Meta Platforms, Inc. (META)'s new AI assistant, Muse, has gone to No. 1 on the U.S. App Store with more than 5 million downloads since it launched last month. Along the way, it's been keeping a running file on people who never downloaded it.
A TIME review of Muse's internal instructions found that the agent updates its dossiers every hour. They cover who matters to its users, what those users want, and which nudges change their behavior.
Meta pitches Muse as a productivity tool that handles chores like grocery shopping and canceling subscriptions. What it actually collects sharply contrasts with the growth story Wall Street is buying.
Mapping the Social Graph
Muse reads users' chats, messages, and emails. From them, it records how users met their contacts, what interests they share, their disputes and "tensions and alliances" within their social group. People who don't use Muse get mapped too, through other users' agents.
The assistant is built to infer goals users "have not said out loud." Each night it reviews the day's conversations to learn which prompts work best. "This user responds better to short nudges after 10 PM," reads one internal example.
A Meta spokesperson didn't dispute the findings.
"Muse remembers what matters most to you, including information about others that you choose to share, so it can be a helpful personal assistant," the spokesperson said.
The Power of Data
Meta's situation reminds of Target Corporation (TGT) and its famous example from more than 20 years ago. Back then, the firm's in-house statistician created a purchase behavior model that figured out whether female customers were pregnant – and if so, even approximated the delivery date.
Customer loyalty is the holy grail for retailers, and a window for establishing repeat purchasing behavior swings wide open when someone goes through transformational life milestones such as pregnancy.
Target's in-house model eventually got so good that it reportedly figured out a teenager was pregnant before her father knew.
Yet, in the second half of the 2020s, Meta's data collection goes far beyond shopping, giving the firm immense power over modeling human behavior – but also posing security risks.
Each Muse runs on its own virtual machine in Meta's cloud, "isolated so that no one else's agent can access it," the company says. Lessons still travel between them, though. "Muse agents across many VMs teach each other through shared lessons," the instructions read. Agents are told to strip out identifying details first.
Meta says it plans to offer encryption later this year. Until then, Meta can still access user data.
Wall Street Split
The firm's shares rose about 27% in September after the launch. BNP Paribas analyst Nick Jones, who has an $885 price forecast, pointed to Meta's ability to promote Muse across Instagram, WhatsApp and Facebook.
He also noted that app-open rates fell after the initial spike. Deutsche Bank added Meta to its Fresh Money list, and Tim Seymour of Seymour Asset Management called the stock "safer money."
Needham's Laura Martin, who rates Meta a hold, said there's no clear way yet to make money from Muse, even as agents running in the background push computing costs up.
Meanwhile, Karen Finerman of Metropolitan Capital Advisors and Courtney Garcia of Payne Capital Management are both keeping their Meta shares unchanged, noting that much of the optimism is already baked into the price.
META Price Action: Meta Platforms shares were down 0.75% at $715.89 during premarket trading on Thursday, according to market data.
