Trustworthy AI Is Not Valuable

ABSTRACT ‘Trustworthy AI’ is a buzzword in AI research, with scores of scholars discussing whether AI systems can be trustworthy. This demand for trustworthy AI is intuitive, especially when AI is used in healthcare or criminal law, where the stakes are incredibly high. In this article, I push back on calls for trustworthy AI. Even if trustworthy AI can be developed, it will not offer people any benefits over AI that is reliable, transparent, non‐discriminatory, and safe. As a result, we should discuss AI in terms of those more specific properties, and think about how the people and institutions that use AI systems can be trustworthy. I offer two sets of arguments. On one hand, many theories of trustworthy AI reduce trustworthy AI to systems that possess other desirable traits, like reliability, transparency, and so on. Thus, the value of trustworthy AI is similarly reducible to the value of those traits. On the other hand, other theories define trustworthiness as reliability combined with some mental property, like goodwill. On those theories, the goods of trust – that is, the reason trust is distinctly valuable – are goods that are enjoyed within interpersonal relationships. They cannot be realized, or need not be realized, in human–AI interactions. Thus, we do not gain any distinct value from having trustworthy AI.