Meta withdrew within 48 hours the feature announced Tuesday that allowed AI image generation by tagging any public Instagram account. The Muse Image feature required no account owner consent and provided no notification of use.

The reaction was immediate and harsh. Haley McNamara of the National Center on Sexual Exploitation called the feature “an obvious tool for sextortion and fraud,” criticizing the opt-out approach. The Screen Actors Guild published instructions for disabling it, and talent agencies like CAA pressured Meta directly.

Why it matters to you. This is the first documented case of a major platform launching an AI image generation feature on real people’s images without prior explicit consent, then withdrawing it under pressure in two days. The precedent matters: if Meta reversed course this quickly, it signals that explicit consent is consolidating as a non-negotiable standard—not just for internal policy but for reputational and legal risk. If you’re designing any system using other people’s content for training or generation, this is the signal that opt-out no longer holds up.

To understand the context. Meta had already launched Muse Spark 1.1 API days earlier with aggressive pricing to attract developers: this feature was an attempt to bring the same generative capability into Instagram for regular users, but the chosen consent model proved unsustainable.

In detail

What came before

Meta had already integrated AI generative capabilities into Instagram and Facebook (Meta AI, available in chat and feed), but so far these tools worked on text prompts or images uploaded by the user themselves. The novelty of this feature was direct reference to public content from other users via tagging: write a prompt, tag a public account, and the model generates an image incorporating visual elements from that account.

The mechanism was opt-out: every user with a public profile was included by default and had to go into settings to disable use of their content. There was no notification when someone used your account in a generation.

The backlash

Criticism centered on three points:

1. Risk of sexual abuse and fraud. McNamara (NCOSE) highlighted that the feature is “an obvious tool for sextortion”: generating images of a person in false contexts (sexual, compromising, illegal) becomes trivial. The risk isn’t theoretical—nonconsensual deepfake cases are already documented and widespread, and a feature built into a platform with billions of users drastically lowers the technical barrier.

2. Burden reversal. The opt-out approach puts responsibility for protecting one’s image on the user, who must learn about the feature, find it in settings, and disable it. By contrast, other systems (like Anthropic and OpenAI) don’t allow generating images of identifiable real people without explicit prior consent.

3. Lack of transparency. There was no way for a user to know if or when someone had used their account to generate an image.

The Screen Actors Guild published a guide for members on disabling the feature, and representation agencies like CAA raised the issue directly with Meta. Press coverage (The Verge, TechCrunch, Puck News) was unanimous in criticizing the choice.

The reversal

On Friday, Meta updated the launch post with this line: “We heard feedback that this feature missed the mark, so it’s no longer available.” The feature was disabled globally.

The fact that Meta withdrew the feature in 48 hours, rather than simply switching the default from opt-out to opt-in, signals the problem wasn’t just implementation but the consent model itself. Dylan Byers (Puck News) reported the decision came after direct pressure from talent agencies.

What remains unclear

Meta hasn’t clarified:

• Whether images generated during the 48 hours the feature was active were retained or deleted.
• Whether public account content used during that period remains in the Muse Image training set or was removed.
• Whether an opt-in version of the feature will return or if the project has been abandoned.

Practical implications

If you’re building or deciding on a system using real people’s content (images, video, voice, text) for training or generation:

• Opt-out is no longer sufficient for content depicting identifiable people. The reputational and legal risk is too high, even if technically the policy respects platform terms of service.
• Notification matters. Even with opt-in, a system that doesn’t notify users when their content is used remains vulnerable.
• Consent must precede use, not follow it: apologizing after launch costs more than designing consent upfront.

This isn’t an isolated case: it’s the first public, rapid withdrawal of an AI feature on a consumer platform at this scale, and it traces a boundary that others will need to follow or justify crossing.