The Necessity and Possibility of Trustworthy AI

Abstract

Calls for trustworthy artificial intelligence (AI) have become increasingly common in recent years. This paper investigates the validity of the concept of trustworthy AI: Is it possible to call AI trustworthy without committing a category mistake, equivocation fallacy or undue anthropomorphization? And if so, is trust in AI a good pathway to ethical AI development and use? We start by analyzing the value of trust through the lens of its constitutive aim: The mitigation of the characteristic kind of risk that arises in the context of interaction with other agents in virtue of their autonomy (agential risk). Then, we describe the notion of trust both formally and substantively with the aim of establishing how well-founded trust can serve this functional role and what being a trustworthy agent amounts to. We identify a dilemma in the AI context: Trustworthy AI is necessary because AI agents are sources of agential risk which needs to be managed. But at the same time, AI agents cannot fulfill the preconditions necessary to be called trustworthy because they lack the capacity to possess normatively well-grounded motivations for their actions. We resolve this dilemma by explaining how AI agents’ aims can be normatively grounded through their creators and the design process, without simply reducing trust in an AI agent to trust in its creators. We conclude that trustworthy AI is a meaningful concept that can help structure our approach to ethically sound human-AI-interaction.