Meta Pulls Its Instagram Tool for Deepfaking Strangers

Meta has switched off one of the more alarming features it has shipped in a while. For a few days this month, Instagram users could generate new images of other people simply by typing an at-mention and a username into its Muse Image tool. The system would pull physical traits from the target's public account and synthesise a fresh picture of them. After roughly three days of mounting criticism, Meta disabled the at-mention capability, as TechCrunch first reported.

The problem was not only that the feature existed. It was that Meta turned it on for everyone. The company made the setting opt-in by default for adult public accounts, which in plain terms means it was switched on without asking, and there was no notification to tell people their likeness was being used to make images for strangers. You could be the subject of an AI-generated picture and never know it happened.

The pushback

Performers' union SAG-AFTRA was blunt about the design. In its view, quoted by Variety, anything short of a clear and conspicuous opt-in for this kind of use of people's images is unacceptable. Privacy researchers and ordinary users piled on through the week, and by the time Forbes wrote it up, Meta had cancelled the tool less than a week after launching it.

Meta's fix went a step further than pausing the feature. The company removed the account setting labelled allow people to reuse your content on Instagram and with AI features on Meta, the toggle that had governed the whole thing. Muse Image still works as a general text-to-image generator, but it can no longer reach into another user's account and build a picture of them.

Consent by default

What makes this worth noting is not the retreat but the starting position. A company with Meta's legal and policy resources decided that the safe default was to enrol every public account in a face-synthesis feature and let people opt out if they noticed. That is the same consent problem showing up across the industry, where the burden of saying no keeps landing on the person whose likeness is at stake rather than the platform building the tool.

The stakes are not abstract. Nonconsensual AI imagery has already produced real harm and real litigation, including the case in which xAI sued a user over abuse images made with Grok. Meanwhile AI firms are moving in the opposite direction on identity, with Anthropic now able to ask Claude users for a government ID. Meta pulled this feature quickly, which is to its credit. The harder question is why it thought shipping it on by default was acceptable in the first place.