Meta's Muse Image let anyone generate AI pictures of public Instagram users by tagging them, with no notification required. Three days and a wave of backlash later, it's gone.
Meta built an AI tool that could turn your public Instagram photos into raw material for anyone's image generation prompt, and it didn't tell you first. That's the part that killed it.
Muse Image launched on July 7, 2026, the first image model to come out of Meta Superintelligence Labs, the unit Mark Zuckerberg built around Scale AI founder Alexandr Wang. Meta pitched it as its most capable image generator yet, one that follows instructions precisely, edits with accuracy, and composes from multiple reference images. It shipped inside the Meta AI app, on meta.ai, in Instagram Stories in the US, and on WhatsApp in select countries.
The feature that caused the trouble was simple to describe and, apparently, simple to underestimate. Users could @-mention any public Instagram account, including those of public figures, to reference that person's likeness in a generated image. Every public account belonging to a user 18 or older was included by default. Opting out meant digging into account settings and disabling it yourself. If someone used your face to generate an image, you weren't notified.
That default is what turned a product launch into a three-day crisis. CAA, the talent agency representing Tom Hanks and Meryl Streep among others, went directly to Meta with its objections. The agency's position, as reported by Deadline and Variety, was blunt: no one's name, image, likeness, voice or creative work should be used by any third party, AI models included, without clear, documented consent. CAA wanted protection to be the default setting, not something a user has to go find and switch on.
SAG-AFTRA moved just as fast, publicly urging its members and Instagram users generally to opt out and, as the union put it, take action to protect their likeness. That's a notable escalation from a union that spent much of its recent contract fights negotiating AI likeness protections into film and television deals. Watching Meta ship a consumer feature with the opposite default, opt-out instead of opt-in, was never going to sit well with them.
Meta pulled Muse Image from Instagram on Friday, July 10. A company spokesperson's statement was short: the feature "missed the mark." No detailed roadmap for a fixed version. No timeline for return. Just an acknowledgment that the design was wrong and a retreat.
A rough week to be Meta's trust and safety team
The Muse Image reversal didn't happen in isolation. On the very same day, the European Commission ordered Meta to dismantle what it calls addictive design elements across Facebook and Instagram, including autoplay video and infinite scroll, warning of fines up to 6% of Meta's global annual turnover under the Digital Services Act if the company doesn't comply. Two different regulatory and reputational fires. One week. One company.
It's also a strange contrast against what Meta did just one day earlier. On July 9, Meta opened Muse Spark 1.1, its agentic reasoning model, to outside developers through a new Meta Model API, pricing it at $1.25 per million input tokens and $4.25 per million output tokens. Zuckerberg said that's roughly a quarter of what OpenAI and Anthropic charge for comparable models. That launch went fine. It was a straightforward commercial move: developers sign up, pay for tokens, build things. Nobody's face was involved.
That's the actual lesson here, and it's not really about image generation technology. Muse Spark 1.1 succeeded because the people affected by it, developers integrating an API, chose to be affected by it. Muse Image failed because the people affected by it, millions of public Instagram accounts, never got a choice at all. You can ship an ambitious AI product and get away with plenty. What you can't get away with is using someone's face without asking first and calling the silence consent.
For Meta Superintelligence Labs, barely months old and still establishing what kind of lab it wants to be, that's an expensive lesson to learn in public. Wang's unit is racing to build models that compete with OpenAI and Google on capability. This week showed that capability was never the constraint. Judgment about consent was.
Also read: Zuckerberg Breaks His Three Year Silence On X To Launch Muse Spark 1.1 • Elliott builds a stake in CCC Intelligent Solutions as a sale looms • SK Hynix's CEO Says the Memory Chip Crunch Will Outlast This Decade
