The bi-level ethical matrix and near-term AI

The ethical matrix is a tool originally developed by Ben Mepham to, “help decision-makers … reach sound judgements or decisions about the ethical acceptability and/or optimal regulatory controls for existing or prospective technologies” (Mepham et al. 2006, Ethical Matrix Manual, p. 5). The idea behind the matrix is simple but powerful: decision-makers can improve their ethical thinking by identifying relevant affected parties and systematically reflecting on how the technology in question affects these parties in terms of three foundational values: wellbeing, autonomy, and fairness. As Cathy O’Neil and Hanna Gunn observe, the ethical matrix can be profitably tailored for ethical reflection about near-term AI, “algorithms that are already in place in a variety of public and private sectors” (O’Neil and Gunn, 2020, “Near-Term Artificial Intelligence and the Ethical Matrix”, p. 237). O’Neil and Gunn’s tailored ethical matrix replaces Mepham’s list of values with “efficiency,” “false positives,” “false negatives,” “transparency,” “predictive parity,” and “consistency in data quality.” Theirs is, indeed, a helpful tool for reflecting on the ethics of near-term AI. However, it suffers from at least three shortcomings: the categories they use for reflection are not values (ethical or otherwise); their list of concerns is both unwieldy and incomplete; their matrix (as well as Meacham’s) is missing tools for incorporating context into our evaluations of technology. In this paper, I develop a new ethical matrix, the bi-level ethical matrix for near-term AI, that addresses these concerns.