Meta’s Muse AI works and creeps me out

Meta launched Muse, its first serious AI productivity agent: shopping, email triage, trip planning, generated podcasts, images, videos, and interactive “artifacts,” running actions on a cloud virtual computer. The Verge’s hands-on found the busywork features mostly worked.

What overshadowed the demo was personal data. Muse built a feed that pulled local news from an Amazon shipping address, then listed specific interests — anime, CrossFit, Labradors, Florida wildlife, ’90s/2000s nostalgia — tied to the reviewer’s Instagram and Facebook signup. Asked where that list came from, Muse said it read API data that “surfaces more than the UI does.”

Meta says Muse only exchanges data third-party apps need to work and that it doesn’t share information with advertisers. The Verge asked Meta what Muse can access inside Instagram and Facebook; no immediate reply.

We want agents that delete promo mail and finish a cart without drama. Muse did those jobs. It even noticed stray cart items before buying workout tops — a small courtesy that signals real tool use, not just chat. On capability, this is Meta finally treating productivity as a product, not a side quest next to Reels.

The creep is the product. An assistant that knows more about you than your settings screens show is a trust bomb, especially from a company with Meta’s privacy history. The reviewer connected Gmail and Amazon with understandable unease; then Muse volunteered interest graph detail that didn’t match the coarser Ad topics list. That’s not a cute personalization flex — it’s a reminder that “helpful context” and “surveillance residue” can be the same database.

If this scales as the default Meta agent, the fight won’t be whether it can cancel subscriptions. It’ll be whether people accept an always-on worker that sees inbox, purchases, and social graph in one loop. Useful AI here needs sharper consent, clearer data maps, and refusal defaults — not another “we don’t sell it to advertisers” line while the agent narrates your psyche.

I’m watching whether Meta publishes a precise Muse data inventory, and whether users get a kill switch that actually severs Instagram/Facebook interest access without killing the product.

Muse refused some cartoon-character image prompts but generated Apple-logo product-launch imagery with hallucinated app labels. It also keeps personalized “ideas” and “goals” lists from conversations — standard agentic sticky memory, now attached to Meta’s identity stack.