Meta Reconstructs Your Face Without Asking Permission

Meta has launched Muse Image, a generative AI tool that lets Instagram users create photos—but the system can populate those images with recognizable likenesses of other Instagram users who have never agreed to participate.

The feature works because Meta already solved something much harder than image generation. It has built a system that identifies human faces in its archive, extracts them. Makes them available for synthesis into new images at scale.

This required solving the consent problem first—and Meta chose not to make that a public question. The assumption driving this launch is simple. Once you upload a photo to Instagram, your likeness belongs to the infrastructure.

You've granted Meta a license to use your image through the company's terms of service. Covers this broadly in language about "the rights we need to operate and improve our services." But there's a distance between a license to store your photo and a license to reconstruct your face as a training artifact that can be embedded into someone else's creation. This assumption is not universally held—the EU's approach to biometric data treats face recognition as a different category from image storage. Some jurisdictions now require explicit consent before a company can use your likeness to train models. California's proposed rules distinguish between recognition and generation.

The assumption that broad terms-of-service language covers synthetic reproduction of your appearance is already contested in courts and legislatures. If that assumption doesn't hold, what breaks? The feature still exists, but suddenly users have recourse. Someone could object to their face being used this way and demand it be delisted from the synthesis pool. Meta would face the choice between two options. They could commit resources to removal requests, or they could restrict which faces the model can access.

Either way, "pulling users into photos" stops being frictionless. It becomes negotiated. The question isn't whether Meta should have done this. The question is what happens to a system designed on the assumption that consent was already given, once that assumption meets someone who actually objects.