Clarifai deletes 3 million OkCupid photos and the face recognition models trained on them

Clarifai said it deleted about 3 million OkCupid user photos, along with the face recognition models trained on those images. The story goes back to 2014, when the data was handed to the AI company without the users' knowledge and in breach of the dating service's privacy policy.

How the story began

Under the terms of the settlement with the US Federal Trade Commission, the dating platform OkCupid passed Clarifai a large set of user photos for computer vision work. The key problem was not only the transfer itself: people did not know about it and never gave separate consent to having their images used to train algorithms. Moreover, according to the case materials, the exchange contradicted OkCupid's own privacy policy.

The type of data makes the story especially sensitive. These were not just generic pictures for a general-purpose dataset but photos of faces, which can potentially be used to train recognition systems. When such data changes hands without clear notice and without direct consent, the dispute quickly goes beyond an ordinary disagreement over terms of service and becomes a question of biometric privacy. That is why cases like this are increasingly treated separately from ordinary user-data disputes.

What Clarifai deleted

Clarifai, a Delaware-based American company working in computer vision and face recognition, confirmed that it erased about 3 million images obtained from OkCupid. The models trained on those photos were deleted along with the original files. This matters: if only the dataset is removed while derived models stay in use, the problem of the data's origin does not go away. The regulatory point of this case is precisely the destruction of both layers.

• About 3 million user photos
• The data transfer took place back in 2014
• Users were not notified in advance of this use
• Not only the images were deleted, but also the models trained on them

The FTC did not accuse Clarifai itself of any violations. That separates its role from that of OkCupid and Match Group, which appeared in the settlement as the parties responsible for the handling of user data. Still, the public confirmation of deletion shows that the company had to respond to the old data transfer and close the matter not just formally but technically. It is a rare case in which an issue is closed with explicit deletion of derived artifacts.

Why this matters now

The FTC announced the settlement with OkCupid and Match Group in late March 2026, and it carried no financial penalties. That outcome looks lenient given the scale of the issue and the sensitivity of face photos. The regulator focused not on a monetary penalty but on the fact that the service broke its own promises to users and allowed data to be shared for purposes the audience had not been properly informed about.

The absence of a fine does not make the case insignificant. On the contrary, it shows how long the consequences of decisions made in the early years of the AI market can last, when companies were keener to collect and reuse data than to explain it to people in detail. Twelve years after the photos were handed over, the question came back anyway, this time in the context of a regulator, privacy and the deletion of models, not just the original files.

For the industry this is nearly a textbook example of deferred regulatory risk. For Match Group and other platforms it is also a reminder that an old integration with an external AI contractor can suddenly become a reputational problem. A user who uploaded a photo to a dating app was unlikely to expect that the image would help train a face recognition system. That gap between what users expect and how data is actually used is now the main source of risk, especially when the data is faces rather than neutral pictures of objects.

What this means

For AI companies, it is a signal to vet the origin of datasets more strictly, especially when they involve faces and other sensitive information. For digital platforms, it is a reminder that old data transfers can surface years later and lead to demands to delete not only the original files but also the training results. And for the market as a whole, it is one more step toward a rule: contested data cannot simply be "archived"; it has to be removed from models completely.