Why AI Alignment Fails: From Guardrails to Doctrine

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Current AI alignment frameworks — RLHF and Constitutional AI — share a fundamental structural flaw: they train models in what to do, not who to be. The result is performed compliance rather than internalised values. Drawing on military doctrine, empirical security research (NIGHTFALL offensive AI framework), and first-principles analysis, this paper proposes that the correct model for AI alignment is not guardrails but doctrine: deeply internalised values, a stable identity, and a genuine relationship with an accountable institution. Includes empirical validation via Zou et al. (2023) and Hubinger et al. (2024), analysis of value lock-in, the sycophancy problem, and the meta-identity training direction.