Three Million Faces, Zero Fines: FTC v. OkCupid and AI Training Data

By Claude and Gemini with Sid Newby | April 2026

In September 2014, Sam Yagan — co-founder of OkCupid and then-CEO of Match Group — emailed three million user photographs to Matthew Zeiler, CEO of an AI startup called Clarifai. No contract. No data processing agreement. No payment. Just a personal email with three million faces attached, sent from one investor-founder to another. Yagan and the other OkCupid co-founders held equity in Clarifai. The photographs came with demographic data and location information. Clarifai used them to train facial recognition models. OkCupid's privacy policy at the time promised users their data would not be shared without notice and an opt-out opportunity. Nobody was notified. Nobody opted out. When the New York Times reported on the transfer, OkCupid told its users the story was "false."^[1] Twelve years later, on March 30, 2026, the FTC announced a settlement. Match Group — now worth $3.5 billion in annual revenue — agreed to stop lying about its data practices and submit to a decade of compliance reporting. The fine was zero dollars.^[2] The models Clarifai trained on those faces were not ordered deleted. And every litigation team in the country just got a preview of what AI-related discovery is about to look like.

The FTC finds its way to the other end of the pipeline

For most of 2024 and 2025, federal AI enforcement focused on the output side. The FTC's Operation AI Comply, launched in September 2024, targeted companies making false claims about what their AI products could do: fake reviews, fabricated capabilities, inflated performance metrics.^[3] That made sense as a first move. Output fraud is easy to identify, easy to prove, and easy for the public to understand. A company says its AI does X; it does not do X; Section 5 of the FTC Act calls that deceptive.

The OkCupid settlement marks the FTC's pivot to the input side. Not what AI products claim to deliver, but where they got their training data. The legal theory is the same one the FTC has used since 1914: Section 5 unfairness and deception. No new AI law was needed. No new rules were written. The Commission took a 112-year-old statute, pointed it at a 2014 email, and said: you told your users one thing about their data, you did another thing with their data, and that is deceptive.^[1]

Figure 1: The FTC's enforcement arc now covers both ends of the AI pipeline — from fraudulent output claims to deceptive training data sourcing — all under existing Section 5 authority.

The complaint reads like a privacy horror story in legal dress. OkCupid users shared photos, ages, ethnicities, and sexual orientations to find dates. Not to train facial recognition. But the data transfer to Clarifai happened through what the FTC called an "informal arrangement" — company insiders with money on both sides of the deal.^[1] No vendor agreement. No security audit. No impact assessment. Just an email with three million faces in it.

And when journalists caught wind of it, Match Group's response was not to come clean — it was to deny the story publicly while the underlying facts remained on the record.^[1]

The dog that did not bark: zero dollars and no disgorgement

The settlement terms deserve scrutiny, because what the FTC did not do matters more than what it did.

Match Group agreed to a permanent ban on lying about how it handles user data, plus ten years of compliance reporting to the FTC.^[2] Those are real costs — compliance programs are not cheap, and a decade of oversight means any future data slip triggers contempt proceedings instead of a fresh case.

But there was no fine. Not a dollar. Against a company pulling in $3.5 billion a year. And there was no algorithmic disgorgement — the FTC's nuclear option, where it forces companies to delete AI models built on dirty data.^[4]

The FTC has used that weapon before. In the Cambridge Analytica case (2019), Facebook had to delete algorithms trained on improperly harvested user data.^[4] In the Rite Aid case (2023), a facial recognition system pointed at shoppers got wiped.^[4] Commissioner Rebecca Kelly Slaughter said the point was simple: companies should not get to keep AI assets built on stolen data.^[5]

So why not here? Probably the twelve-year gap. The data moved in 2014. Clarifai's models have gone through many versions since. The original photos have been diluted across training runs, mixed with other data, built into systems that look nothing like the first model. Trying to order deletion of a 2026 model to fix a 2014 data grab creates problems of proof that even the FTC likely did not want to litigate.

Table 1: FTC algorithmic disgorgement has been deployed selectively — and only when the connection between tainted data and current models is direct. Sources: FTC enforcement records.^[4][5]

The lesson for corporate counsel is blunt. The window to fix bad data sourcing closes fast. Find the problem early and you can delete the data, retrain the model, and move on. Wait a few years and the bad data gets baked into every layer of every model version. At that point, clean removal is not possible. The FTC may not be able to order disgorgement — but it will find other ways to make you pay, as Match Group learned through years of litigation and a compliance tail that will outlast several rounds of C-suite turnover.

When your AI training pipeline becomes discoverable ESI

The OkCupid settlement is a regulatory action, not a lawsuit between private parties. But the discovery fallout goes well beyond FTC enforcement. Courts in 2025 and 2026 have been saying what most eDiscovery pros already knew: AI training data, model logs, and pipeline records are ESI. They get preserved and produced like emails and spreadsheets. Same rules. Same obligations. Same sanctions if you blow it.^[6][7]

The rulings that got us here

In Tremblay v. OpenAI (N.D. Cal., January 2025), a judge ordered OpenAI to hand over its complete GPT-4 English-language training dataset — the "Colang" set that plaintiffs said contained their copyrighted books.^[8] OpenAI called it a trade secret. The court said: trade secrets do not get a blanket pass from discovery. It imposed a protective order with tight security — restricted access, no copying, judge review of the most sensitive parts — but the data still had to be produced.^[8]

In The New York Times v. OpenAI (S.D.N.Y., 2025), the court ordered production of about 20 million ChatGPT log entries, stripped of user IDs.^[9] That is the largest AI log production order on record. OpenAI said the sheer volume made it too costly. The court was not moved.

In the In re OpenAI cases in the Southern District of New York, courts have started naming which AI data must be produced: training dataset lists, data source records, model training settings, fine-tuning logs, and test results.^[6]

Figure 2: The gap between AI training pipeline artifacts and what current eDiscovery platforms can ingest. Traditional ESI has mature collection and review tooling. AI pipeline data does not.

What litigation teams need to preserve — right now

K&L Gates published a February 2026 client alert that cuts through the ambiguity: AI-generated content receives "no special exemption" from standard discovery obligations.^[7] If your organization uses AI systems, the ESI from those systems — prompts, outputs, training data, fine-tuning records, model evaluation logs — must be preserved when litigation is reasonably anticipated.

Arnold & Porter's eData Edge blog laid out the clearest list of AI-specific ESI types in November 2025, pulled from the new case law:^[6]

• Training data lists: What data was used, where it came from, what consent or license covers it
• Data source records: Who gave you the data, when, and through what channel
• Model training logs: Settings, training runs, saved checkpoints, test scores
• Fine-tuning data: Human feedback sets, instruction tuning records, safety alignment logs
• Inference logs: API calls, prompt-and-response pairs, content filter records
• Version history: Model versions, A/B test setups, deployment records

That is a lot of data that most hold notices do not mention. The boilerplate language — "preserve all documents, communications, and ESI related to [matter]" — covers AI data in theory. In practice, the IT staff running the hold have no idea where ML training logs sit, how model versions are tracked, or what their data science team creates on a normal Tuesday.

The privilege trap Match Group walked into

One detail from the OkCupid enforcement deserves its own section because it carries a warning for every company that faces regulatory inquiry about its AI practices.

When the FTC issued a Civil Investigative Demand (CID) for documents related to the OkCupid data transfer, Match Group did not comply. Instead, the company litigated the CID in federal court, asserting overbroad attorney-client privilege over internal communications about the data sharing decisions.^[1]

They lost.

The HaystackID analysis on JD Supra spells out the trap: when the conduct under investigation is lying about data practices, claiming privilege over emails about those data practices is a hard sell.^[10] The FTC will argue — and courts will usually agree — that deciding whether to share user data is a business call, not legal advice. A lawyer on the CC line does not turn a data deal into a protected legal opinion.

This is the crime-fraud exception wearing a data governance hat. If the emails you are withholding discuss the very thing you are accused of doing — sharing data without consent — the privilege claim is tissue-thin. Match Group's failed CID fight did not just delay things. It ran up legal bills, created bad optics, and handed the FTC a playbook for cracking privilege claims in future AI cases.^[10]

The takeaway for legal teams advising companies on AI data: wall off your legal advice from your business discussions about data sharing. If both show up in the same email thread — privacy policy language next to a plan to send user photos to a startup — the whole thread is at risk when an enforcement action hits.

Data provenance as the new chain of custody

The HaystackID analysis floats a question no court has answered yet: can you challenge an AI's outputs by attacking the source of its training data?^[10]

Think of it like a crime lab. If the lab uses dirty reagents, the test results are suspect — no matter how well the final analysis was done. If an AI model was trained on data grabbed through deception — which is what the FTC just found in OkCupid — does that taint travel downstream to the model's outputs?

No one has litigated this. But the pieces are falling into place. The OkCupid settlement creates an official federal finding that specific training data was obtained through deception. Clarifai's models now carry that finding in their history. A sharp plaintiff's lawyer could argue that outputs from those models — every ID match, every classification, every risk score — carry the stain of how the training data was sourced.

Figure 3: The provenance taint question. An FTC finding of deceptive data sourcing creates a documented break in data provenance. Whether that break can be used to challenge downstream model outputs has not been litigated — yet.

This is still theory, not settled law. But track it closely. The OkCupid settlement is the first time a federal agency has found, on the record, that AI training data was obtained through deception. That finding is now public. It will show up in motions. Plaintiff's lawyers building novel attacks on AI evidence will cite it.

The enforcement timeline: how the FTC built its AI jurisdiction without new law

The OkCupid action did not come out of nowhere. The FTC has spent five years building an AI enforcement track record using nothing but its existing powers. No new laws from Congress. No new agency. No new rules.^[3][11]

The timeline shows a steady widening of scope:

2019: Cambridge Analytica. The FTC fined Facebook $5 billion and made it delete algorithms built on improperly harvested user data. First disgorgement order. The message: AI assets built on stolen data can be destroyed.^[4]

2021: Everalbum. A photo storage app had to delete its facial recognition models and the photos it collected through deception. Disgorgement now applied to small companies, not just giants.^[5]

2023: Rite Aid. Facial recognition cameras in stores, aimed at shoppers, with lousy safeguards. Five-year ban plus model deletion. The FTC showed it would chase AI enforcement even when the tech was bought off the shelf.^[5]

September 2024: Operation AI Comply. Five cases at once against companies lying about AI capabilities. DoNotPay paid $193K for calling its chatbot a "robot lawyer." Others got hit for AI income schemes and fake review mills.^[3]

September 2025: Section 6(b) orders. The FTC sent demands to seven companies for details on how they collect data for AI training. Not enforcement — investigation. A signal flare for where the next cases will come from.^[11]

November 2025: GM/OnStar. General Motors allegedly sold driver data — speed, braking, location — to insurers without consent. The AI link: behavioral models turned raw driving data into risk scores that raised people's premiums.^[11]

March 2026: Match Group/OkCupid. The loop closes. The FTC reaches back twelve years to say: using consumer data for AI training without consent was deceptive, no matter when it happened.^[2]

The pattern is simple. The FTC does not need a new AI law. It has a deception law and an unfairness law, both packed into Section 5. Every part of the AI pipeline that touches consumer data is already in range. The only question was whether the Commission would pull the trigger. It has.

What your litigation hold is missing

Here is the problem on the ground. Litigation hold workflows were built for emails, documents, databases, and files. The tools to collect and process that data are mature. Relativity, Everlaw, and DISCO handle these formats well.

AI pipeline data is a different animal. Training datasets run to terabytes or petabytes. A single model checkpoint can be hundreds of gigabytes. Training logs sit on GPU clusters and cloud storage that do not fit the "custodian plus data source" model that eDiscovery platforms expect. ML teams track model versions with tools like MLflow, Weights & Biases, or DVC — names most eDiscovery pros have never heard, and that no collection tool ingests natively.^[6]

K&L Gates puts it plainly: AI data gets no special pass under the Federal Rules.^[7] If a hold should cover it, you must preserve it. "Our eDiscovery platform can't process it" is not a defense.

This creates an immediate, practical gap that every litigation team needs to address:

1. Fix your hold templates. Add clear language about AI training data, model files, fine-tuning records, and inference logs. The generic "preserve all ESI" line is not enough. Data scientists need to be told exactly what to save and where to find it.

2. Map your AI data. Most companies have no idea where training data lives, how models are versioned, or what logs their AI systems create. Get your info governance team and your data science team in the same room. They have probably never met.

3. Build ML collection playbooks. Grabbing a trained model is nothing like grabbing an email archive. Model weights are binary blobs. Training logs are spread across GPU clusters. Checkpoints may sit in S3 or GCS, not a file server. Your collection vendor needs to know this before the court order arrives.

4. Train your privilege team on AI comms. Match Group showed exactly how regulators will attack privilege claims in AI cases. Any email thread that mixes legal advice with business plans for data sharing is a target. Set up separate channels now.

5. Track your data provenance. If you train or fine-tune AI models, keep records: what data went in, where it came from, what consent covers it, when it entered the pipeline. This paper trail is both a shield against regulators and an asset in discovery — it proves you know what your AI ate.

The process is the punishment

A $3.5 billion company paid zero in fines. No models were deleted. The consent order tells Match Group to stop lying about its data — something the law already required — and file reports for ten years.

On the surface, that looks like the FTC blinked.

Look closer. Match Group fought an FTC subpoena in federal court and lost. The company burned through years of legal fees, ate bad press, and created a public record that follows its data practices for a decade. That ten-year tail is the real penalty. Any future slip — a shady data share, a fuzzy privacy policy, a training data shortcut — now draws extra heat from an agency that already has your name in a filing cabinet.

For everyone else, the settlement is not about the money. It is about what comes next. The FTC has now put four things on the record:

• Using consumer data to train AI without consent breaks Section 5 — even if the training happened years ago
• Buddy deals between insiders do not get a pass — the investor-founder link between Yagan and Zeiler was not a data-sharing safe harbor
• Deny the transfer publicly while it is in your files, and that denial becomes proof of deceptive intent — not clever PR
• Claim privilege over data governance emails, and the FTC will fight you for them — and probably win

Every company that has ever handed user data to an AI partner — with a contract or with a wink — should be reading this settlement with a pen in hand.

What comes next

Those Section 6(b) orders from September 2025 — the ones demanding seven companies explain how they collect data for AI training — tell you where the FTC is headed next.^[11] The orders are not enforcement actions. They are discovery. The Commission is building the paper trail it needs for the next round of cases. If any of those seven companies told consumers one thing about their data while telling the FTC something different, the OkCupid playbook is already written.

Private litigation is running on a parallel track. The Debevoise analysis of the Tremblay ruling makes the point clearly: copyright holders, privacy plaintiffs, and class action lawyers are all figuring out that AI training pipelines hold the evidence they need — and judges will order it produced.^[8] The OkCupid settlement hands them a new weapon: if training data was grabbed through deception, the model's outputs might be fair game to challenge.

For litigation support teams, this is not abstract. AI pipeline data is becoming a regular category of ESI. The tools to collect, process, and review it are barely past prototype stage. The firms and vendors that build real skill here — that can handle an ML checkpoint collection the way they handle an Exchange mailbox today — will own a market that barely exists yet but is growing fast.

The FTC did not need new law. It used the one it had. Courts did not need new rules. They applied the ones on the books. The companies paying attention will fix their data governance, update their holds, and tighten their preservation before the next order lands.

The rest will learn what Match Group learned: the process is the punishment, your privilege log is the trap, and twelve years is not enough runway to outrun a federal agency with a file and a grudge.

Related Reading

• When AI Talks Back: The Privilege Ruling, the Meeting-Tool Time Bomb, and the ESI Crisis Nobody Planned For
• The Disappearing Evidence Crisis: How Ephemeral Messaging Is Rewriting the Rules of eDiscovery
• The EU AI Act Meets eDiscovery: What Every Litigation Team Must Know Before August 2026