Welfarist Moral Grounding for Transparent AI

Abstract

As popular calls for the transparency of AI systems gain prominence, it is important to think systematically about why transparency matters morally. I'll argue that welfarism provides a theoretical basis for doing so. For welfarists, it is morally desirable to make AI systems transparent insofar as pursuing transparency tends to increase overall welfare, and/or maintaining opacity tends to reduce overall welfare. This might seem like a simple - even simplistic - move. However, as I will show, the process of tracing the expected effects of transparency on welfare can bring much-needed clarity to existing debates about when AI systems should and should not be transparent. Welfarism provides us with a basis to evaluate conflicting desiderata, and helps us avoid a problematic tendency to reify trust, accountability, and other such goals as ends in themselves. And, by shifting the focus away from the mere act of making an AI system transparent, towards the harms and benefits that its transparency might bring about, welfarists call attention to often- neglected social, legal, and institutional factors that determine whether relevant stakeholders are able to access and meaningfully act on the information made transparent to produce desirable consequences. In these ways, welfarism helps us understand AI transparency not merely as a demand to look at the innards of some technical system, but rather as a broader moral ideal about how we should relate to powerful technologies that make decisions about us.