Instagram shuts off the AI tool that made deepfakes of public accounts

The feature lasted only a few days. Meta introduced it this same week and, after considerable pushback, switched it off: it let users generate AI images from the content of any public Instagram account just by tagging it. As it was set up, anyone could feed the generator with the photos of someone else's profile without asking the owner for permission, and that was exactly the trigger.

The Verge summed it up in its headline with no euphemisms: the tool let users "make AI deepfakes of public accounts." The Verge detailed that content from any open profile could be used in AI creations without authorization from the account owner, and that the reaction was strong enough for Meta to turn the feature off in the same week it had announced it.

The design that sparked the backlash

Generating an image from someone else's photo is not new, and the quality of these models has been high for a while. What changed here was who decided. The system did not consult the source profile: tagging it was enough for its content to enter the process. In other words, the person whose image was being remixed was the last to find out, if they found out at all.

That distribution of power is what turns a "creative tool" into a serious problem. When permission is presumed rather than requested, the practical result is that anyone can produce synthetic versions of another person from their public photos. That is where the word deepfake stops sounding like an exaggeration and describes precisely what the feature enabled.

Why "public" does not mean "free to remix"

The underlying misunderstanding is old and still current. Having a public account means accepting that others see and share what you post, not that your face can become the base material for images you have neither created nor approved. For creators, journalists, activists or anyone with a public presence, that distinction marks the line between distribution and impersonation.

The risk is not only reputational. A synthetic image built from real photos can be taken out of context, attribute to someone gestures or situations that never happened, and circulate with no visible mark that it is artificial. That the starting point is "a public account" does not reduce that harm: it amplifies it, because there is more material available and easier to find.

A pattern we have seen before

The cycle feels familiar. A large platform launches a generative feature in a hurry not to fall behind, the permission design falls short, criticism arrives, and the feature is pulled or trimmed. It already happened with other image and voice tools over the past year. The novelty here is the speed: days passed between the announcement and the reversal, not months.

That speed suggests the problem was not caught in time inside the company, or that it was caught and shipped anyway. Neither reading reflects well on the prior review process. For users, the uncomfortable message is that the rules about the use of their content can change from one update to the next.

What the episode leaves behind

Meta has backtracked, but it has not explained whether it is preparing a version with explicit consent or what controls it would add. For the sector, the lesson is concrete: in features that generate images of people, the subject's permission cannot be an optional setting hidden in the configuration. It has to be the starting point of the design.

At ElephantPink we see it often when we help integrate generative models into real products: the hard part is almost never the generation, but deciding who has the right to use whose image. Meta has just reminded everyone, in record time, how costly it is to skip that question.