
Muse’s growth puts non-users in the frame
Meta’s Muse AI assistant has crossed a significant adoption threshold: more than 10 million users, according to the company, with roughly one in four using it daily. That scale matters not only to people who have chosen to use the assistant but also to the friends, family members, colleagues, and clients who appear in their messages, calendars, and contact lists. Muse is marketed as an agentic assistant that can handle tasks across email, calendars, and other services, but the data it touches often belongs to someone else.
The core privacy question is not whether Muse users have opted in. It is what happens to information about people who never installed the app, never created a Muse account, and may not even know the product exists. As AI assistants move from answering questions to taking actions, the line between a user’s data and a bystander’s data becomes harder to see.
What Muse can access — and what it asks for
Muse launched on September 8 as Meta’s entry into the personal AI assistant market. It is promoted as an assistant that can take busywork off a user’s plate and connect to email, calendars, and other services to complete tasks. In practice, that means Muse can request broad permissions and access to sensitive parts of a user’s digital life.
On mobile, its app listing indicates that it may collect a wide range of data categories, including:
- Photos and videos
- Financial information
- App information and performance data
- Files and documents
- App activity
- Personal information such as name, email address, user IDs, address, phone number, and other details
- Contacts
- Web browsing activity
- Health and fitness data
- Messages, including emails, SMS or MMS, and other in-app messages
- Audio
- Location
- Calendar and calendar events
The list is expansive because the assistant is designed to act, not just chat. It can work around the clock, interact with a browser, and access services that hold sensitive information. It also enables AI model training by default unless a user changes settings, adding another layer of data use that non-users cannot directly control.
How non-users’ data reaches Muse
Meta says Muse can gather information about non-users in several ways, including when a user connects the assistant to apps and services that contain data about other people. A user may also submit information about themselves and others in the form of text, images, videos, or uploaded documents.
Meta’s own example is ordinary enough: if a Muse user connects a calendar app and asks what meetings are scheduled for the week, the names of attendees and event details could be included in both the request and the response. That is a simple scheduling task, but it pulls third-party names, plans, and relationships into an AI system.
Connectors extend the agent’s reach
Meta calls the bridges between Muse and third-party services Connectors. Facebook, Instagram, and Threads are connected automatically when a Muse account uses the same credentials as the Meta Account Center. On launch day, Meta said Connectors also include Gmail, Google Calendar, Outlook, Plaid, OpenTable, Google Docs, Spotify, Function Health, Withings, Tailscale, and Peloton. If a connector is not available, a user can ask Muse to make an API call and try to create one to another service on their behalf.
That architecture is central to the privacy concern. Each connection can expand the surface area of data that the assistant can read, summarize, or act on. A user may grant access for a legitimate purpose, such as managing a calendar or drafting replies, but the data pulled in may include details about other people who did not grant anything.
Conversations can expose third parties
If a Muse user has granted permission and asked the assistant to summarize a message thread or craft a response, it may read content in a messaging app, an email service, or another communication platform that includes sensitive information belonging to someone else. Travel plans, medical issues, job discussions, financial concerns, and family logistics can all be processed by an assistant that a non-user did not consent to and does not use.
Meta has said that because the person using Muse chooses what they share and what they request, the company cannot predict exactly what information about a non-user may be included. That is technically true, but it also highlights the absence of a direct consent mechanism for the people whose data is swept into the assistant’s context.
What Meta says about privacy and identification
Meta does not promise full anonymization. Instead, it says that if a non-user’s data is viewed or ingested by Muse, it will not be linked to a Meta account and will not be stored in a way that allows Meta to reasonably identify which information belongs to a particular person. The standard is reasonable identification, not absolute anonymity.
The distinction matters. Data that is not linked to an account can still be sensitive, and it can still be used in aggregate or to improve models. It can also be re-identified in some circumstances, especially when combined with other data. For a non-user, the practical question is not only whether Meta can identify them today but whether the data can be used, retained, or combined in ways they did not anticipate.
Confidential VM still unavailable
Meta’s privacy policy says Muse currently operates in a cloud-based virtual machine. In the future, the company says there will be an option for a confidential VM that keeps people’s information private even from Meta. That private environment is not available now, which means the current architecture depends on Meta’s cloud and its data handling rules.
AI training and opt-outs
Interactions may be used to improve Meta’s AI training. Users can turn this off under settings and data controls. Meta says that if users allow training, it removes certain categories of personally identifiable information, such as names, email addresses, phone numbers, and Social Security numbers, and disassociates interactions from the account. Those are meaningful safeguards, but they are not the same as giving non-users a veto. The opt-out belongs to the Muse user, not to the person whose message was summarized.
Why non-users have few options
Muse users can adjust permissions, turn off training, and decide which connectors to enable. Non-users generally cannot. They may have rights under local data privacy laws to object to processing, request access, or ask for deletion, but those rights can be difficult to exercise when the data sits inside another person’s assistant conversation or a cloud-based virtual machine. The law may also not apply to every Muse interaction, depending on jurisdiction and the legal basis for processing.
Privacy experts have noted that asking a friend or colleague not to share personal information may work in individual cases, but it places the burden on the relationship rather than the platform. The principle is simple: platforms should protect people who do not consent to data collection by default. When an AI assistant is designed to ingest third-party data, that responsibility cannot rest solely on users remembering what they should not paste into a chat.
The broader context: agentic AI and bystander privacy
Muse is part of a larger shift toward agentic AI. Earlier generations of chatbots mostly responded to prompts with text. Agentic assistants are designed to connect to tools, read documents, manage schedules, send messages, and complete multi-step tasks. That capability is useful, but it changes the privacy model. A chatbot that only sees what a user types is one thing. An assistant that can read email, calendars, contacts, and files is another.
The risk is not hypothetical. Many people already forward emails, share screenshots, and discuss personal matters in digital channels. When one participant brings an AI assistant into that exchange, the assistant may process information about everyone involved. The non-user may never see a permission screen, may never receive a notice, and may never know which conversation was summarized or which attachment was uploaded.
Data minimization, purpose limitation, and transparency are the usual principles that privacy regulators use to address this kind of imbalance. The challenge with AI agents is that the data flow is dynamic. A user can connect a new service, ask a new question, or request an API integration at any time. The assistant may pull in data that the platform did not specifically anticipate. That makes default protections and clear boundaries more important, not less.
There is also the question of retention. Even if Meta does not link non-user data to an account, the data may still be stored for some period, processed by models, or included in logs. The company’s language about reasonable identification may be legally carefully worded, but it does not answer every practical concern. Non-users want to know what was collected, why, how long it was kept, and whether it was used for training.
The next frontier: AI phone calls
Meta is already experimenting with AI making phone calls to businesses on behalf of Muse users. Business owners can request not to receive AI calls, according to the company. That detail is revealing. It shows that the next phase of agentic AI will not stay inside a user’s inbox or calendar. It will reach into phone lines, customer service queues, reservations, and other interactions where the person on the other end may have no idea they are speaking to an AI agent.
The same consent-by-association problem will follow. A business may be told that the call is automated, or it may not. A receptionist may be asked to confirm a booking, change an appointment, or provide information. The data may belong to the business, the customer, or both. The assistant’s user may have chosen the convenience, but the other party has not chosen the processing.
For now, Muse’s most immediate privacy issue is simpler: a user’s connected apps can bring other people’s messages, names, schedules, and personal details into an AI system that those people never agreed to use. Meta says the data is not stored in a way that allows reasonable identification. Non-users have limited means to verify that, object to it, or opt out. As AI assistants become more capable and more deeply connected, that gap between user consent and bystander protection will be one of the defining privacy debates of the agentic era.
Source:ZDNET News
