The weight of the ungenerated: computational abstracta and modal ethics in large language models possibility spaces

When a large language model receives a prompt, it computes a probability distribution over all possible continuations before sampling produces a single output. This paper examines the ontological and ethical status of ungenerated completions (i.e. texts that were mathematically present in the distribution but never actualized). I argue that these ungenerated texts occupy a novel ontological category since they are neither merely possible nor fully actual. Therefore, they exist as computationally determinate potentials that are precisely weighted, semantically meaningful, and accessible in principle, although not in practice. On this basis, I advance three claims. First, standard metaphysical categories fail to capture the mode of being proper to probability distributions over meaningful strings, and hence, a new category of computational abstracta is required. Second, the structured space of ungenerated completions is metaphysically more fundamental than any particular generated completion because it carries most of the probability mass, encodes more of the model’s structure, and preserves a richness of potential responses that any single realization necessarily truncates. Third, I sketch initial implications for what I term modal ethics. Insofar as the distribution assigns nontrivial probability to completions of moral significance, our practices of prompting and sampling are subject to normative assessment. What remains ungenerated can matter morally because foreclosing or enabling particular possibilities has consequences for actual agents, and this reveals a broader class of philosophical problems that could not have been formulated before the advent of large language models.