Alignment as Jurisprudence

Technical Summary: Alignment as Jurisprudence

Problem
The paper identifies a structural isomorphism between the field of jurisprudence (the study of how judges decide cases) and AI alignment (the science of getting AI models to conform to human values). Both fields face the fundamental challenge of governing powerful decision-makers (judges or AI agents) in the "unknown future." They must extrapolate from past data (precedent or training data) to novel situations where explicit rules are ambiguous or non-existent.

Current alignment techniques, primarily Reinforcement Learning from Human Feedback (RLHF) and its derivatives like Constitutional AI and Case-Based Reasoning (CBR), face two critical limitations:

1. Representational Deficits: Current methods often rely on majoritarian inputs or the values of small, elite groups (e.g., lab engineers), failing to adequately represent pluralistic societal values or protect minority rights.
2. Specification Deficits: Natural language principles (e.g., "be helpful," "be fair") are inherently vague. Without concrete context, models interpret these principles based on their underlying training data, leading to unpredictable applications in novel scenarios. The paper argues that existing alignment approaches struggle to scale these principles effectively because they lack a robust mechanism for specifying the content of values through examples and deliberation.

Methodology
The paper employs a comparative theoretical analysis rather than empirical experimentation. It maps specific technical alignment approaches onto established jurisprudential theories to identify structural parallels and potential cross-pollination.

• Alignment Techniques Analyzed:
  • RLHF: Training models via human preference rankings.
  • Constitutional AI (CAI): Using a set of human-written principles (a "constitution") to guide an AI's self-critique and feedback generation (RLAIF).
  • Case-Based Reasoning (CBR): Using a repository of specific cases and expert-identified dimensions to guide model extrapolation.
• Jurisprudential Frameworks Analyzed:
  • H.L.A. Hart's Positivism: The concept of "open texture" in language, where rules have a core of clear meaning but vague peripheries requiring judicial discretion.
  • Ronald Dworkin's Interpretivism: The theory that legal decisions must "fit" past precedents while also presenting the legal system in its "best light" according to abstract moral principles (justice, fairness).
  • Cass Sunstein's Analogical Reasoning: The theory that law relies on "incompletely theorized agreements" where actors agree on specific outcomes in particular cases without agreeing on high-level general principles.

Key Contributions
The paper proposes a bidirectional framework where legal theory informs AI alignment and AI tools inform legal theory.

1. Alignment as Jurisprudence (Applying Law to AI):

   • Constitutional AI + Dworkin: The paper argues that Constitutional AI should adopt a Dworkinian hierarchy. Instead of a flat list of principles, models should be imbued with "meta-principles" that dictate how to balance conflicting values (e.g., prioritizing minority rights over majoritarian preferences in specific contexts). To solve the vagueness problem, the paper suggests supplementing general principles with diverse, human-deliberated examples of their application, allowing the model to learn the "through-line" of a value across changing contexts.
   • CBR + Sunstein: The paper critiques current CBR approaches for relying on expert-selected dimensions of similarity. It proposes a modified CBR approach where the dimensions of analogy are not pre-selected by experts but emerge from public deliberation. By generating diverse cases based on selected principles and having humans deliberate on the outcomes, the system can form "incompletely theorized agreements." This allows the AI to learn low-level principles of similarity and distinction that are representative of a pluralistic society, rather than the biases of a specific expert group.

2. Jurisprudence as Alignment (Applying AI to Law):

   • Supplementing Existing Law: The paper suggests AI can act as an objective "aligner" for the legal system. Large Language Models (LLMs) trained on case law could analyze judicial opinions to determine if a judge's reasoning is a plausible extrapolation of precedent or merely a post-hoc rationalization of personal policy preferences. This could constrain judicial discretion and increase the representativeness of the law.
   • Complementing Existing Law: The paper posits that agentic AI systems could act as quasi-legal agents, empowering individuals to contract, negotiate, and participate in public deliberation at a scale and cost previously impossible. This would extend human capabilities and align the law more closely with individual preferences by lowering the friction of legal participation.

Results
As a theoretical essay, the paper does not present experimental results or quantitative data. Instead, its "results" are conceptual mappings:

• It demonstrates that Constitutional AI currently suffers from the same "open texture" problem Hart identified in law: principles are too vague to scale without concrete specification.
• It demonstrates that CBR currently lacks the pluralistic grounding Sunstein advocates for, as it relies on expert-defined dimensions rather than emergent public reasoning.
• It identifies that Dworkin's requirement for "fit" and "best light" offers a solution to the majoritarian bias in AI, suggesting a need for constitutional-level meta-principles that protect minority rights against majority rule.
• It identifies that Sunstein's focus on low-level agreements offers a path to pluralistic alignment where diverse groups can agree on specific AI behaviors without agreeing on deep philosophical principles.

Significance
The paper claims its primary significance lies in opening a dialogue between two fields that are structurally identical but currently operating in isolation.

• For AI Alignment: It argues that legal theory provides a mature toolkit for solving the scaling and specification problems of alignment. By treating AI models as judges and alignment as jurisprudence, researchers can adopt proven methods for handling ambiguity, pluralism, and the evolution of rules over time.
• For Jurisprudence: It argues that AI offers a new "black box" for analyzing the law. Because LLMs can process vast amounts of legal text and identify patterns of similarity and distinction that humans might miss, they can help clarify how legal reasoning actually functions, potentially resolving long-standing debates about the nature of legal interpretation.
• Societal Impact: The paper concludes that as AI systems take on increasingly essential adjudicatory and decision-making roles, understanding the relationship between law and alignment is essential. It suggests that a hybrid approach—combining high-level principles with case-based examples and public deliberation—is necessary to ensure that powerful AI systems remain aligned with human values in a democratic, pluralistic society.