face. Write down the ten questions customers actually ask, test a plain text bot against them, and only add voice and video once the answers hold up. Generative AI customer service: a worked example walks through that kind of test. An avatar will not fix a bot that gives three different refund amounts to the same question. It will only make the inconsistency harder to spot.

A worked example

Say your bank’s website offers a video assistant to help you dispute a card charge. The avatar is friendly, explains the process clearly and tells you the dispute window is 90 days. Before you rely on that:

1. 01Ask directly: “Am I talking to an AI?” A well-run service will say yes without hedging.
2. 02Ask where the 90 days comes from, and request the policy link or a written confirmation by email.
3. 03Check that figure against the bank’s own terms page, not the assistant’s summary of it.
4. 04If the assistant asks to turn on your camera for a task that does not need it, decline. Explaining a form works fine with the camera off.

None of that is specific to Google. It is the same routine as with any agent that acts on your behalf, and the difference between an agent and a plain model is laid out in AI agent vs LLM.

Faces and voices are no longer proof of anything

This launch landed a day after Google’s new speech models, which can copy a voice from a short sample under a consent check; is Gemini 3.8 TTS safe covers that side. Put the two together and a convincing face and a convincing voice are both off-the-shelf enterprise features. The practical rule for anyone at work follows from that: a face on a screen or a voice on a call is not identity verification. Payment changes, password resets and urgent requests still need a call back on a number you already have.

Is it safe to use?

• —For a business: reasonably, if you use the preset avatars, tell customers they are talking to an AI, and test the answers as text first. Pricing sits on the platform pricing page and data residency choices in the Live API documentation, both of which your security team should read before anything else.
• —For a customer: yes, with the same scepticism you would give a chatbot. Get anything important in writing.
• —For your own likeness: the allowlist is a meaningful barrier, but it guards this one product. It is not a general protection against deepfakes made elsewhere.

The same questions came up with Meta’s agent, and is Meta Muse safe works through them for a tool that acts on your accounts. How Google’s wider line-up compares with Meta’s is in Muse Spark vs Gemini.

With a face or without one, the skill that matters is the same: ask a clear question and check the answer before you act on it. That is what Coursium teaches, on your phone.