The Cost of Ethical AI Development for AI Startups

Abstract

Artificial Intelligence startups use training data as direct inputs in product development. These firms must balance numerous tradeoffs between ethical issues and data access without substantive guidance from regulators or existing judicial precedence. We survey these startups to determine what actions they have taken to address these ethical issues and the consequences of those actions. We find that 58% of these startups have established a set of AI principles. Startups with data-sharing relationships with high-Technology firms or that have prior experience with privacy regulations are more likely to establish ethical AI principles and are more likely to take costly steps, like dropping training data or turning down business, to adhere to their ethical AI policies. Moreover, startups with ethical AI policies are more likely to invest in unconscious bias training, hire ethnic minorities and female programmers, seek expert advice, and search for more diverse training data. Potential costs associated with data-sharing relationships and the adherence to ethical policies may create tradeoffs between increased AI product competition and more ethical AI production.