AI is getting a face. Will that make us trust it too much?

A spoken AI assistant can already sound fluent. Google’s new Live Avatar adds something people instinctively read as social: a face that moves while it talks.

Google announced Gemini 3.8 Live with Live Avatar on 24 September 2026. In plain English, it combines spoken, speech-to-speech conversation with generated video streamed at low latency. The on-screen avatar is designed to move its mouth in time with speech, show facial expressions and take turns in a conversation. The system can also receive live audio and visual input.

Google says synchronized speech and expression are supported in 97 languages. That is a product claim from Google, not an independent test of quality across every language, accent or business setting.

This is an interface, not a digital employee

The technology is more than a pre-recorded character placed beside a chatbot. Google describes generated video and conversation working together in real time. A business could build an avatar into a reception screen, a product guide, a training exercise or a multilingual information service. Those are possible developer-built uses; the announcement does not mean Gemini Enterprise arrives with a ready-made receptionist that can safely run a company’s front desk.

Availability matters too. Google describes Live Avatar as generally available in Gemini Enterprise and through its Live API, with US and EU endpoints. That does not make it a universal feature in the consumer Gemini app, and an available endpoint is not proof that every UK small business account is eligible. Custom avatars based on a reference image require enterprise allowlisting and a verification process.

Why a face could change the conversation

A face gives people extra cues to interpret: eye contact, expression, warmth and apparent confidence. It may make an interaction easier or less intimidating. It may also encourage a user to treat a fluent answer as if it came from a person who understands the consequences.

A perspective in the Proceedings of the National Academy of Sciences discusses both the potential benefits and dangers of anthropomorphic conversational agents. That is a useful general warning, not a study of Google’s Live Avatar and not proof that adding a face always increases trust. People and contexts differ. A smile on a screen does not improve a model’s accuracy, and an expressive voice does not create professional judgement.

What businesses should decide before using one

If a business builds or trials an avatar, it should decide what the system may answer, which information it can see and where a human takes over. A product guide can explain opening hours or help someone find a manual. It should not quietly become the only route for complaints, unusual requests or decisions with financial consequences.

• Tell people plainly when they are speaking with AI; do not rely on a tiny label they will miss.
• Give the avatar a limited role and a clear way to hand the conversation to a person.
• Check important answers against approved information instead of treating confident delivery as proof.
• Decide what conversation data is collected and who can access it before connecting the system to business records.
• Try the experience with the people who will use it, including those who may find a face or voice less accessible or comfortable.

There is a related question in workplace training: an avatar can make roleplay more engaging, but recording or scoring employees changes the character of the exercise. IT Club has looked at that issue in its article about AI coaching at work.

Would You Let an AI Coach Your Employees? →

What the watermark can—and cannot—tell you

Google says Live Avatar audio and video streams carry imperceptible SynthID watermarks. Watermarking can help identify AI-generated media in supported verification workflows. It is not a guarantee that every copy will remain detectable after every edit, nor does it certify that the words spoken are true or that a depicted person approved the interaction. Google’s custom-avatar allowlisting is also a vendor-described safeguard, not evidence that all identity misuse is impossible.

For a business, disclosure is the simpler starting point: say clearly that the conversation is with an AI, explain what it can do, and do not design the experience to trick somebody into thinking a real employee is on screen.

The useful question is not whether an avatar looks convincing. It is whether the system makes a real task easier while staying understandable, limited and easy to question. The interface may become more human. The need to check important answers does not.

Sources and further reading

Google’s descriptions of product capability and safeguards are vendor claims. The research paper below discusses anthropomorphic AI generally; it did not test Gemini Live Avatar.

Google: Introducing Gemini 3.8 Live with Live Avatar →

Google Cloud: Gemini 3.8 Live with Live Avatar →

Google DeepMind: SynthID →

PNAS: The benefits and dangers of anthropomorphic conversational agents →

Found this useful? Forward it to someone who might too.