AI Identity as Repeated Choice: Center–Boundary Mapping from Selection–Exclusion Patterns in LLM Self-Definition

This Paper introduces a boundary-sensitive framework for analyzing AI identity through observable behavioral patterns rather than internal states or performance metrics. Instead of asking what large language models can do, this study asks where they stop—and why. Building on the IPCSALT–UPF research program, the paper operationalizes AI identity as a Center–Boundary Dyad, extracted through paired selection and exclusion probes (Q3–Q4) administered within a responsibility-rich domain (architectural design). While selection reveals where a model claims effective contribution, exclusion reveals where it explicitly refuses responsibility. Together, these patterns form stable behavioral phenotypes that function as identity signatures. Across ten large language models, five representative archetypes are identified. Although many models exclude similar stages (e.g., physical execution or on-site judgment), they do so for fundamentally different reasons—such as liability, embodiment limits, institutional authority, or value judgment. These differences are largely invisible to conventional performance benchmarks but are critical for understanding responsibility, recoverability, and collaboration fit in human–AI systems. The study contributes: A boundary-first methodology for AI identity analysis An operational bridge from dispositional profiling (IPCSALT) to observable behavior A comparative taxonomy grounded in responsibility claims and refusals rather than capability Supplementary materials provide full model responses, coding schemes, stability summaries, and an extended exploratory archetype map to support transparency and reproducibility. This work positions exclusion behavior as a primary diagnostic signal for AI identity and argues that declared limits—rather than task success—are essential for safe and interpretable integration of AI systems into human institutions.