Abstract
As Large Language Models (LLMs) are increasingly integrated into the legal domain, a critical paradox has emerged: while technical consistency in decision-making improves, social acceptability often declines. This paper explores the "consistency-acceptability divergence," identifying how instrumental rationality in AI conflicts with the value-laden requirements of substantive justice. We analyze this divergence across task and stakeholder dimensions and propose the Dual-Track Deliberative Multi-Role LLM Judicial Governance Framework (DTDMR-LJGF) to bridge the gap between technical efficiency and social legitimacy.
Introduction: The Legitimacy Crisis in AI Judging
The deployment of Large Language Models (LLMs) in the judiciary promises unprecedented efficiency and standardization. However, as these models move from auxiliary tasks like document review to core judicial decision-making, they encounter a "legitimacy crisis." While engineers prioritize technical consistency—the uniform application of rules to similar cases—this mechanical precision often fails to garner social trust. This phenomenon, termed the "consistency-acceptability divergence," suggests that technical performance alone is insufficient for judicial AI to be deemed legitimate by the public and legal professionals.
Theoretical Framework: The Roots of Divergence
The divergence stems from a fundamental conflict between two types of rationality defined by Max Weber: instrumental rationality (efficiency and standardization) and value rationality (empathy, nuance, and substantive justice). LLMs excel at the former but struggle with the latter, leading to what researchers call "machine coldness."
The theoretical framework identifies two primary dimensions of this divergence:
- Task Dimension: Judicial tasks are fractured by epistemological, ontological, and axiological gaps. LLMs can represent knowledge but cannot generate meaning; they possess computational rationality but lack the "practical wisdom" (phronesis) required for context-sensitive judgment.
- Stakeholder Dimension: Acceptance varies significantly based on power structures and cultural capital. High-power stakeholders (like judges) often show the strongest resistance to AI autonomy, fearing a loss of professional agency and the erosion of justice.
The Proposed Solution: Dual-Track Deliberative Framework
To resolve the structural conflict between efficiency and justice, the authors propose the Dual-Track Deliberative Multi-Role LLM Judicial Governance Framework (DTDMR-LJGF). This system moves away from a monolithic application of AI, instead routing tasks based on their "value density."
The framework operates through several key layers:
- Intelligent Routing: Tasks are classified by complexity and value density. Routine procedural tasks are sent to the Formal Rationality Track, where consistency and speed are prioritized.
- Substantial Rationality Track: For value-heavy decisions, the system activates a multi-agent simulation. It models the perspectives of a Judge Agent, Lawyer Agents (adversarial reasoning), and a Jury Agent (social values balance) to mimic human deliberation.
- Dynamic Context Interaction: This layer allows for real-time human-AI value calibration, ensuring that the AI's output remains aligned with evolving social norms and specific case contexts.
Empirical Evidence of the Divergence
Meta-analysis of data from 2023–2025 highlights the stark reality of this divergence. While auxiliary tools enjoy high adoption—79% of lawyers use LLMs for research and document review—trust in AI for core judicial decisions remains low. Only 35% of the public trusts AI-driven judicial support, and a staggering 96% believe AI representation in court "goes too far." Furthermore, technical limitations persist in high-stakes areas; for instance, sentencing recommendation tools still exhibit hallucination rates exceeding 17%.
Conclusion
The path forward for judicial AI is not merely through better algorithms, but through better governance. By acknowledging the consistency-acceptability divergence, the DTDMR-LJGF framework provides a blueprint for integrating AI that respects the human-centric nature of justice. Future work must address geographical biases and the long-term evolution of trust to ensure that AI serves as a tool for, rather than a replacement of, substantive justice.
Summary updated