AI is learning to perform presence.
Google’s latest audio and avatar tools point to a new creative constraint: the interface can feel human, but the operating system behind it still needs to be unmistakably yours.
The next AI interface is not just a chatbot with a face. It is a directed performance.
Google announced Gemini 3.8 Flash TTS on September 23 with custom voice design, line-by-line direction, two-speaker scene staging and long-form generation. Google says creators can design voices across more than 100 languages and dialects, use 2,000+ production-ready voices, and replicate a voice from a 30-second sample when they have the rights and consent.
The following day, Google introduced Gemini 3.8 Live with Live Avatar: near-real-time video paired with speech, expressive lip-sync, asynchronous tool calls and multilingual switching across 97 languages. It is available in Gemini Enterprise; custom avatar creation is currently enterprise-allowlisted.
It is a better point of view.
Style becomes infrastructure.
For creators, the breakthrough is not that a machine can sound expressive. That was already obvious. The important shift is that voice, pacing, character, visual identity and conversational timing are becoming reusable production assets.
That lowers the cost of making many versions of an idea: a founder briefing in English and Thai, a product walkthrough with a consistent narrator, or a short-form series whose host can answer questions without a camera crew. But it also raises the cost of being generic. If everyone can generate a polished voice and face, “polished” stops being differentiation.
The business question is therefore not “Can we make this sound real?” It is “What should this presence help a person do?” A tutorial can optimise for clarity; a brand character can optimise for recognition; a support avatar can optimise for speed and escalation. Those are different jobs, with different scripts, metrics and failure modes. Treating them as one universal persona is how teams accidentally produce content that is technically fluent but strategically empty.
For a small team, this suggests a useful division of labour. Let the model handle variation and performance; keep editorial judgment, customer truth and escalation with people. Build a small library of approved pronunciations, claims and examples so every new output starts from a known standard rather than a blank prompt.
JUFU inference: the winning creator workflow will treat AI presence like a brand system, not a novelty effect. Define the point of view, approval boundaries and recurring formats first; select the voice or avatar second.
Direct the performance.
What must remain recognisably yours?
Write the tone, vocabulary, pacing and things the persona will never claim.
What deserves line-level direction?
Use stage cues for emphasis, pauses, dialect, reactions and turn-taking—not one giant prompt.
What can you prove?
Keep consent, likeness permissions, provenance and disclosure attached to every reusable asset.
Where does a human decide?
Approve claims, sensitive answers, translations and anything that could be mistaken for a real person.
One format. Three surfaces.
Take one useful founder or creator idea and produce three versions: a narrated short, a two-speaker dialogue, and a live Q&A prototype. Keep the message constant. Change only the performance layer. Measure completion, comprehension and trust—not just whether the output looks impressive.
Google’s own launch materials emphasise watermarking and provenance controls, including SynthID and C2PA for voice replication. Those are useful rails, not a substitute for disclosure. Make “AI-assisted” part of the experience design wherever a reasonable viewer could otherwise infer a human recording.
Own the point of view. Rent the performance layer.
AI can now supply voices, faces and conversational motion at unusual scale. Your durable advantage is deciding what the presence stands for, what it is allowed to say, and when a human must take over.
JUFU DAILY / RESEARCH NOTE