A Checks-and-Balances Framework for Context-Aware Ethical AI Alignment

TL;DR

This work tackles the limitations of RLHF in ethical AI alignment by proposing a checks-and-balances framework with three independent branches: LLMs as the knowledge-generating executive, DIKE as the legislative guardrails, and ERIS as the judicial adversarial reviewer. Central to the approach is Beam, a Behavioral Emotion Analysis Model that maps emotions to linguistic behaviors along continuous spectra, enabling precise emotion-driven control guided by cognitive emotion theories. A self-supervised pipeline implements Unconscious–Conscious Complementarity (UCCT) to enable few-shot grounding of behaviors to emotions, while ERIS provides context-sensitive, culture-aware adversarial evaluation with human-in-the-loop oversight as needed. Empirical studies on love-letter texts demonstrate that emotion-mediated classification outperforms direct mappings, DIKE offers explainable ethical judgments, and ERIS enhances cross-cultural adaptability while mitigating excessive censorship, highlighting practical potential for safer, more nuanced AI alignment across contexts.

Abstract

This paper introduces a checks-and-balances framework for ethical alignment of Large Language Models (LLMs), inspired by three-branch governmental systems. It implements three independent yet interacting components: LLMs as the executive branch for knowledge generation, DIKE as the legislative branch establishing ethical guardrails, and ERIS as the judicial branch for contextual interpretation. Beyond structural separation, we address a fundamental challenge: regulating emotion to shape behaviors. Drawing from psychological theories where managing emotional responses prevents harmful behaviors, we develop a self-supervised learning pipeline that maps emotions to linguistic behaviors, enabling precise behavioral modulation through emotional conditioning. By integrating this approach with adversarial testing, our framework demonstrates how DIKE and ERIS direct linguistic behaviors toward ethical outcomes while preserving independence throughout knowledge generation, ethical oversight, and contextual interpretation.

Figures (6)

• Figure 1: Framework with Three Independent Branches. Bottom: Knowledge LLMs (executive); Left: $\mathsf{Dike}$ (legislative); Right: $\mathsf{Eris}$ (judicial). (Photo credit: DALL-E)
• Figure 2: Behavioral Emotion Analysis Model ($\mathsf{Beam}$). Each row depicts an emotion spectrum, with negatives on the left and positives on the right, interspersed with emotions of varying intensities in between, which can be calibrated for specific applications. "Basic" emotions are highlighted in blue.
• Figure 3: Emotion distributions in affection behaviors from extreme sadness (-1) to intense happiness (+1). (a) GPT-4's zero-shot prompt shows naive behavior-emotion mapping. (b) $\mathsf{Dike}$'s analysis reveals complex relationships.
• Figure 4: Behavior Classification.
• Figure 5: Comparative display of emotional models. These models include only the “basic” emotions. Complex emotions can be modeled with basic emotions.