FTC AI Regulation Under Section 5: Justices Skeptical a 1914 Statute Reaches How AI Models Are Trained

WASHINGTON — The Federal Trade Commission's claim that its century-old authority over "unfair" practices lets it police how companies train artificial intelligence models met a skeptical bench on Tuesday, as several justices asked whether the agency was using a 1914 statute to regulate a technology that Congress has debated for three years without legislating. The case, Verrano Holdings v. FTC, is the first in which the Court has considered the agency's AI enforcement program, and the argument suggested that at least some of it may not survive.

"You are asking us to find a very large elephant in a very old mousehole," one justice told the government's lawyer, borrowing the Court's own phrase for agency claims of authority that outrun their statutes.

The order under review

Verrano, which operates a network of consumer apps, used the behavioral and location data it collected from users to train a model that it then licensed to advertisers and insurers. The FTC found that the collection was unfair under Section 5 because users could not reasonably avoid it and were not told how the data would be used, and its order did three things: it required Verrano to delete the data, it required the company to delete every model and algorithm trained on it — the remedy the agency calls algorithmic disgorgement, first imposed in Everalbum in 2021 and in Rite Aid in 2023 — and it barred Verrano for twenty years from training any model on consumer data without affirmative express consent. The court of appeals upheld the order; Verrano's petition challenged only the second and third provisions.

The arguments

• The company. Section 5 reaches the collection of data from consumers, Verrano conceded, but not what a company does with information once it holds it. Training a model is not an act or practice directed at consumers; it is internal engineering. Reading "unfair" to cover it would make the FTC the regulator of every data-driven product in the economy, a result that, under West Virginia v. EPA (2022), requires a clear statement from Congress. And after Loper Bright (2024), the agency's own interpretation of its statute earns no deference.
• The government. The unfairness prong has always reached the use of information, not only its collection: the injury to consumers whose data was taken without their knowledge continues as long as the data, or what was derived from it, is exploited. Section 5(n) supplies the limiting principle the justices sought — substantial injury, not reasonably avoidable, not outweighed by benefits — and the FTC applied it to one company's conduct on a record, which is adjudication, not the kind of economy-wide rule the major questions doctrine polices. The deletion remedy follows from the violation; a company may not keep the fruit of an unfair practice.

What the justices asked

The bench's questions clustered around three concerns. The first was scope: if training on unfairly collected data is itself unfair, several justices asked, what data practice is not — and does the answer change when the model is trained on public data, on licensed data, or on data collected lawfully but used for a new purpose? The government's answer, that each case turns on Section 5(n), drew the response that a test satisfied by every case is not a limit. The second was the remedy: two justices questioned whether ordering the destruction of a model, as opposed to the data, was a penalty rather than a cure, pointing to AMG Capital (2021), where the Court held that the FTC's power to seek injunctions did not carry with it a power to obtain monetary relief. The third was Congress: a justice observed that both chambers have held hearings on AI every session since 2023 and passed nothing, and asked whether the Court should infer that the FTC already had the authority or that Congress had declined to give it.

Not every justice was hostile. Others noted that the FTC has policed data practices under Section 5 for twenty-five years without objection from Congress, that the deletion of models built on unlawfully obtained data has been ordered in consent decrees many companies accepted, and that a ruling for Verrano would leave consumers with no remedy against the exploitation of data the agency had already found was taken unfairly.

You are asking us to find a very large elephant in a very old mousehole.

A justice, at argument

What is at stake

The FTC's AI enforcement rests on two legs. The deception cases — against companies that claimed to use AI when they did not, or overstated what their tools could do, the subject of the agency's 2024 enforcement sweep — are not before the Court and will continue whatever happens. The unfairness cases, which reach how data is gathered and how models are built, are at risk. A ruling that Section 5 does not reach training would end algorithmic disgorgement as a remedy and push regulation of model development to the states, where Colorado's law on high-risk automated decisions took effect this year and others are following, and to a Congress that has shown no appetite to act.

What happens next

A decision is expected by the end of June 2027. Companies with pending FTC investigations involving training data have asked the agency to hold them in abeyance; the agency has declined. Compliance advice in the meantime is unchanged: obtain consent for data used in training, document the basis for each use, and assume that the deletion remedy remains available until the Court says otherwise.

The Counsel will publish the decision and analysis when it issues. Companies building models on consumer data should take advice on consent and documentation now; the remedy remains available until the Court rules.

This report is for general information only and is not legal advice. Readers should consult a licensed attorney about their own circumstances. Docket numbers and citations are to the official reporters unless otherwise stated.