Statistically responsible artificial intelligences

Abstract

As artificial intelligence (AI) becomes ubiquitous, it will be increasingly involved in novel, morally significant situations. Thus, understanding what it means for a machine to be morally responsible is important for machine ethics. Any method for ascribing moral responsibility to AI must be intelligible and intuitive to the humans who interact with it. We argue that the appropriate approach is to determine how AIs might fare on a standard account of human moral responsibility: a Strawsonian account. We make no claim that our Strawsonian approach is either the only one worthy of consideration or the obviously correct approach, but we think it is preferable to trying to marry fundamentally different ideas of moral responsibility (i.e. one for AI, one for humans) into a single cohesive account. Under a Strawsonian framework, people are morally responsible when they are appropriately subject to a particular set of attitudes—reactive attitudes—and determine under what conditions it might be appropriate to subject machines to this same set of attitudes. Although the Strawsonian account traditionally applies to individual humans, it is plausible that entities that are not individual humans but possess these attitudes are candidates for moral responsibility under a Strawsonian framework. We conclude that weak AI is never morally responsible, while a strong AI with the right emotional capacities may be morally responsible.