What changes when AI starts looking human?
A face can make a conversation feel more natural. It does not make the system more capable, better informed or accountable. If an organisation gives AI a human-like interface, it also needs to design the expectations that come with it.
A more human interface changes the interaction, not the intelligence
A text box asks people to read and type. A speaking face adds voice, expression and a social presence. That may suit a guided product demonstration, a multilingual welcome or training where listening and watching is easier than navigating a long page. It may also prompt people to read warmth or confidence as signs of understanding.
In a 2024 experiment with 2,165 US participants, Cohn and colleagues compared a pseudo-LLM presented as text or as speech plus text. The speech-plus-text condition received higher ratings for anthropomorphism and perceived accuracy. The study did not test generated video avatars, real business services or objective accuracy. It is a reason to test interface effects, not proof that every user will over-trust a face. Read the study in the ACM Digital Library.
Start with the user problem, not the avatar
A small training provider might use a video guide to walk new employees through a safety process in several languages. The useful question is whether the visual format helps people understand and complete that process. If it adds delay, distraction or misplaced confidence, a narrated guide or well-designed text interface may work better. A face is not a substitute for reliable service, accessible alternatives or a route to a person.
Transparency should be understandable to the user
Google Cloud says Gemini 3.8 Live with Live Avatar became generally available for enterprise use on 24 September 2026. Google describes synchronised speech and video, support for 97 languages, pre-built avatars and custom-avatar creation subject to allowlisting and verification. It also says generated audio and video carry imperceptible SynthID watermarks. These are product safeguards and provenance measures, not a replacement for clear disclosure in the experience. An imperceptible watermark cannot tell a user, in the moment, that they are speaking to AI. See Google Cloud’s availability and safeguard details.
Identify the system plainly in the interface and audio, avoid implying a real employee is present, and make the human handover route easy to find. If the system handles a sensitive or consequential request, explain what it can do before collecting information.
Review the interface as part of the AI system
- Who is using it, and what task should a face make easier?
- Can people tell that they are interacting with AI before sharing information or acting on advice?
- Can the system show evidence, express uncertainty and distinguish facts from suggestions?
- What can it influence, what may it collect and when is a person available?
- Will you measure task completion, comprehension, errors and appropriate trust—not only session length?
Compare the avatar with a text or voice-only version. Ask users what they think the system can do, whether they know it is AI, which information they relied on and how certain they feel. Track task completion, corrections, complaints and handovers as well as satisfaction. If confidence rises without better outcomes, change the interface or workflow.
A face is not a trust policy
The decision is whether this interface solves a real user problem, tells the truth about what is behind it and helps people judge the answer on evidence. For how much to check an AI recommendation, see The more authority you give AI, the more verification matters. For use-case selection, read Stop asking what AI can write. Start asking what problems it can solve..
The more human AI looks, the more important it is to make clear that the system is still software—with limits, evidence and a person accountable for the service.