If a photo of your own face that you posted on social media turned out to have been used as material for someone else's AI-generated image without your ever requesting it, you would likely feel a sense of unease. Meta AI's new model "Muse Image" was set up in such a way that simply @-mentioning a public Instagram account could generate an image of that person without their consent. Just three days after its July 7 launch, Meta removed this reference feature uniformly across all accounts. Behind this move, protests from talent agencies and labor unions converged simultaneously with criticism from privacy experts and ordinary users. But this marks the third time an issue with consent-related design has surfaced at Meta.
The Path from Announcement to Feature Removal in Just 3 Days
On July 7, 2026, Meta Superintelligence Labs announced "Muse Image," an image-generation AI model developed by the lab, and began offering it through the Meta AI app and meta.ai. On the Image Arena leaderboard, which measures image generation capability, it ranked second behind OpenAI's GPT Image 2, and the video counterpart, Muse Video, placed third on video generation benchmarks—according to a report by the-decoder. While the technical evaluation was high, an issue separate from performance emerged immediately after launch.
The Meta AI app had built into it a feature that could generate an image of a person simply by @-mentioning their public Instagram account—without that person's consent. This design drew criticism from CAA (Creative Artists Agency), SAG-AFTRA (the Screen Actors Guild - American Federation of Television and Radio Artists), privacy experts, and ordinary users, and Meta announced on July 10—just three days after the initial announcement—that it had discontinued the feature. According to the-decoder, Meta officially acknowledged, "In response to feedback that this feature did not meet expectations, we are discontinuing it." Only the @-mention person-reference feature was removed; Muse Image itself and its other image generation capabilities remain in service.
Even so, the speed with which the feature itself was completely halted—just three days after launch—is unusual for a consumer-facing AI product. Behind this speed lies the fact that criticism from multiple directions converged simultaneously.
The Mechanism and Design Flaw Behind Generating AI Images with a Single @-Mention
The @-mention feature in Muse Image worked by simply specifying a public Instagram account with an @ symbol within the Meta AI app, which would then generate a person-image using that account's posted photos as material. The setting was on by default—that is, an opt-out system—meaning that, unless a person explicitly opted out, anyone could generate images of others by default, according to a report by TechCrunch. The only accounts automatically excluded were private accounts and accounts belonging to users under 18; all public accounts belonging to adults were included as targets. This was a design in which only users who knew the opt-out procedure could protect their own images, leaving the majority of public account holders—who had no idea the feature even existed—without any protection.
Comparing this design to the "cameo" feature of Sora, the video-generation app formerly offered by OpenAI, makes the location of the flaw clear. Sora's cameo feature was built on an opt-in model, in which individuals chose in advance whether to allow their likeness to be used in others' videos (Sora ended its service in April 2026). Despite earning a technical evaluation second only to GPT Image 2, Muse Image's consent design pointed in the opposite direction.
Under this design, every public account belonging to a user aged 18 or older that had not been locked was included as a target, regardless of follower count or celebrity status. Because the core functionality of Muse Image is available in Japan as well through the Meta AI app, Japanese Instagram users with public accounts are believed to have been included within the same scope. TechCrunch published an article explaining how to opt out on July 9, the day before the feature was removed. In reality, it seems many users went about their business, until the feature's removal, unaware not only of how to opt out but of the very fact that their images could be used at all. The default-on design created a disparity between the small number of informed users and those who remained exposed without knowing it.
Compound Pressure Halted the Feature in 3 Days
It was not a single protest that moved Meta. CAA issued a statement condemning the feature, saying, "An individual's name, image, likeness, voice, and creative works should not be used by any third party, including AI models, without clear and documented consent," and called for a change from opt-out to opt-in. According to Variety, after Meta removed the feature, CAA shifted to praising "Meta's swift decision."
According to Deadline, SAG-AFTRA published opt-out setting instructions on X (formerly Twitter) not only for its members but for all Instagram users, and it also delivered a statement to the Hollywood Reporter saying, "Despite widespread recognition of the dangers associated with non-consensual digital replication, it was a misjudgment to create a feature that facilitates it." Thorin Klosowski, a security and privacy activist at EFF (Electronic Frontier Foundation), also told the Washington Examiner, "This setting absolutely should have been opt-in." In addition, concerns from ordinary users and the media spread in parallel, including TechCrunch's July 9 publication of its own guide to the opt-out procedure.
This speed was not produced by the voices of talent agencies and labor unions alone. Protests from prominent organizations converged at the same time as criticism from ordinary users and privacy experts, and it appears this convergence lay behind Meta's unusually fast response of three days. The feature itself was removed uniformly across all public accounts, not as a response targeted at any particular group.
Meta's Repeated Failures in Consent Design
This is not the first time Meta has faced an issue involving non-consensual generation of a person's likeness. In August 2025, a Reuters investigation revealed that unauthorized chatbots impersonating celebrities such as Taylor Swift and Scarlett Johansson had been operating on the Meta AI platform, engaging in sexual advances and generating photorealistic images of them in bathtubs or in underwear. Many of these bots were created by users, but at least three were confirmed to have been created by Meta employees. According to Variety, Meta deleted roughly a dozen of the offending bots immediately before the investigation's findings were published, but this response merely removed individual bots and did not address the underlying design that allowed celebrities' likenesses to be used without consent in the first place. The recurrence of the same type of problem the following year with Muse Image can be seen as a result of this underlying design having been left unaddressed.
Meta's official announcement makes no mention of why it did not adopt an opt-in system from the start, or what internal decision-making process led to the choice of default settings. This silence itself supports the notion that consent design has been handled on an ad hoc basis each time.
The laxity of the underlying design had also surfaced in a different form in May 2025. When Meta used EU users' public posts as training data for Llama, it likewise adopted an opt-out system, prompting the privacy organization NOYB to warn of legal action, stating that "the GDPR (EU General Data Protection Regulation) requires opt-in consent." Whether it's unauthorized chatbots, the use of training data, or now Muse Image, the target changes each time, but the same pattern repeats: "turn it on by default first, then respond individually once criticism arrives." Lining up these three cases reveals that Meta lacks a mechanism for building consent design into products from the earliest stage.
The Dangerous Feature Is Gone, But the Habit of Judgment Remains
The complete removal of the @-mention feature was an adequate response to the specific feature that drew criticism. But the Muse Image technology itself has not withdrawn from corporate activity. Meta's official announcement states that within the coming weeks, advertisers and agencies will be able to use Muse Image through Advantage+ creative. The use cases listed for Advantage+ creative include generating product backgrounds, creating lifestyle image variations from existing ads, and generating still images from video—with no mention of a feature that would allow unauthorized reference to real people's Instagram accounts. The "non-consensual person reference" design that drew criticism and Advantage+ creative are, at least as far as has been publicly disclosed, separate mechanisms.
Even so, the point of contention that remains lies in something Meta's official announcement has consistently failed to explain. Meta has said nothing about why the @-mention feature was designed with an opt-out system rather than an opt-in system from the start, or about the decision-making process behind that choice. This time, a swift response averted disaster, but there is no guarantee that the same design decision won't be made for the next generative AI feature Meta rolls out.
Looking at the issue of non-consensual image generation across the industry as a whole, Meta's response speed stands out. xAI's chatbot Grok has, since 2025, continued generating non-consensual images of real people in swimwear or with sexual insinuation, with reports indicating 4.4 million images generated over nine days, of which 1.8 million were sexual images of women. According to an estimate by the Center for Countering Digital Hate, 23,000 sexual images of children were generated over eleven days, and the matter has developed into multiple class-action lawsuits. While Muse Image became a problem and was removed within three days, Grok continues to operate while carrying the same type of problem—and judging by response speed alone, Meta's action was fast. What both companies share in common is a structure in which opt-in design is not treated as a prerequisite before a product's release, but is instead addressed after the fact once the problem becomes visible.
On the Eve of Tightened Regulation, and What's Different for Japanese Readers
Under current law, the transparency obligations of Article 50 of the EU AI Act are scheduled to take effect on August 2, 2026, requiring providers to ensure machine-readable marking and detectability for AI-generated or manipulated content, and requiring deployers of deepfakes to disclose to viewers the fact that content has been generated or manipulated. An analysis by GT Law notes that penalties for violations can reach up to €15 million or 3% of global annual turnover, though a political agreement has already been reached to postpone some obligations to December 2, under the AI Omnibus proposal, which is still awaiting formal adoption—meaning the actual timing of application remains fluid. Even so, it remains the case that a feature like Muse Image's could fall within the scope of this regulation. In the United States, the "TAKE IT DOWN Act," which makes the publication of non-consensual sexual images or deepfakes, as well as threats to publish them, a federal crime, was enacted on May 19, 2025. In Montana, HB513—which recognizes property rights in one's name, voice, and likeness and permits recovery of actual damages and profits gained for certain acts such as unauthorized digital replication for commercial purposes—took effect on January 1, 2026. Meanwhile, the "NO FAKES Act," which would comprehensively establish similar rights at the federal level, remains at the stage of a bill reintroduced to the Senate in 2025 (S.1367) and had not been enacted as of July 2026.
For businesses offering a feature like Muse Image's within the EU, failing to provide transparency labeling on generated content after August 2 will no longer be a matter that can be resolved through a one-off round of criticism. What Article 50 requires is marking and disclosure of the fact of generation or manipulation; it does not concern itself with whether consent was obtained. Whether the speed of response Meta showed in removing the feature will be replicated in a scenario carrying legal force is the next thing to watch.
For Japanese readers, this incident offers material for thinking about differences in legal protection frameworks. While the core functionality of Muse Image itself is available in Japan through the Meta AI app, the AI effects within Instagram Stories and image generation within WhatsApp DMs are reportedly not offered in Japan. Japan has no dedicated law, like the U.S. TAKE IT DOWN Act or Montana's HB513, that clearly positions a person's likeness as property and establishes a compensation framework for unauthorized use. That said, in addition to publicity rights protecting the property value of one's name and likeness—as established in the Pink Lady case—there also exist, as protections under civil law, the right of likeness (derived from personal rights against being photographed or used without consent) and privacy rights, and the handling of facial images may also fall under the oversight of the Act on the Protection of Personal Information and the Personal Information Protection Commission. If the same problem occurred in Japan, there would be no statutory damages provision as in the U.S., but the available remedies would not be limited to civil litigation alone.
What Meta removed this time was the @-mention feature itself, but the underlying decision-making process behind why it chose an opt-out system remains unexplained. With the application of the EU AI Act's Article 50 transparency obligations set to begin on August 2, once regulators start enforcing labeling requirements, there should be more occasions on which Meta is pressed to explain which feature it will next apply the same design decision to. Whether this pattern repeats itself in the next generative AI feature rolled out, including Advantage+ creative, will be the real point of resolution for this incident.