Meta Is Training Its AI on Your Chats With a Plush Mascot

A Plush Mascot in a Toga, Banned for Anyone Under 18

The character at the center of the Muse app is called Jolly, a round, fuzzy creature you dress up like a doll. Mark Zuckerberg introduced it at Meta Connect in a toga and laurel wreath. In Instagram ads spotted by WIRED, it wears workout gear and pitches itself as a fitness coach. By year's end, Meta plans to sell a small Tamagotchi-style gadget to carry it around.

Alexandr Wang, the company's chief AI officer, posts memes of Jolly nonstop, some of them suggestive. And the app itself is restricted to adults, 18 and up. Which is the question WIRED put in its September 26 headline: why does a product built for grown-ups look like a well-designed kid's toy?

Fair question. The last line of that same piece raises a more checkable one: Meta trains its models on your Muse conversations unless you opt out. We went and read the actual instructions. One thing before we dig in: Muse isn't available in Europe yet, and by the end we'll see that this detail says something about the setting itself.

What Meta Says It Trains On, and What It Doesn't Say

Meta's help pages are clear on the principle and vague on the scope. They state, "You decide whether we can use your interactions with Muse to train and improve our AI models," and add that "This setting is on when you first use Muse." The word doing all the work here is "interactions." Meta never defines it anywhere.

And Muse does a lot more than chat. It's an agent that keeps a history, stores files and uploaded documents in a cloud virtual machine, and browses the web for you. The help pages never say whether "interactions" covers those files, that machine, or that browsing. The only thing Meta actually lists is what it strips out before training: "certain categories of personally identifiable information like names, email addresses, phone numbers and Social Security Numbers."

That's a list of what gets removed, not what gets kept. It's like a mover promising not to touch your houseplants, saying nothing about the rest of the living room. Meta does add two guarantees: neither the conversations nor the virtual machine's data feed its advertising systems, and interactions are disassociated from your account.

Opting Out Takes Three Taps, and It Reaches Into the Past

The path exists, and it's short. Go to Settings, then tap Data controls. Toggle off Help improve our AI models. Select Turn off to confirm.

Here's the surprise: it favors Meta. The setting doesn't just look forward. The help page states plainly that "Changes to this setting also apply to previous interactions." Most settings like this only govern what comes next. This one claims to cover what already happened, the more generous version of the gesture.

What no page addresses is what's already baked into a shipped model. The setting governs how interactions get used; it doesn't promise to pull anything back out. The closest Meta comes to admitting this is a line acknowledging that Muse "may still remember information it learned" from something you deleted. Opting out isn't an eraser.

Meta also runs a page for people who've never touched Muse. It notes that "When others use Muse, information about you may be provided to Muse, even if you don't use Muse directly." The only recourse is a privacy request, and Meta can only act on it for information tied to an account. No Meta account, no way to even ask.

In July, the Same Company Pulled an Opt-Out in Two Days

The most telling precedent sits in the product's own history. A feature called Muse Image let people tag a public Instagram account and generate images from its content. It applied automatically to every public account run by an adult, excluding private accounts and minors.

CAA, the talent agency representing Tom Hanks and Meryl Streep, went public on Wednesday, July 8. Its position: nobody's name, image, voice, or creative work should be used by a third party, AI models included, without clear, documented consent, and the protection should be the default. Two days later, on Friday, July 10, Meta pulled the feature, saying it had heard the feedback, that the feature had missed the mark, and that it was no longer available.

Forty-eight hours. Same company, same product, the same opt-out mechanics. Which makes the opt-out look less like a fixed principle than a judgment call, one that gets revisited fast when the pushback comes from people with lawyers on retainer.

On Age, Meta Has an Answer, and Nobody's Measured It

Asked about it by WIRED, Meta spokesperson Daniel Roberts opened by dismissing the premise: "This is a story in search of a story." He went on to say people enjoy building their own avatar, with its own personality and style, because it's fun. Then the mechanism: Meta says it asks everyone using Muse for their date of birth, blocks anyone it flags as possibly under 18, and runs additional checks.

Nobody outside the company has evaluated any of it. An audit from Gen Digital flags the terms-of-service clause that bars anyone under 18 from Muse, and finds that Meta's documentation describes no verification process. No outside researcher has tested the filter, and Meta has published no results.

The only numbers ever attached to age checks at Meta come from somewhere else entirely. On August 26, the company settled with 51 US state attorneys general for $17 billion over ten years, committing to detect users under 18. That settlement covers Instagram and Facebook, not Muse. And according to two specialized readings of the ruling, the thresholds it sets are tolerated error rates: at most 10% of 16 and 17 year olds misidentified as adults by a third-party tool, and 3% for 13 to 15 year olds.

The Setting Exists, the App Doesn't

Which brings us to what makes this whole question theoretical on this side of the Atlantic. Muse isn't available in Europe. Meta's September 8 announcement says the rollout covers "the US on iOS, Android, and muse.ai," coming soon to AI glasses, and never mentions Europe once. The help pages simply note that the app and website "are not yet available everywhere." No timeline, no reason given.

Meanwhile, the same maneuver is playing out in Brussels. The European Commission wants to write a new Article 88c into the GDPR that would treat AI training as a legitimate interest, and had originally placed an unconditional right to object inside it. The Council presidency's compromise text from September 3 renumbers the clause as Article 88a and moves that right into the preamble instead, the part of the text that states intent without binding anyone. On February 11, the European Data Protection Board and the European Data Protection Supervisor had already found that no specific provision was needed, and warned that some of the proposed changes risked weakening the level of protection people currently enjoy.