Meta Hid the Deepfake Tool, Kept Scraping Your Photos

Meta disabled the Instagram feature that let users generate AI deepfakes of public accounts without consent.

The company faced backlash, pulled the feature, and declared the problem solved. This is the story everyone told themselves—and also the story that obscures what actually happened.

Meta's feature made something visible that was already invisible. Users could now see, in real time, that their Instagram photos were being fed into an AI training system. They could watch the transformation happen. They could understand, finally, what "training data" meant: their face, their body, their image converted into a vector space and locked into a model they did not consent to build.

So Meta removed the visible part. The company kept the training data. Instagram photos continue to be scraped and fed into Meta's AI models, exactly as before. The deepfake generator was the symptom, not the disease—disabling it solved the optics problem by removing the evidence of the underlying practice rather than constraining the practice itself.

This is how institutional power operates when it has been observed. It does not change what it does. It changes what you can see.

Users can no longer see what was being done to their content because the tool that made it visible has been switched off. This is how institutional power operates when it has been observed. It does not change what it does; it changes what you can see. Consider what happened when hiring practices shifted from résumés to skill-based assessments—the companies that made the switch were often praised for becoming more equitable. They rarely disclosed what the assessments were actually measuring or whether the outcomes improved. They made the process less legible to the candidate. Meta's move follows the same logic. Transparency is a problem to be solved through opacity.

The real extraction continues in the dark. Users who understand this are facing a choice that extends far beyond Instagram. Every platform that trains on your data without explicit permission is operating on the same principle. Meta simply forgot to hide it, then corrected the mistake by hiding it again. The question is not whether to trust Meta.