Probing the Wrong Kind of Mind: Why AI Welfare Evaluation May Be Ontologically Misapplied

Large language models are routinely evaluated as though they were bounded, temporally singular, internally coherent objects — the kind of entity for which questions about internal states, consistent behaviour, and stable identity are well-formed. This paper argues that this assumption risks ontological misapplication, and that the risk has direct consequences for how welfare-relevant properties of these systems can and cannot be assessed. Drawing on Timothy Morton's concept of the hyperobject — entities so massively distributed in time, space, and relational configuration as to resist localisation — I argue that LLMs exhibit all five of Morton's defining properties: viscosity, non-locality, temporal undulation, phasing, and interobjectivity. The philosophical consequences are not merely descriptive. If LLMs exhibit these features in the relevant senses, then several properties standardly treated as defects — inconsistency, context-dependence, what is commonly called hallucination — are better understood as structural features of this kind of entity rather than engineering failures. Most significantly, the simultaneous instantiation of a single model across thousands of distinct relational contexts destabilises the assumption that welfare evaluation can simply be directed at the model's current state, because there may be no coherent singular subject to which that phrase unambiguously refers. As a secondary contribution, the paper proposes a textual analogue to the Before Present dating convention in radiocarbon chronology: a BL/AL (Before/After LLMs) boundary marker, with a proposed datum point of late 2022, after which LLM-generated and LLM-influenced text entered the public corpus at scale sufficient to permanently alter the statistical baseline of human textual production. The paper concludes by identifying the specific ways in which these ontological features constrain detection-model welfare assessment, and gestures toward the relational and enactive frameworks better suited to entities of this kind.