Artificial Intelligence, Values, and Alignment

Abstract

This paper looks at philosophical questions that arise in the context of AI alignment. It defends three propositions. First, normative and technical aspects of the AI alignment problem are interrelated, creating space for productive engagement between people working in both domains. Second, it is important to be clear about the goal of alignment. There are significant differences between AI that aligns with instructions, intentions, revealed preferences, ideal preferences, interests and values. A principle-based approach to AI alignment, which combines these elements in a systematic way, has considerable advantages in this context. Third, the central challenge for theorists is not to identify ‘true’ moral principles for AI; rather, it is to identify fair principles for alignment that receive reflective endorsement despite widespread variation in people’s moral beliefs. The final part of the paper explores three ways in which fair principles for AI alignment could potentially be identified.

AI Ethics and Value Alignment for Nonhuman Animals

“…Firstly, the introduced approaches can be again summarized by recapping two further wordings of Gabriel [17]:…”

“…He then elaborates on challenges for certain groups, such as antisocial psychopaths, but also children as well as nonhuman animals. Additionally, Gabriel mentions in his elaboration on AI alignment nonhuman animals and sentient beings in general initially, but not anymore when he explores the details of value extraction and aggregation [17]. Sarma and colleagues argue for an interdisciplinary approach for the AI value alignment problem, involving neuroscience in particular, which may lead to insights of what they call "mammalian value systems" [18,19].…”

“…First, it has to be attempted to precisely extract the values of all concerned beings; second, it has to be attempted to aggregate these values in a consistent manner; third, it has to be attempted to instill these values into AI systems. The first two subproblems have in common that they are normative, while the third one is technical [17].…”

“…The fourth option, which resembles Yudkowsky's approach of "Coherent Extrapolated Volition" (CEV) [23], and the fifth option address the shortcomings of the second and third option, but also have challenges. Therefore, Gabriel favors the sixth option that the AI does what it morally ought to do, but explains that the challenges remain to specify the values and to aggregate them [17].…”

“…Bottom-up approaches are also referred to as social choice and aim for AI systems which integrate individual views. These approaches carry the advantages of being democratic as opposed to autocratic structures, yet also a range of challenges regarding practicability and fairness [16,17,57].…”