Meta's AI Image Tool Lasted 96 Hours

Key Takeaways

• Meta launched Muse Image inside Instagram, WhatsApp, and the Meta AI app, letting anyone generate images by @-mentioning a public Instagram account, and the person referenced received no notification.
• Despite defending the tool's "strong controls and safety guardrails," Meta pulled the @-mention feature within 96 hours after Creative Artists Agency and SAG-AFTRA condemned its opt-out design.
• The content-reuse setting that fed the tool remains active by default on public accounts, so the consent problem the backlash targeted did not disappear with the feature.
• Meta's invisible Content Seal watermark may not satisfy the EU AI Act's visible-labeling requirement, which takes effect August 2, 2026.

A Tag Was All It Took to Generate Someone Else's Face

Meta rolled Muse Image into Instagram, WhatsApp, and the Meta AI app this month, and the mechanism behind it fits in one sentence. Type an @-mention of any public adult Instagram account inside the Meta AI chatbot, and the tool pulled that person's public photos as reference material to generate new AI images involving their likeness, with no message sent and no permission asked, as Rappler reported this week.

Private accounts and users under 18 were excluded automatically. Everyone else was opted in the moment the feature shipped, and the toggle to opt out sat buried inside Instagram's settings menu, a design enabled for all adult public accounts by default, as Tech Times detailed. Meta described this as "strong controls and safety guardrails from day one." The guardrails did not include asking anyone.

Hollywood Didn't Wait for Meta to Notice the Problem

Creative Artists Agency said artists deserve to decide if and how their likeness is used, with consent and the ability to set their own terms, according to a statement Deadline obtained from the agency. SAG-AFTRA, representing more than 160,000 film and television workers, called the opt-out default "an utter miscalculation of public sentiment" and urged its members to disable the setting immediately.

Meta folded fast. By Friday, three days after launch, the company updated its own announcement to admit the feature "missed the mark" and pulled the @-mention capability entirely. I'd call that responsive if the response addressed the actual failure, but it didn't, because Meta never explained why a company built on facial recognition and ad targeting somehow failed to predict that letting strangers generate images of people who never agreed to it would go badly.

Meanwhile, the pattern is a familiar one for Meta specifically. The company quietly removed end-to-end encryption from Instagram chats by updating a four-year-old help article, another case of a privacy-relevant default changing with no direct notice to the people it affected.

The Toggle That Survived the Backlash

The @-mention capability is gone, but the setting that fed it survived the backlash intact. The content-reuse toggle governing whether someone's public photos can be used in any current or future Meta AI feature remains active by default, and turning it off now does nothing to remove images already generated before the switch was flipped.

There is also a labeling problem waiting on the other side of summer. Meta's Content Seal watermark is invisible, embedded only in machine-readable form, and the EU AI Act's Article 50 requires visibly labeled AI-generated content starting August 2, 2026. An invisible watermark does not meet a visible-labeling standard, whatever else it accomplishes.

Under GDPR, the privacy group NOYB has argued for years that opt-out consent does not meet the regulation's threshold, and that platforms must obtain explicit opt-in agreement before processing someone's data this way. Meta built Muse Image as though that argument did not exist, then acted surprised when the argument showed up anyway.

Pulling a feature after the damage is already generated is cleanup dressed up as accountability, and Meta should be named for that. The tool is gone, but the setting that made it possible is still running quietly on every public account that never agreed to any of this.