Clarifai deletes 3 million OkCupid photos and trained facial-recognition models under FTC deceptive-practices case

Clarifai was ordered on April 21, 2026 to delete 3 million OkCupid user photos and destroy any facial-recognition models trained on them, settling an FTC deceptive-practices case. The remedy is the first significant US enforcement action that requires deleting the trained AI model, not just the training data.

What happened

• 2014: OkCupid provided user photos to Clarifai for what users understood to be research. Scale: ~3M photos.
• 2015-2023: Clarifai trained internal facial-recognition models on the dataset. Models were used in commercial products.
• FTC complaint: user consent did not cover the broader model-training uses. OkCupid’s user disclosures were narrow; Clarifai’s downstream use was not.
• April 21, 2026 order: delete the raw photos AND delete any models that were trained on them. Commercial products relying on those models must be withdrawn or retrained from scratch on consented data.

Why “delete the model” is the remedy that matters

Deleting training data after the fact is a standard remedy that rarely hurts. The model has already captured the pattern.

Deleting the trained model is structurally different. It forces the company to:

1. Destroy the commercial IP derived from the contested data.
2. Retrain from scratch on consented data.
3. Accept any capability regression from the smaller or differently-distributed replacement dataset.

This is the remedy AI ethics researchers have pushed for since 2019. April 21 is the first time it lands in a US enforcement action against an AI-first company.

Downstream ripple

Every AI company that scraped, licensed, or ingested data under one set of user-consent terms and later trained on it under different terms now faces a concrete precedent. The short list of exposures:

• Image generators trained on web-scraped imagery (Midjourney, Stability AI, Ideogram).
• Voice cloning models trained on podcast or YouTube audio under terms narrower than current commercial use.
• LLMs trained on forum, reddit, or user-generated-content corpora where original-poster consent is contested.

None of those cases are at the FTC action stage. Clarifai is. Plaintiffs in the pending Stability and Midjourney class actions will cite this remedy by Q2 2026.

What it changes for tool buyers

Near term, nothing on pricing or availability. Medium term, expect:

• Clearer training-data provenance disclosures from commercial vendors (Adobe Firefly’s “trained on licensed data only” position looks even more valuable).
• More retrained-from-scratch models where a vendor chooses to preempt rather than litigate.
• Higher enterprise procurement scrutiny on training-data licensing terms, especially for regulated industries (legal, health).

Open questions

• Does OkCupid face separate exposure for the 2014 data sharing? The FTC order addresses Clarifai’s downstream use; OkCupid’s original disclosure practices were the precondition.
• Does Clarifai retain non-facial-recognition models trained on the same photos (general CV embeddings, for example)?
• Precedent scope. Does the remedy carry to class-action private enforcement, or only to FTC cases?

Related

• AI chatbot privilege court rulings split