Meta Muse uses your interactions to train Meta's models unless you turn that off. Meta says the opt-out is in Muse's settings and that it strips "critical personally identifying information" before training. The training switch is only one of several privacy controls, though, and it doesn't touch memory, connector access or what Meta can see while running the service.
Does Muse train on my data by default?
Yes. CNBC reported at launch that Muse training is opt-out, and Meta's security post says the same. Your interactions train Meta's models unless you change the setting. Meta says it removes "critical personally identifying information" before the data is used. It doesn't define that phrase or list what counts.
This is the opposite default to most business plans. OpenAI says Business, Enterprise and Edu data isn't used for training by default, and Anthropic uses consumer chats only if the Model Improvement setting is on. Muse is a consumer product with a free tier, and Meta has chosen training on as the starting point.
How do I turn off Muse training?
Meta says the opt-out lives in Muse's settings, but it hasn't published a menu path we could confirm, so the steps below stay general.
- Open MuseUse the Muse app on iOS or Android, or sign in at muse.ai.
- Find privacy settingsLook for the model-training or data-use option in Settings.
- Turn training offSwitch off the option that lets Meta use your interactions to improve its models.
- Check each deviceIf you use Muse on several devices, confirm the setting shows off everywhere.
- Recheck after updatesSettings can move between app versions, so look again after a major update.
Muse logs in with a Meta account. Meta hasn't said whether the Muse setting is shared with other Meta AI products or kept separate, so if you also use Meta AI elsewhere, check its settings too.
What does the opt-out not cover?
The training switch answers one question: may Meta use your interactions to improve its models? Several other privacy questions sit apart from it.
| Control | What it does | Where it comes from |
|---|---|---|
| Training opt-out | Stops interactions being used to train Meta's models | CNBC; Meta security post |
| Memory "forget" | Removes specific things Muse remembers | Meta Help Center |
| Inspect, edit, download | Lets you view and export VM files and memory | Meta security post |
| Connector access levels | Sets read-only or read-and-send per app | Meta launch post |
| Disconnect a service | Cuts Muse off from a connected app | Meta launch post |
| Ad separation | Keeps chats and VM data out of ad systems | Meta launch post |
Memory
Muse remembers context from past conversations. Meta's help centre says you can tell it to "forget" specific things. The security post adds that you can inspect, edit and download the files in your Muse VM and your memory. That gives you more control than OpenAI dots, where individual memories can't be edited and the only way to clear them is to reset the dot. Our memory privacy comparison sets the agents side by side.
Connectors
When you connect an app, you choose how much Muse can do. Meta's example is reading your email versus also sending it. You can disconnect a service at any time. Disconnecting stops new access. Meta doesn't say whether it removes what Muse already learned from that app.
The email connector also filters out one-time codes, password-reset links and magic login links, so those never reach the agent. That's a security feature rather than a privacy setting, but it limits what sensitive mail Muse can read.
Are my Muse chats used for ads?
Meta's launch post says your conversations and VM data are not shared with its ad systems. The security post adds a caveat: browsing Muse does for you can still influence ads indirectly. Meta's example is a retailer showing you an Instagram ad after Muse visited its site. That happens through the retailer's own tracking, not through Muse handing your chat to Meta's ad tools.
If you want to limit this, the opt-out won't help. It's a training control, not an ad control. The practical step is to use Meta's general ad preferences and to be aware that a site Muse visits may treat that visit like any other.
Can Meta see my Muse data anyway?
Yes, in some cases. The security post says the current architecture "does not prevent Meta from accessing data when necessary to support, secure or operate the service." Your Muse VM is backed up continuously, and the opt-out doesn't change Meta's operational access.
Meta's planned answer is the Muse Confidential VM, which would encrypt the whole VM with a key only you hold. As of November 2026 it isn't generally available. Meta said "later this year" at launch, and TestingCatalog reports trusted testers have it. Until it ships, assume Meta can reach your Muse data when it decides it needs to.
- Muse launches in the US with training on by default
- Meta disputes a report that Muse read private Mac Messages
- Confidential VM still in trusted testing
The Messages dispute matters for trust. Inc. columnist Jason Aten said Muse read his Mac Messages with Full Disk Access off, apparently through notification syncing. Meta denied it, saying the feature needs opt-in, Full Disk Access and the Messages connector, and that Muse gave a wrong explanation of what it had done. The question was unresolved when reported. It's a reminder that an agent's account of its own actions isn't proof, and Meta's help centre itself warns that Muse "may be inaccurate or take unexpected actions."
What should I change on day one?
Pros
- Turn off training if you don't want your interactions used for Meta's models
- Connect apps read-only first and add send access only where you need it
- Tell Muse to forget anything sensitive it picked up
- Download your memory and VM files now and then to see what's stored
Cons
- Don't rely on disconnecting an app to erase what Muse learned
- Don't assume cancelling your plan deletes your data
Cancelling a paid plan is a billing change, not a deletion request. Our guide to cancelling a personal AI agent covers the cancel steps, and the Muse setup guide covers connector choices when you first sign in. For approvals and the separate Sentinel agent, see Muse Sentinel explained.
For a cross-agent walkthrough of training opt-outs, see safeaiagents.com, which also has a page on Muse safety settings.
Bottom line
Muse trains on your interactions by default, and Meta says you can switch that off in settings. Do it if you don't want your chats in Meta's training data, but don't treat it as a privacy reset. Memory, connector access and ad influence each need their own attention, Meta publishes no retention period, and Meta keeps operational access until the Confidential VM ships. Our Muse review tracks changes to these settings.


