Is Muse Realtime Avatar Safe? Voice, Faces and Data

Muse Realtime Avatar is the talking, moving face Meta added to its Muse app. Meta started rolling it out on 5 November 2026, together with real-time voice, in the US first. Is it safe? It is as safe as the Muse app it lives in, plus one new question: what happens to your voice and to any image you give the avatar. Meta has documented how the technology works. It has not, as of 8 November 2026, published how long it keeps voice conversations or reference images.

This post sets out what Meta has said, what it has not, and a short setup routine. We have not used the feature ourselves; it is US-only for now. Everything below comes from Meta’s own posts and launch coverage, checked on 8 November 2026.

What Meta actually launched

Two pieces of technology, shipped as one feature. Muse Realtime Voice lets you talk to your Muse and interrupt it, the way you would a person. Muse Realtime Avatar turns that voice into a face that reacts while you speak. In the app you tap the phone icon and choose how present the avatar is: voice only, an avatar beside the chat, or full screen. Meta says there are hundreds of voices to pick from, and that voice is coming later to its AI glasses and the Muse Charm device.

The technology itself is older than the rollout. Meta’s research team introduced Muse Realtime Avatar on 23 September 2026, at Connect, as a preview. That post is worth reading for one detail that matters for privacy: the avatar is “conditioned on reference media”. In plain words, it animates a picture. Meta’s examples include a photographic portrait, an illustration, an animal and an everyday object. The research post says it streams portrait video at 448 by 768 pixels and 25 frames a second, with about 870 milliseconds from the end of your turn to the reply.

If you are new to Muse itself, it is Meta’s agent app, built on the Muse Spark model. The difference between the two is covered in what Muse Spark is, and the app setup in the Meta Muse app guide.

What Meta has said about your data

Not much that is specific to voice or avatars. The 5 November announcement covers what the feature does and where it is available. It does not mention recordings, transcripts, retention or human review. Gadget Review’s launch report noticed the same gaps: no word on data retention, review practices or parental controls.

What does exist is the Muse app’s general position. When Meta launched Muse in September, it said you can opt out of your interactions being used to train its models, and that conversations are kept away from its ad systems. Those are Meta’s own statements, not an audit. Above them sits the Meta Privacy Policy, which is long and general and covers all of Meta’s products. It does not answer the voice question for this feature in particular.

So the honest summary is this. A voice conversation is still a conversation with Muse, and the September rules presumably apply to it. “Presumably” is doing work in that sentence. Until Meta says how it handles audio and reference images, treat both as data you are handing over and cannot easily take back.

The three risks worth weighing

First, your voice is personal data in a way text is not. It identifies you, it carries mood and health signals, and it is what someone needs to imitate you. Speaking to Muse sends that to Meta. That is no different from any voice assistant, but it is worth deciding on purpose rather than by habit.

Second, the avatar works from an image. If the app lets you choose or upload one, a photo of a real person is that person’s data, not yours. Do not animate a colleague, a child or a public figure without their permission. Meta’s research post shows what the model can do with a portrait; it does not say what the consumer app allows or blocks.

Third, a face changes how much you trust what you hear. A friendly avatar that nods while you talk feels more reliable than a text box. It is not. It runs the same model, with the same habit of sounding confident when it is wrong. If you use voice to rehearse an interview or talk through a decision, the advice still needs checking, the way how to check an AI answer when you are not the expert describes.

How it compares with Google’s avatar

Google got to live avatars first, for developers rather than consumers. Is Gemini Live Avatar safe covers that launch, where the main questions were consent for the face and what businesses must disclose to the people talking to it. Google also documents a watermark on its generated media. For Meta’s avatar we found no equivalent statement in the launch posts, so we make no claim either way. The same consent rules for faces apply to Google’s new video model, set out in is Gemini Omni 1.2 safe.

A setup routine that keeps the risk small

1. 01Check the training opt-out in the Muse app before your first voice call, not after.
2. 02Start with voice only. Add the avatar once you know what you are getting from voice.
3. 03Use one of Meta’s own avatars or an illustration, not a photo of a real person.
4. 04Keep voice calls away from other people’s private details: clients, patients, colleagues.
5. 05Do not use it on a work account or for work data unless your employer has said yes in writing.
6. 06Treat anything it tells you out loud as a draft to check, especially facts, figures and dates.
7. 07Recheck Meta’s posts in a few weeks. A retention policy for voice may well appear, and it changes this advice.

The wider point about agents, which is what Muse is underneath the face, is in is Meta Muse safe: give the least access that does the job, and read every approval. Voice makes it easier to say yes without reading. Slow down when Muse asks to send, buy or book.

If your organisation is deciding whether to allow tools like this, the NIST AI Risk Management Framework gives a sensible order: map where it is used, measure what could go wrong, then manage the risks you accept. For a voice avatar, the first questions are whose voice and whose face, and where the recordings go.

Is it worth trying?

For thinking out loud, probably. Talking through a half-formed plan is faster than typing it, and rehearsing an interview answer out loud is genuinely useful. The avatar adds less than the voice does, and it is the part with the open questions. Many people will be fine with voice only.

Whatever the interface, the skill that matters is the same: give clear instructions and check what comes back. Coursium teaches people to use AI at work, including how to spot a confident wrong answer. You can see how Coursium works or try the free practice example.