Mark Zuckerberg is betting that Meta’s next major artificial intelligence product will become a personal agent capable of handling much of everyday life online. Muse, unveiled at Meta’s developer conference in Menlo Park, can book hotels, flights and restaurant tables, write emails, manage calendars, negotiate contracts and carry out online purchases. But the technology also demands something Meta has struggled to secure in recent years: the willingness of users to give it extensive access to their personal information and digital accounts.
Muse is designed as the first Meta AI agent aimed at a mass market, including people with little technical knowledge. Unlike conventional chatbots, it is intended to act on a user’s behalf rather than simply answer questions. It is currently available only to adults in the United States. Zuckerberg said on Wednesday that he expected “Muse to develop into a personal superintelligence”, adding that “billions of people around the world will use it to achieve their goals and improve their lives”. He has also described using it for “personalised baking recipes” to try with his daughter at weekends.
Meta’s ambitions extend beyond the Muse application itself. The company is working with businesses including PayPal and travel company Expedia, while planning to incorporate the agent into its AI glasses. By the end of the year, it also plans to launch Muse Charm, a wearable device with a screen and microphone, no larger than a key ring, designed to keep users connected to their agent wherever they are.
For Meta, the technology arrives at a critical point in its business. The company’s user numbers have stagnated, although they remain enormous at 3.6 billion. It is facing numerous legal cases and financial penalties, while governments in several countries are considering restrictions on social media use by children; Australia has already introduced such a ban. At the same time, the artificial intelligence race is consuming billions of dollars, including large payments to recruit developers from competitors.
Meta is spending $137 billion on AI development this year alone. The company can still finance much of that spending through advertising revenue, but investors will eventually expect the investment to generate returns. The initial response to Muse has been encouraging: the app was downloaded more than 900,000 times during its first seven days and Meta’s share price rose 11 per cent. Sustaining that momentum, however, depends partly on whether users are prepared to give the agent the access it needs.
The problem is straightforward. A genuinely useful digital agent needs extensive permissions. Users can give Muse access to their email inboxes and their Instagram and Facebook accounts. Without those credentials, an agent would repeatedly stop to request logins or two-factor authentication, potentially creating more work than it saves.
Early demonstrations suggest how much the system can already do. One journalist asked Muse to order healthy, relatively non-perishable groceries, resulting in sweet potatoes and bananas being delivered. Another user said the agent negotiated discounts on insurance. In one test described by the source, Muse booked a hotel, unsubscribed from unwanted newsletters and solved website CAPTCHAs designed to keep automated systems out. Asked to find a second-hand bag within a specified budget, it created an eBay account, negotiated with sellers and correctly supplied a delivery address found in an Airbnb booking confirmation.
That degree of access creates a particular problem for Meta. In 2019, the company paid $5 billion to the US competition authority over privacy violations connected with the Cambridge Analytica scandal, in which data from more than 80 million Facebook users was used for political advertising without their consent. The following year, data from 533 million Facebook accounts became public following a hacking incident.
Meta says Muse will use the data it collects to train its artificial intelligence unless users actively object. Each user is supposed to operate within a separate, protected cloud environment inaccessible to other users, and Meta says the data will not be used for advertising. Zuckerberg has repeatedly stressed the importance of safety testing, saying the company held Muse back for several weeks to test it further, although that process cannot independently be verified.
The company also plans to charge frequent users for additional computing power, with subscriptions costing $20 or $100 a month depending on requirements. Zuckerberg has indicated that Meta may also take commissions on purchases made by Muse on users’ behalf. That business model could face resistance from online retailers concerned about losing their direct relationship with customers. Amazon currently blocks agents such as Muse from its website, citing its terms and conditions.
The problem emerged during the conference itself. While Zuckerberg was still delivering his keynote, an eBay account used by Muse to negotiate the purchase of a second-hand bag was suspended. The company cited “serious security concerns”. The user was told that outstanding payments remained due but could no longer see which agreements Meta’s agent had made on their behalf. When asked what had happened, Muse could not explain it.
