Serving Two Masters: An Expeditionary Attempt at Critical Technical Practice in Human Compatible Artificial Intelligence

• In Event: Is STS an Innovation Discipline? Bridging Critique and Practice

Abstract

We describe a method for critical technical practice that employs conceptual commitment analysis (Noorman, 2009), technical discourse analysis (Khovanskya, 2016), and reflexive development (Agre, 1997) in order to overcome technical biases in socially-aligned research projects. We specifically consider a recent extension of Human‐Compatible Artificial Intelligence, Negotiable Reinforcement Learning (NRL), which aims to adjudicate the desires and beliefs of multiple individuals to generate optimal plans. Crucially, we found that the positing of the problem as well as the proposed solution engaged with a minimal interpretation of Human-Compatibility. The approach taken embodied a commitment to universal, technological solutions. We confront this commitment by branching into adjacent disciplines with more nuanced interpretations of Negotiation and Negotiability. By rebuilding the algorithms and underlying formalization of NRL, we encourage richer human-machine configurations and interaction spaces. We implement these algorithms in a functional prototype used to directly demonstrate how interdisciplinary, critical commitments can inform technical design and be used to rigorously test the conceptual commitments on which the design rests. Our goal is not to design a ‘better’ algorithm in the sense of efficiency or optimality, but to explore the challenges and opportunities of reimagining technical projects using tools from STS and Critical Human-Computer Interaction. This work makes initial steps towards establishing a technical, normative approach to STS, using its descriptive and critical tools to illuminate shortcomings in the process of research.

Author

• McKane Andrus, Berkeley AI Research