Pay With Your Data: Optimal Data-Sharing Mechanisms for AI Services

Abstract

Rapid advances in Machine Learning (ML) have led to a proliferation of Artificial Intelligence (AI) services offered by firms. To develop a valuable AI service, a firm must build an accurate ML model which, in turn, requires a large amount of training data. Present-day firms usually obtain this data by offering incentives to consumers to share their data during the initial development phase of the AI service, and then use that data to re-train the ML models to improve the quality of the service. Consumers, on the other hand, incur privacy costs for sharing their data. Inspired by AI services such as speech-to-text conversion offered by Google, and ChatGPT and DALL-E offered by OpenAI, we analyze two popular data-sharing mechanisms that firms employ in practice: manual data-sharing and algorithmic data-sharing. In the former approach, consumers decide the amount of data to share with the firm, whereas in the latter, the firm uses algorithmic data-redaction – an established approach used by technology firms such as Amazon, IBM, and Oracle – to identify and censor sensitive segments of data, and determine the amount of data collected from consumers. For both the data-sharing approaches, we obtain revenue-maximizing mechanisms for the firm and analyze the fundamental differences between the two approaches in terms of the revenue accrued by the firm, the consumer surplus, and the volume of data collected. Our analysis uncovers several interesting economic effects: For instance, we show that the firm can obtain a higher revenue with an inferior data-redaction algorithm and highlight two nuanced effects – namely, a weak privacy-cost-compensation effect and a strong data-collection effect – underlying this behavior.

Keywords: data-sharing mechanisms, artificial intelligence services, mechanism design

JEL Classification: D47, D40