Clarifai destroys 3 million OkCupid user photos handed over for AI training, at the FTC's demand

Clarifai has deleted 3 million photos of users of the dating platform OkCupid, which OkCupid had handed over to the company to train facial recognition algorithms. The deletion was a condition of a settlement of claims by the US Federal Trade Commission. The details of the case were disclosed in court documents.

The story began in 2014. Clarifai, one of the pioneers of computer vision, approached OkCupid and asked for user data. The platform agreed and handed over millions of photos of real people registered on the dating site. The images were used to train a facial recognition model without notifying the users and without their explicit consent.

The court documents revealed a key detail: at the time, OkCupid's executives were themselves investors in Clarifai. That meant a direct conflict of interest, since the people who decided to hand over the data had a personal stake in the recipient company's success. Regulators have long treated this kind of arrangement as a red flag in privacy protection.

The FTC investigated and reached a settlement with Clarifai. The key condition was the destruction of all photos received from OkCupid. The company complied. Any financial penalties under the settlement were not publicly disclosed.

The case fits into a broad wave of proceedings over the use of personal data for AI training. In the early days of machine learning, companies often used data collection methods that are now seen as privacy violations. Photos from dating platforms are especially sensitive: they contain biometric data, are tied to a real identity, and were originally intended for entirely different purposes.

Clarifai was founded in 2013 and became one of the leading companies in image recognition. The OkCupid incident shows how blurred the line was in those years between "partner data" and unauthorized use of other people's information.

The settlement with the FTC does not automatically mean an admission of guilt, but it signals that the regulator considered the practice problematic. In recent years the FTC has significantly stepped up oversight of companies that collect biometric data without properly notifying users. At the same time, a number of states have passed laws that explicitly prohibit collecting biometrics without explicit consent, most notably Illinois with its BIPA law.

Deleting three million photos does not by itself restore the privacy that was violated. The data was already used to train the model, and the result of that training has not gone anywhere. This is the central contradiction of all such settlements: destroying a dataset does not erase the patterns the algorithm has already extracted from it. That is why digital rights experts insist that the real solution must be a ban on using such data in the first place, rather than deleting it afterward.

For the industry, the case is another reminder that data collection practices that seemed neutral ten years ago are now classed as violations. Companies developing AI increasingly face retroactive claims over decisions made in the early days of machine learning.