Moral Programming: Crafting a flexible heuristic moral meta-model for meaningful AI control in pluralistic societies

Abstract

Artificial Intelligence (AI) permeates more and more application domains. Its progress regarding scale, speed, and scope magnifies potential societal benefits but also ethically and safety relevant risks. Hence, it becomes vital to seek a meaningful control of present-day AI systems (i.e. tools). For this purpose, one can aim at counterbalancing the increasing problem-solving ability of AI with boundary conditions core to human morality. However, a major problem is that morality exists in a context-sensitive steadily shifting explanatory sphere co-created by humans using natural language – which is inherently ambiguous at multiple levels and neither machine-understandable nor machine-readable. A related problem is what we call epistemic dizziness, a phenomenon linked to the inevitable circumstance that one could always be wrong. Yet, while universal doubt cannot be eliminated from morality, it need not be magnified if the potential/requirement for steady refinements is anticipated by design. Thereby, morality pertains to the set of norms and values enacted at the level of a society, other not nearer specified collectives of persons, or at the level of an individual. Norms are instrumental in attaining the fulfilment of values, the latter being an umbrella term for all that seems decisive for distinctions between right and wrong – a central object of study in ethics. In short, for a meaningful control of AI against the background of the changing contextsensitive and linguistically moulded nature of human morality, it is helpful to craft descriptive and thus sufficiently flexible AI-readable heuristic models of morality. In this way, the problem-solving ability of AI could be efficiently funnelled through these updatable models so as to ideally boost the benefits and mitigate the risks at the AI deployment stage with the conceivable side-effect of improving human moral conjectures. For this purpose, we introduced a novel transdisciplinary framework denoted augmented utilitarianism (AU) (Aliman and Kester, 2019b), which is formulated from a meta-ethical stance. AU attempts to support the human-centred task to harness human norms and values to explicitly and traceably steer AI before humans themselves get unwittingly and unintelligibly steered by the obscurity of AI’s deployment. Importantly, AU is descriptive, non-normative, and explanatory (Aliman, 2020), and is not to be confused with normative utilitarianism. (While normative ethics pertains to ‘what one ought to do’, descriptive ethics relates to empirical studies on human ethical decision-making.) This chapter offers the reader a compact overview of how AU coalesces elements from AI, moral psychology, cognitive and affective science, mathematics, systems engineering, cybernetics, and epistemology to craft a generic scaffold able to heuristically encode given moral frameworks in a machine-readable form. We thematise novel insights and also caveats linked to advanced AI risks yielding incentives for future work.