Meta has shut down the feature within its new Muse Image AI model that let Instagram users generate AI images by @-mentioning any public account, pulling it days after launch. The company updated its original announcement post to state the capability is "no longer available." The Verge

Muse Image debuted as part of Meta AI's image-generation stack, detailed in a blog post at about.fb.com on July 7. Among its capabilities was a reference tool: tag a public Instagram handle in a prompt, and the model would draw on that account's posted imagery — face, style, likeness — to generate new pictures. Critically, this worked on any public account, not just those of the person issuing the prompt. Account owners had no way of knowing when their content had been pulled into a generation unless they checked settings themselves.

Meta's initial response to concerns was an opt-out toggle buried in Instagram settings, added before the company decided to disable the feature outright. The Verge That opt-out shifted the default burden onto users and creators to discover the feature existed and act on it — an arrangement SAG-AFTRA flagged quickly, telling members to opt out and publishing step-by-step instructions for doing so. The union's involvement signals the feature's exposure for professional performers, whose faces are both public-facing by career necessity and now, evidently, machine-referenceable without consent.

Meta did not notify users whose accounts were used to generate images through the tool. The New York Times That absence of notification is the crux of the controversy: the mechanism operated silently, with the only signal of exposure being the eventual existence of AI-generated images bearing someone's likeness circulating on or off the platform. NBC Bay Area's reporting confirms the feature's name as Muse Image and situates the opt-out mechanism within the wider Instagram privacy-settings discourse that followed. NBC Bay Area

The timeline is tight. Muse Image launched July 7. Backlash over the @-mention referencing capability built within days, centered on the deepfake-adjacent risk of generating images of individuals, including minors and public figures, without permission or awareness. By July 10, Meta had disabled the feature and revised its own announcement to reflect that.

The mechanics matter here for anyone evaluating consent architecture in generative tools. An opt-out model assumes affirmative, informed non-consent is sufficient cover — it is not, particularly when the default state is "your public content is fair game for reference generation" and the discovery burden falls entirely on the user. Contrast this with an opt-in framework, where no account's imagery enters a reference pipeline until its owner explicitly permits it. Meta chose the former at launch. The gap between those two models is where most of this week's criticism concentrated, and it's the gap Meta ultimately closed by pulling the feature rather than reworking the consent flow.

Worth flagging: Meta's decision to disable rather than patch is itself informative. Reworking an opt-in flow, adding notification hooks, or restricting the reference tool to first-party content would have preserved the feature in some form. Removing it outright suggests either that no quick technical fix satisfied the legal or reputational exposure, or that the company judged the feature not worth defending in its current form while it reworks the underlying trust model. Meta has not detailed which.

This is not Meta's first collision with likeness rights on generative tools — the company has iterated repeatedly on AI features tied to real people's images across its platforms, each time recalibrating after public reaction rather than anticipating it in design. In this author's view, the recurring pattern — ship broad, opt-out by default, retreat under pressure — reflects a product culture still treating consent as a settings-page afterthought rather than an architectural constraint baked in before launch. Public accounts are, by definition, viewable by anyone; that has never meant their content is licensed for arbitrary AI recombination, and platforms building reference-based generation tools will need clearer legal and product answers to that distinction before the next launch, not after.

For enterprise and consumer AI teams watching from outside Meta, the episode is a useful data point on where regulators, unions, and users are drawing lines around generative reference features in 2026: public availability of an image is not being treated as equivalent to consent for AI reuse, and silent operation — no notification, no audit trail visible to the subject — is treated as an aggravating factor rather than a neutral implementation detail. Meta has not indicated whether or when a revised version of Muse Image's reference capability might return.