What values should an agent align with?

Abstract

The pursuit of values drives human behavior and promotes cooperation. Existing research is focused on general values (e.g., Schwartz) that transcend contexts. However, context-specific values are necessary to (1) understand human decisions, and (2) engineer intelligent agents that can elicit and align with human values. We propose Axies, a hybrid (human and AI) methodology to identify context-specific values. Axies simplifies the abstract task of value identification as a guided value annotation process involving human annotators. Axies exploits the growing availability of value-laden text corpora and Natural Language Processing to assist the annotators in systematically identifying context-specific values. We evaluate Axies in a user study involving 80 human subjects. In our study, six annotators generate value lists for two timely and important contexts: Covid-19 measures and sustainable Energy. We employ two policy experts and 72 crowd workers to evaluate Axies value lists and compare them to a list of general (Schwartz) values. We find that Axies yields values that are (1) more context-specific than general values, (2) more suitable for value annotation than general values, and (3) independent of the people applying the methodology.

Public engagement in decision‐making regarding the management of the COVID‐19 epidemic: Views and expectations of the ‘publics’

“…With regard to the expressed issues on how to handle the different existing opinions amongst the publics, natural language processing and active learning techniques could be used 48,49 . For instance, Liscio et al 50 used these techniques to identify a list of values concerning lifting COVID‐19 measures in the Netherlands. All comments made during engagement could be used to create a value list concerning the designated topic to take into account by policy‐makers.…”

Contextualising Computational Thinking Disposition Framework From an Affective Perspective

“…The first suffers from asking too much of the humans who might not be ready to think about their values and specify them as we expect them to, while the second is prone to errors in the learning mechanism. While impressive value learning [Liscio et al, 2021] and value aggregation [Lera-Leri et al, 2022] mechanisms are being proposed, they are not errorfree and they do not deal with the complexity of value taxonomies. The introduction of these taxonomies introduces new challenges in this relatively young field, such as learning the importance of values, learning the relations between value nodes, learning the property nodes for some value concepts, and designing mechanisms to aggregate value taxonomies.…”

“…The increased interest in engineering values into AI has resulted in a range of technical challenges including how AI learns human values [Liu et al, 2019;Lin et al, 2018;Wilson et al, 2018;Liscio et al, 2021], how individual values can be aggregated to the level of groups [Lera-Leri et al, 2022], how arguments that explicitly reference values can be made [Bench-Capon and Atkinson, 2009], how decision making can be value-driven [di Tosto and Dignum, 2012;Chh-ogyal et al, 2019;Cranefield et al, 2017], and how norms are selected to maximise value-alignment [Serramia et al, 2020;Montes and Sierra, 2021;Sierra et al, 2021].…”

Identifying Context-Specific Values via Hybrid Intelligence

“…Contribution In this extended abstract, we summarize our work previously published at AAMAS and JAAMAS [9,10,12]. Our contribution in these papers is two-fold.…”