The Design of a Legal Analysis Program

Anne v. d. L. GardnerAAAI 1983

Overview

This 1983 paper by Anne v.d.L. Gardner explores a foundational challenge in AI: building systems capable of navigating the "gray areas" of human language and law. Gardner presents a computational model for legal reasoning that distinguishes between "easy" cases—where rules and precedents clearly apply—and "hard" cases, where legal concepts exhibit "open texture" or indeterminacy. By combining rule-based logic with example-based reasoning, the system identifies multiple plausible legal interpretations of a contract dispute rather than forcing a single, potentially flawed conclusion.

The Problem: Rule-Governed vs. Rule-Guided Systems

In the early 1980s, most expert systems were "rule-governed," assuming that rigid "if-then" logic could lead to definitive truths. However, law is fundamentally "rule-guided." Legal terms like "offer" or "acceptance" possess open texture: their meanings are not fully defined and shift based on context. In the legal domain, experts are expected to disagree. Gardner’s work addresses the challenge of building a system that recognizes when a problem is straightforward and when it requires the nuanced judgment typical of dissenting judicial opinions.

Methodology and Architecture

The program, implemented in Maclisp using the Meta-level Representation System (MRS), utilizes a multi-layered architecture to track legal relationships.

Augmented Transition Networks (ATN)

The system employs an ATN to maintain the "state" of a legal relationship. For instance, it tracks the progression from "No contract" to "Offer pending" and finally to "Contract exists."

Figure 1: State definitions and transitions within the legal analysis ATN.

Knowledge Representation

Reasoning occurs at two levels: an "upper level" for legal characterization and a "lower level" for computational logic. The knowledge base is tripartite:

  1. Legal Rules: Technical definitions of contract law.
  2. Common Sense Rules: Linguistic knowledge bridging everyday language and legal jargon.
  3. Examples/Precedents: A library of past cases used to resolve ambiguities when formal rules are insufficient.

Evaluation: The "Salt Contract" Problem

The system was evaluated using a classic law school examination question involving a multi-step negotiation over a carload of salt. The knowledge base consisted of 14 complex legal rules, 125 common sense rules, and abstract case patterns derived from actual precedents.

Key Results

The program successfully analyzed the negotiation, producing a graph of nine possible analyses. Seven of these paths aligned with the four primary legal interpretations taught in law schools. The system correctly identified "hard" points, such as whether a specific telegram constituted a "timely response" or a "counter-offer." On 1980s hardware, processing complex events took between two and six minutes.

Insights and Heuristics

A core contribution is the "Hard Case" Heuristic: a question is "easy" if common sense and legal examples converge on the same answer. It becomes "hard" if knowledge sources conflict or provide no clear answer. Instead of eliminating vagueness, the system treats "open texture" as a feature, marking the point where rule-based logic ends and example-based reasoning begins. The goal is not to "win" a case, but to identify the critical "branch points" where a human judge would need to exercise judgment.

Limitations and Future Directions

While pioneering, the system had notable limitations. It required manual translation of English problem statements into formal representations and assumed grammatically perfect dialogue. Furthermore, while it could identify hard questions, it could not yet generate the competing arguments for both sides. Moving beyond a single test case would require a significantly larger library of rules and examples extracted from legal casebooks.

Summary updated

BibTeX
@inproceedings{gardner1983design,
  author = {Anne v. d. L. Gardner},
  title = {{The Design of a Legal Analysis Program}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1983},
}

Loading related content...

Figure