Microsoft is testing a Copilot experiment that gives the AI a face — actually, 40 faces. The new feature, called Portraits, lets users select from a set of animated 3D avatars and hold voice conversations with them. According to a report by TestingCatalog, the avatars are powered by VASA-1, a real-time facial animation model from Microsoft Research. But there are significant catches: only users in the US, UK, and Canada will get access, they must be 18 or older, and conversations are capped at 20 minutes per day.

The Portraits Experiment: What Testers Are Seeing

The details come from an internal description that surfaced via testers. The feature lives in Copilot Labs, Microsoft’s sandbox for experimental AI tools. Unlike static profile pictures, these avatars react visually and emotionally in sync with spoken conversation. TestingCatalog reports that 40 distinct portraits will be available, all in a non-photorealistic, cartoonish or 3D style. Users can pair an avatar with a choice of voices.

The engine under the hood is VASA-1, which Microsoft Research unveiled as a model capable of generating lifelike talking faces from a single image and an audio clip. It handles lip sync, head movements, and micro-expressions at interactive frame rates — up to 40 frames per second at 512×512 resolution. That technical prowess is what makes a real-time voice chat with an animated avatar feel fluid enough for practical use.

Microsoft is targeting specific learning and rehearsal tasks: practicing difficult conversations, preparing for public speaking, role-playing job interviews, and even study sessions. Testing notes hint at a “study mode voice” flag for educational contexts. This isn’t just a toy; Microsoft envisions it as a coaching and productivity tool.

Who Gets Access — and the Guardrails

The rollout follows Copilot Labs’ traditional pattern: phased, regional, and limited. For now, only users in the United States, United Kingdom, and Canada are eligible, and even then not all at once. The feature is gated behind an 18+ age requirement, and the reported 20-minute daily cap on portrait conversations is likely designed to manage both health considerations and computational cost.

These guardrails are notable. Real-time animation of a high-fidelity talking avatar is expensive on server GPUs. A per-user time limit makes sense while Microsoft assesses demand and infrastructure. It also serves as a safety measure: limiting continuous use reduces the risk of over-reliance or misuse during this experimental phase. The non-photorealistic style further mitigates impersonation risks — you won’t mistake a cartoon avatar for a real person.

All of these specifics — the 40-portrait count, the 20-minute cap, the exact regional gating — come from a single testing report and have not been officially confirmed by Microsoft. Treat them as credible but provisional until the company posts a formal announcement.

Why Microsoft Is Being So Cautious

The caution is rooted in both technical and ethical realities. VASA-1’s ability to generate convincing facial animation from just audio and a single image is exactly the technology that fuels deepfake concerns. Microsoft’s own research team warned about the potential for impersonation and misuse when the model was first released as a research demonstration. By enforcing non-photorealistic styling, age gating, and time limits, Microsoft is trying to make the feature engaging but safe — keeping it as a coaching tool rather than a deepfake engine.

The Copilot Labs context provides another layer of control. Experiments in Labs are temporary playgrounds where Microsoft collects telemetry and user feedback before deciding whether to graduate a feature. Previous Labs entries like Copilot Vision or 3D image generation followed a similar pattern: limited rollouts, short retention windows for generated content, and clear usage terms. Portraits inherits that tested-but-not-production nature.

Privacy also looms large. Voice chats with an animated avatar could generate a trove of data: audio streams, emotional affect estimates, usage patterns. Microsoft has not yet published a data-use policy specific to Portraits. In the absence of documentation, users should assume that conversations are monitored for quality and safety, possibly retained short-term, and potentially used for model improvement — unless explicitly stated otherwise.

What It Means for You, Depending on Who You Are

Everyday Windows users: If you’re in the US, UK, or Canada and over 18, you might spot Portraits in your Copilot sidebar or Labs menu when it rolls out. It’s an optional experiment — you don’t have to use it. If you’re curious, try it for interview prep or a speech rehearsal. The 20-minute limit means you won’t lose hours to it. Outside those regions, you’ll have to wait; there’s no indication of broader release yet.

Power users and early adopters: This is a chance to test a genuinely novel multimodal interaction. VASA-1’s expressiveness could change how you perceive an AI assistant’s empathy and coaching ability. But keep in mind that the feature is a Labs experiment, not a finished product. Expect glitches, occasional lag, and possibly awkward animation artifacts. Provide feedback through official channels — it’s the whole point of Labs.

IT professionals and admins: For organizations managing Windows devices, Copilot Labs features are typically not available to commercial users or can be restricted via policy. Portraits is unlikely to appear on enterprise tenants during the Lab phase. Still, its existence signals where Microsoft is heading: audiovisual AI assistants that are deeply integrated into the OS. Consider the data-handling implications now. If and when such features graduate to broader release, you’ll need clear answers about what audio is stored, how identity is verified, and whether the model can be trained on user interactions. Watch for Microsoft 365 admin documentation.

Developers: VASA-1 is a research model, not currently a public API. But the product integration hints at future tooling possibilities — perhaps Microsoft will eventually offer avatar animation as a service within Azure or for Copilot extensions. For now, it’s a showcase of what the tech can do, not a dev platform.

How We Got Here: The Road to Animated Copilot Avatars

Copilot started as a text-centric assistant. Over the past two years, Microsoft bolted on image generation (DALL‑E integration), vision capabilities (Copilot Vision), and 3D object creation. The Labs program was the proving ground for each of these expansions. Portraits is the next logical step: coupling a voice mode with visual feedback, moving from one-way multimodal (you ask, it shows) to synchronous, conversational interaction.

VASA-1 was published in early 2024 by Microsoft Research as a breakthrough in audio-driven facial animation. Unlike earlier systems that only synced lip movements, VASA-1 generated full expressive sequences — head tilts, eye blinks, subtle emotional shifts — from a static portrait and an audio track. The research team demonstrated the model’s ability to animate paintings, drawings, and even stylized avatars, all with real-time performance. That flexibility made it an ideal candidate for Copilot’s Labs, where non-photorealistic styles could be used without triggering deepfake concerns.

Meanwhile, industry competition has been simmering. Google Labs launched a “Portraits” experiment with AI coaching from verified experts, but not real-time animation. Startups and Chinese tech giants like Alibaba demonstrated similar face-synthesis models. Microsoft’s advantage is owning the entire stack: the Copilot platform, the Windows user base, and a research arm producing state-of-the-art animation.

What to Do When Portraits Appears in Your Copilot

First, confirm it’s in your Labs menu. If you’re in an eligible region and over 18, you may see a new card or prompt. Before you dive in:

  • Treat it as experimental — avoid sharing sensitive personal information, financial details, or anything you wouldn’t say aloud in a public beta.
  • If you produce any practice recordings or transcripts you want to keep, export them immediately. Labs features have a history of short retention windows; your data might vanish after the experiment ends or be purged regularly.
  • Expect the 20-minute limit. Plan your practice sessions accordingly. If you hit the cap, you’ll need to wait until the next day (or whatever reset period Microsoft uses) to continue.
  • Watch for a consent screen. Microsoft will likely present terms of service that clarify data handling. Read them. If they mention model training on your conversations, consider whether you’re comfortable with that.
  • Give feedback. Microsoft’s Labs program relies on user input. If the avatar feels uncanny, the lip sync breaks, or you encounter a safety gap, report it.

The Outlook: What to Watch Next

Official confirmation is the next milestone. Look for a Microsoft blog post or a Copilot Labs update announcing Portraits publicly. That should clear up the exact number of avatars, regional availability, and whether the 20-minute limit is a temporary research constraint or a permanent design choice.

Shortly after, expect hands-on demos or video previews from early testers. These will reveal how polished the animation really is — whether the avatars achieve truly synchronous lip sync and natural head movement, or if latency makes them feel like marionettes on a delay.

Privacy and data documentation will be critical. If Microsoft publishes a support page detailing retention, telemetry, and consent, that will be a strong signal the feature is moving toward general availability. Absent that, skepticism is warranted.

Longer term, Portraits could graduate from Labs into a standard Copilot feature, perhaps integrated into Microsoft 365 for coaching scenarios in Teams, PowerPoint rehearsal, or language learning. But that depends on the feedback, the cost of scaling real-time VASA-1 inference, and how well Microsoft navigates the ethical tightrope of animated AI avatars.

For now, the message is clear: Copilot is getting a face, but only for those who are patient enough to chat with it in 20-minute bursts, and only in a style that makes it unmistakably artificial. The experiment is as much a test of our comfort with expressive AI as it is a test of the technology itself. Watch the Labs announcements, try it if you can, and decide for yourself whether an emotionally responsive avatar makes an AI assistant more helpful — or just more unnerving.