Meta’s Muse AI Can Handle Email and Shopping — and Exposes How Much It Knows About You

Updated September 10, 2026 6:53 pm

Meta’s new Muse assistant is the company’s first serious attempt at an AI productivity tool, but a hands-on test showed that the bigger story is not what it can do — it is how much it can infer about a user from Instagram, Facebook, email, and shopping data. The assistant can sort inboxes, shop for products, and generate summaries and media, but its access to personal information raises immediate privacy questions.

In testing, Muse successfully deleted promotional emails, bought workout clothes through Amazon, and assembled a personalized feed of current events. But it also surfaced a highly specific profile of the tester’s interests that went well beyond what is visible inside Instagram’s own settings, underscoring why trust may become Meta’s toughest challenge as it pushes deeper into agentic AI.

What Meta’s Muse is trying to become

Muse is Meta’s answer to the growing wave of AI agents that can do more than chat. Instead of simply responding to prompts, the assistant is designed to take actions on a user’s behalf inside connected services, using a cloud-based virtual computer to complete tasks across the web.

Meta says the tool is meant to remove routine work from users’ schedules. That includes online shopping, email cleanup, trip planning, and other repetitive chores that are easy to describe but time-consuming to do manually.

The launch matters because Meta has spent years building products that reward engagement and entertainment. Muse signals a more direct push into productivity, a market that is becoming increasingly crowded as major tech companies race to make AI assistants useful enough to become part of everyday workflows.

Why this launch is different for Meta

This is Meta’s first clear move toward an AI agent built around action rather than conversation. That shift is important because it changes the product from something that merely answers questions into something that can access services, move through webpages, and complete transactions.

That also means Muse sits closer to a user’s most sensitive digital spaces. Once an assistant can read messages, connect accounts, and make purchases, convenience and risk become tightly linked.

How Muse handled my inbox and shopping tasks

To test Muse, the assistant was first asked to clean up a Gmail inbox by deleting messages that were no longer needed. That required connecting the tool to a Google account and granting it permission to read and delete mail.

The setup was not seamless. On mobile, the Google login process repeatedly failed and bounced back to the Muse website instead of completing inside the app. The process eventually worked on a laptop, where Muse deleted thousands of promotional messages and routine update emails.

Functionally, the task was a success. Emotionally, it was less straightforward. Allowing a Meta product to inspect an entire inbox is a major trust leap, especially for users already wary of how the company has handled data over the years.

Meta says Muse is built to share only the information necessary for connected third-party apps to function, and the company says it does not hand user data to advertisers. Even so, the experience of giving the assistant broad access to email and shopping accounts can still feel intrusive.

The Amazon test went more smoothly. After connecting an account, Muse was asked to select workout tops in a specific size, style, and color palette. Before checkout, it noticed other items already sitting in the cart and asked whether those should be removed. After approval, it completed the purchase and bought only the requested tank tops.

That small interaction revealed one of Muse’s most useful traits: it can handle mundane tasks with enough awareness to avoid obvious mistakes. It also demonstrated the product’s central promise — a system that can move from recommendation to action without requiring the user to click through every step.

What can Meta’s Muse actually do?

Muse is built around a mix of agentic automation and generative AI features. Beyond shopping and inbox cleanup, the assistant can generate AI podcasts, create images and videos, and produce interactive documents or webpages Meta calls artifacts.

Like several other next-generation assistants, Muse also maintains a running list of ideas and goals based on user behavior and prior requests. Those suggestions can include things like product comparisons, subscription cleanup, and reminder-style prompts for niche interests.

• Email management: Connects to Gmail and can read or delete messages.
• Shopping assistance: Can browse and buy products through linked retail accounts.
• Media generation: Produces images, video, and podcast-style outputs.
• Artifacts: Builds interactive pages or documents from prompts.
• Personalized ideas: Suggests tasks and goals based on user behavior.

Where Muse still feels rough around the edges

Despite the breadth of its features, Muse is not free of boundaries or mistakes. It refused several image-generation prompts that involved cartoon characters, including requests that resembled well-known copyrighted figures. Yet it was willing to generate promotional visuals that borrowed heavily from Apple’s branding.

In one case, it created an image of a fictional product launch scene featuring an iPhone-like foldable device, an Apple Watch, and AirPods, all arranged with Apple logos and a keynote-style backdrop. The image also showed fake native app icons on the phone, though some labels were inaccurate and one app name was misspelled.

That combination of caution and inconsistency is typical of many current AI systems. They may block obvious copyrighted characters while still producing content that closely resembles branded products or familiar corporate identities.

Why the feed feature is the most revealing part

The most surprising part of Muse may be its “Feed” tab, which generates a personalized news stream from a text prompt. Using Meta’s default instruction — asking for a clear, direct feed that is quick to skim and avoids clickbait — Muse produced AI-generated summaries of current events.

Those summaries included stories such as Apple’s reported iPhone Duo launch event, dangerous surf conditions along California’s coast, and tanker ships destroyed in the Iran war. It also surfaced news from the state where the tester was visiting family, information Muse said it learned from the shipping address linked to an Amazon order.

That detail is important because it shows how the assistant can knit together data from different services into a broader picture of a person’s life. Even seemingly minor account information, like a shipping address, can become a clue in a much larger personalization system.

How Muse explained its knowledge

When asked where it got the detailed list of interests it produced, Muse responded that the exact list is not shown in Instagram’s interface. Instead, it said it pulled the information through account API data, which can expose more than what the normal app interface displays.

According to the assistant’s explanation, the closest visible user-facing equivalent is Instagram’s ad-topics controls, which show broad categories advertisers may use and allow some of them to be reduced or removed.

That explanation is telling even if a user cannot independently verify every underlying data source from the app itself. It suggests a gap between what people think they are sharing with Meta and what connected services can actually surface behind the scenes.

What exactly did Muse know about the tester?

Muse rattled off a remarkably specific profile that included anime, CrossFit, Labrador retrievers, Florida wildlife, and nostalgia for the 1990s and early 2000s. Those topics echoed the kinds of content that often appear in Instagram Reels and related engagement systems.

Those details are not especially sensitive on their own, but together they create a portrait of a user that feels more intimate than a standard interest list. The unsettling part was not merely that Muse guessed correctly — it was that it appeared able to access and organize that picture in a way the user could not easily see in the app.

That is the tension at the center of Meta’s AI strategy. Personalization can make an assistant seem helpful, but once the system starts surfacing hidden data connections, it can feel less like a tool and more like a surveillance layer wrapped in convenience.

How does Muse compare with other AI agents?

Muse belongs to the same broad category as other agent-style products that aim to act on behalf of users rather than simply answer questions. Some competing tools can summarize research, draft documents, or move through browser tasks, but many still struggle when a workflow requires real-world judgment or multi-step account access.

Meta’s version stands out because of the sheer amount of first-party and cross-platform data it can potentially draw from. The combination of Facebook, Instagram, email, shopping history, and browser-like task completion gives it a personalization advantage that rivals may not match.

That advantage could also become a liability. The more context an assistant uses, the more likely users are to ask where that information came from and how much of it is stored, inferred, or shared.

1. It can perform real tasks rather than only offering suggestions.
2. It draws from multiple accounts and data sources.
3. It can expose hidden or inferred personal interests.
4. Its usefulness depends heavily on user trust.

Why privacy will determine whether Muse succeeds

Convenience alone is not enough for an AI assistant that wants access to inboxes and payment methods. Users also need confidence that the system will not overshare, misuse, or overinterpret their data.

Meta’s own history makes that especially difficult. Even when a company says information is not going to advertisers, many users will still hesitate before granting an assistant broad permissions across the services that contain their most personal communications and purchases.

That hesitation is not theoretical. In practice, the assistant’s ability to infer intimate interests from multiple accounts is exactly what makes it useful — and exactly what may prevent people from relying on it for sensitive tasks.

Why trust matters more than features

The core technical question is no longer whether an AI agent can click through a website or organize messages. The harder question is whether users are willing to let a platform with Meta’s reputation sit in the middle of their digital lives.

For many people, the answer may depend on whether Muse can prove that it respects boundaries better than the company’s previous products. If it cannot, even strong performance on everyday chores may not be enough to overcome the feeling that the assistant knows too much.

What comes next for Meta’s AI push?

Muse is likely only the beginning of a broader effort to integrate more autonomous AI into Meta’s ecosystem. If the assistant proves reliable, it could become a bridge between Meta’s social apps and a more utility-focused future where the company competes directly with productivity software, shopping tools, and smart assistants.

But the early test suggests that the path forward will be shaped less by technical novelty than by user comfort. An agent that can clear a mailbox and place an order is useful. An agent that can also reconstruct a detailed profile from hidden account data is powerful — and potentially alarming.

That is the paradox Meta now faces. Muse works, at least for many everyday jobs, but its real challenge is not to prove that it can help. It is to convince users that help will not come at the cost of control.

Timeline of the Muse test

The following sequence captures the main stages of the hands-on experience and how the assistant behaved at each step.

Bottom line

Muse shows that Meta can build an AI assistant that does real work, not just chat. It can clean inboxes, help with shopping, summarize information, and generate media, all while tailoring its behavior to the user.

But the same personalization that makes Muse feel helpful also makes it feel invasive. The assistant’s ability to extract and synthesize detailed personal interests from connected accounts may be impressive, yet it also explains why many users may hesitate before giving Meta that much access.

For now, Muse’s biggest obstacle is not whether it works. It is whether people will feel comfortable letting it know so much.