Loading page…
A successful search was a one-time task. More useful financial errands could require deeper account access, while Meta says Muse data may help train future models unless users opt out.
Listen to this story
Meta’s Muse produced a tangible win for TechCrunch reporter Sean O’Kane, who says a check is coming after the agent suggested he search unclaimed property. But he sees little repeat value so far, while the recurring jobs he can imagine—flagging duplicate charges or canceling subscriptions—require linking services such as Gmail or a credit card. Meta says users can control actions and opt out of using inference data for future training; its planned Confidential VM is not yet available, leaving trust a central hurdle
O’Kane says Muse initially treated him like a stranger, which made him more willing to try it; later, he felt it encouraged him to connect more accounts.
Meta says each user’s agent runs in a dedicated cloud computer and cannot see connected services’ real credentials; a separate system can allow, deny, or prompt for actions.
Muse conversations and tool-use records may train future models, but users can opt out through a setting.
Meta’s Muse prompted TechCrunch’s Sean O’Kane to look for unclaimed funds. It found some, and he says a check is on its way. In a new Equity podcast discussion, O’Kane called that result useful but questioned whether he would trust the agent with the accounts needed for more regular help.
O’Kane did not go looking for a way to test unclaimed-property searches. Muse suggested checking for money in his name, and he followed the prompt. He said he would not have done the search otherwise. The result gave him a concrete reason to appreciate the agent on his first day using it.
Finding that money is not an errand he can repeat every day. O’Kane described it as more of a “party-trick type thing” than a reason to keep using Muse. He also said he had not found another task nearly as useful. That is his experience with the product, not a measure of what other users have found.
Muse is Meta’s consumer-focused personal AI agent, introduced earlier in September and featured at the company’s Connect event. O’Kane saw a more repeatable use in the financial chores Meta has discussed: spotting double charges or canceling subscriptions. Those jobs would mean connecting information such as a credit card account or Gmail. They also set a different test from finding money through a suggested search: whether a person is comfortable letting the agent work inside accounts they use regularly.
O’Kane expected Muse to pull in context from his Meta accounts when he installed it. Instead, he said, it initially treated him like a stranger. That made him more willing to try it. As he continued using the app, though, he felt it encouraged him to bring more of those accounts into the system. His hesitation was not that Muse had failed at the first task; it was about what a more capable version of that help might ask him to share.
He drew a personal distinction between Meta and Apple. O’Kane said he would be more willing to give a Siri-like agent sensitive information because he trusts Apple more and sees Meta’s advertising business differently. That is his judgment about the companies, not evidence that Muse used his financial information for advertising.
Meta’s account of Muse’s design offers a more precise way to examine that concern. The company says each user’s agent runs in a dedicated cloud computer. It says the agent cannot see real credentials for connected services, while a separate system decides whether to allow an action, deny it or ask the user. Those are safeguards against the agent taking an unauthorized action or exposing credentials; they do not answer every question about Meta’s use of information.
Meta says Muse inference data, including conversations and records of tool use, can train future model versions. Users can opt out through a setting. The company has also described a planned Confidential VM, intended to cryptographically prevent Meta from accessing data in a user’s cloud computer. It is a plan, not a protection users can count on in the current product.
O’Kane’s test leaves two questions distinct. Muse found him money without becoming a fixture in his life. The chores he considers more lasting would ask him to connect accounts he is not ready to share. Meta’s permission controls and training opt-out give users choices about parts of that process, but the podcast discussion does not show whether those choices are enough to win his trust—or make the agent useful again tomorrow.
Editorial analysis
The unclaimed-funds search is an unusually clear demonstration of an agent doing something useful without first becoming part of every corner of a person’s life. That may also be its limit. The financial tasks Sean O’Kane sees as more durable would ask for access to accounts he is reluctant to connect. Meta’s planned Confidential VM addresses one form of access by the company, while its current training opt-out addresses another data decision. Neither, on its own, shows that people will make Muse a daily habit. The next useful evidence would be whether users keep returning for tasks that require connected accounts, not just whether they try a suggested search.
Citation desk / original work
The unclaimed-funds search is an unusually clear demonstration of an agent doing something useful without first becoming part of every corner of a person’s life.
/posts/techcrunch-host-finds-money-with-meta-s-muse-but-questions-trust-needed-for-daily-use#finding-1
Loading discussion...
Join the conversation
What makes that default fair or unfair to you?
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.