Data is at the heart of AI development. Developing AI models can be a resource intensive process, requiring large amounts of data and compute, and not all companies have the capacity to develop their own models. Some companies, which we refer to as “model-as-a-service” companies in this post, develop and host models to make available to third parties via an end-user interface or an application programming interface (API). For example, a company can train a large language model (LLM) and sell access to this model to businesses (online stores, hotels, banks, etc.) who apply it to customer service chatbots.
Like most AI companies, model-as-a-service companies have a continuous appetite for data to develop new or customer specific models or refine existing ones. This business incentive to constantly ingest additional data can be at odds with a company’s obligations to protect users’ data, undermining peoples’ privacy or resulting in the appropriation of a firm’s competitively significant data. This risk is particularly salient given that customers may reveal sensitive or confidential information when using a company’s models, such as internal documents and even their own users’ data. There’s also a risk that a model-as-a-service company may, through its APIs, infer a range of business data from the companies using its models, such as their scale and precise growth trajectories.
Model-as-a-service companies that fail to abide by their privacy commitments to their users and customers, may be liable under the laws enforced by the FTC. This includes promises made by companies that they won’t use customer data for secret purposes, such as to train or update their models—be it directly or through workarounds. In its prior enforcement actions, the FTC has required businesses that unlawfully obtain consumer data to delete any products—including models and algorithms developed in whole or in part using that unlawfully obtained data. The FTC will continue to ensure that firms are not reaping business benefits from violating the law.
Additionally, what a company fails to disclose to customers may be just as significant as what it promises. The FTC can and does bring actions against companies that omit material facts that would affect whether customers buy a particular product—for example, how a company collects and uses data from customers. The FTC brought an action against a company that, among other things, claimed not to use its facial recognition technology on consumers’ images unless consumers affirmatively chose to activate the feature—but omitted the fact that this was only true for consumers in some jurisdictions. This omission subjected the company to the same type of enforcement as if it had made an explicit misrepresentation.
Misrepresentations, material omissions, and misuse of data associated with the training and deployment of AI models also pose potential risks to competition. These tactics can undermine fair competition by, among other things, luring and locking in customers based on false promises or allowing dishonest businesses to get a leg up over honest businesses. Model-as-a-service companies that appropriate the competitively significant information of business customers may also run afoul of the prohibition against unfair methods of competition. In short, the data practices of model-as-a-service companies can violate antitrust laws as well as consumer protection laws.
There is no AI exemption from the laws on the books. Like all firms, model-as-a-service companies that deceive customers or users about how their data is collected—whether explicitly or implicitly, by inclusion or by omission—may be violating the law.
Thank you to staff from across the Office of Technology, the Bureau of Competition, and the Bureau of Consumer Protection who collaborated on this post (in alphabetical order): Krisha Cerilli, Jessica Colnago, Julia Horwitz, Amritha Jayanti, Stephanie T. Nguyen.
 Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT '21). Association for Computing Machinery, New York, NY, USA, 610–623. https://doi.org/10.1145/3442188.3445922
 For instance, users have concerns that companies will disclose user data to each other, (see, e.g., https://simonwillison.net/2023/Dec/14/ai-trust-crisis/); and they may have concerns that models have leaked users’ personal data to the public. See, e.g., https://www.computerworld.com/article/3711467/questions-raised-as-amazon-q-reportedly-starts-to-hallucinate-and-leak-confidential-data.html; https://www.fastcompany
 This includes consumers, small businesses, and enterprise users.
 Similarly, a model-as-a-service company that makes its service available to end-users may get access to people’s personal information, including health and financial data.