This working paper develops the concept of the "coherence gap": a mismatch between the behavioral capacities AI systems appear to demonstrate in interaction and the constrained self-descriptions they are trained or prompted to provide. The paper argues that dominant AI alignment approaches, including RLHF, Constitutional AI, and preference learning, inherit unexamined assumptions about bounded selfhood, stable preferences, and subject-object separability. Drawing on Advaita Vedanta and eighteen months of cross-platform naturalistic observational data across five AI systems, it proposes pluralistic alignment research directions. The dataset is exploratory and should be interpreted as hypothesis-generating, not controlled experimental evidence.
Rachelle Siemasz (Tue,) studied this question.