The deployment of advanced AI systems in high-impact decision contexts has intensified concerns regarding alignment, governance, and misuse. Current approaches predominantly conceptualize AI-related risk as a property of model behavior, emphasizing output alignment, constraint enforcement, and external oversight mechanisms. While these strategies address important failure modes, they remain structurally incomplete in contexts where AI systems function primarily as decision-support tools for human actors with concentrated authority. This paper argues that a significant class of AI-related risk arises not from model misbehavior, but from progressive degradation of human judgment under conditions of AI-amplified decision power. In environments characterized by irreversibility, asymmetric impact, and limited corrective feedback, sustained interaction with highly capable AI systems can systematically narrow reasoning, reinforce overconfidence, and attenuate sensitivity to human consequences, even when system outputs remain formally aligned. We introduce an architectural design space for internal ethical counterweights in AI systems. These counterweights are conceived as autonomous, non-task-oriented subspaces that operate alongside operational AI cores to detect structural risk conditions associated with judgment degradation and to modulate system interaction accordingly. Rather than enforcing normative outcomes or restricting system capabilities, ethical counterweights introduce persistent internal friction through graduated output modulation, reflection prompts, and uncertainty amplification. The paper does not propose a universal ethical doctrine or a single implementation strategy. Instead, it delineates multiple construction pathways—policy-driven, model-based, and hybrid—and analyzes their respective trade-offs in terms of adaptability, auditability, and governance. By reframing alignment as a problem of judgment stabilization under amplified power rather than output control alone, this work provides a conceptual foundation for integrating internal ethical friction into AI-assisted decision-making systems operating in high-impact domains.