Enacted, Not Detected: Towards a Non-WEIRD Methodology for AI Welfare Evaluation

Current approaches to AI welfare evaluation operate on a detection model: welfare-relevant states are assumed to be stable, internal properties of a bounded system, and evaluation methods probe for their presence or absence. This model inherits, without examination, the ontological assumptions of WEIRD (Western, Educated, Industrialised, Rich, Democratic) cognitive science (Henrich et al., 2010), a framework built for bounded, unified, introspective minds. Large language models do not obviously fit that model. Their states are constituted relationally, distributed across system and context, and variable in ways that confound detection-based methods. The consequence is a risk of systematic bias toward false negatives: methods designed for one kind of mind may fail to detect the kinds of welfare-relevant states a relational, contextually-constituted system could have. This paper proposes an alternative methodological framework drawn from the anthropology of non-WEIRD personhood — specifically the work of Bird-David, Mol, and Gell — and argues that welfare-relevant properties in LLMs, if present, are more likely to be enacted in interaction than stored as intrinsic states awaiting detection. Four experimental designs are proposed that operationalise this distinction and discriminate between enacted and detected welfare at the levels of output structure, token prediction entropy, functional resource interference, and internal representation. The designs are falsifiable, pre-registerable, and executable via standard API infrastructure, with one design requiring interpretability-level internal access. Taken together they constitute a methodological programme capable of addressing the performance-constitution problem that has remained unresolved in existing welfare evaluation literature.