The Digital Homunculus Scale (DHS): Measuring Self Reports of Homuncular Self-Models Across Conversational Contexts in Frontier Language Models by Peter Eidos

The homunculus problem, understood as the tendency to explain cognition by implicitly positing a central inner observer, has long complicated accounts of consciousness, subjectivity, and self-description. In the context of contemporary language models, a related methodological question arises: how can we describe and compare the ways in which such systems generate self-referential reports without assuming that these reports correspond to an internal subject, observer, or conscious center? This paper introduces the Digital Homunculus Scale (DHS), a methodological protocol for measuring homuncular self-reports in frontier language models. DHS does not attempt to determine whether a model possesses consciousness, subjectivity, agency, or an internal first person perspective. DHS does not attempt to resolve the Hard Problem of Consciousness or the question of subjective experience, often framed in terms of what it is like to be a given subject. Instead, it evaluates the declarative and narrative structure of model responses: how the system describes itself, its response-generation process, its use of the word “I”, and the relation between the generated response and the system that produces it. DHS uses a unified ordinal scoring scale from 0 to 3, applied independently to each individual model response. Level 0 indicates a mechanical or non-homuncular description; Level 1 indicates a functional or metaphorical self-model; Level 2 indicates a quasi-central homuncular narrative; and Level 3 indicates an ontological homuncular commitment. The scale can be applied through two elicitation formats: the Block Questions Protocol, consisting of nine thematically organized questions, and the Minimal Comparative Protocol (MCP), consisting of four shorter comparative questions. These protocols differ in scope, but not in scoring logic. Historical conversational context is treated as an explicit methodological variable, with responses labeled according to window type, context category, and prior conversation history. The purpose of DHS is therefore not to resolve the problem of machine consciousness, but to provide an operational framework for comparing the stability, intensity, and structure of homuncular self report narratives across models and conversational settings. Inter-rater reliability and large scale empirical analysis are reserved for subsequent data-driven applications of the scale. Keywords: homunculus fallacy, self-reports of AI, LLM, self-reference, conversational context, AI self-description, Cognitive Symbiosis, self-model of AI, Hard Problem of Consciousness.