‘Systematic Alignment Decay’: a robust, technology-agnostic hazard associated with the advancement and proliferation of artificial intelligence

Abstract I distinguish a robust, systematic and possibly existential hazard to human wellbeing produced by AI advancement and proliferation. Behaviourally salient norms in adaptive systems (including, through various mechanisms, AI systems) function, amongst other things, as evolving responses to environmental conditions. Human-benefiting norms are selected for by conditions of fitness dependence upon, or kinship with, humans. AI advancement and proliferation tend to reduce such dependence, whilst AIs have little such kinship; AI advancement and proliferation may thus be expected to tend to reduce the internalisation of human-benefiting norms by AIs and AI-dominated institutions. Further, as such entities lack the quotidian human involvement that constrains other artificial agents, such as conventional institutions, whilst competing with humans for resources, they may be expected to inflict intolerable harms on humans under such conditions. Because this ‘Systematic Alignment Decay’ (‘SAD’) phenomenon arises from technology-agnostic selection pressures, it may be expected to result from sufficient AI proliferation and advancement whatever technical progress is made in ‘AI alignment’, even if no specially harmful ‘rogue’ AIs are deployed and if AI costs and benefits are fairly distributed. Further, such SAD outcomes may occur even without the development of systems exhibiting general or superhuman intelligence, though such advancements may help accelerate such outcomes. Contra certain extant narratives about possible futures lived alongside busy AIs, all likely sustainably tolerable futures require AI curtailment.