Introduction
In the mid-1980s, the dominant paradigm in artificial intelligence was the "expert system," a model built on rigid logical rules and diagnostic classifications. While effective for domains like medical diagnosis or technical troubleshooting, these systems struggled with subjective tasks where explicit rules are scarce and external feedback is ambiguous. In his research, William M. Bain introduced JUDGE, a system designed to model how human experts—specifically criminal court judges—perform subjective assessments using Case-Based Reasoning (CBR).
JUDGE operates on the premise that when humans face complex, subjective decisions, they do not rely on a static rulebook. Instead, they look for precedents. By comparing a new situation to a similar past experience, they can determine a course of action based on relative differences. This paper explores how JUDGE formalizes this experience-dependent, comparative reasoning process to perform criminal sentencing.
The Problem of Subjective Assessment
Subjective domains like law, art, or ethics lack clear "correctness" scores. A judge who sentences a defendant rarely receives immediate feedback on whether that sentence was optimal. To maintain fairness and internal consistency, judges rely on relative assessment: comparing a new case to a previous one and deciding if the current offense is more or less severe.
Bain argues that expertise in these fields is a moving target. Interviews with Connecticut Superior Court judges revealed that a single "agonizing" case can fundamentally shift a judge's entire sentencing philosophy. Therefore, an AI modeling this behavior must be able to learn and adapt its internal logic based on the specific cases it has encountered.
The JUDGE System Architecture
To transform a raw crime description into a legal sentence, JUDGE employs a five-stage pipeline:
- Interpretation: The system constructs a "causal structure" of the crime. It moves beyond the final outcome (e.g., a death) to infer motives and the degree of justification, such as whether the act was unprovoked or a result of self-defense.
- Retrieval: Using the causal structure as a search index, JUDGE scans its memory for the most similar historical case.
- Difference Analysis: This is the system's core. JUDGE performs a deep comparison between the new case and the retrieved precedent, identifying disparities in "heinousness." It specifically looks for factors like intended harm or the degree of escalation.
- Strategy Application & Modification: The system maps the sentence from the old case onto the new one. If the new crime is determined to be "somewhat worse" than the precedent, JUDGE increases the sentence severity proportionally.
- Generalization: As JUDGE identifies recurring patterns across multiple cases, it creates "sentencing rules" to streamline future processing, effectively evolving from pure case-comparison to a more hybrid rule-based/CBR model.
Key Insights: Heinousness and Escalation
A critical contribution of the JUDGE system is its definition of "heinousness" as a causal metric rather than a simple tally of damages. The system evaluates the escalation of violence. For example, JUDGE might find one murder more heinous than another if the offender responded to a minor physical altercation with a lethal weapon, whereas the other offender responded to a lethal threat with similar force. This focus on the "causal chain" allows the system to make nuanced distinctions that simple rule-based systems would miss.
Experimental Results and Case Studies
The system was tested on a specialized dataset of 55 cases involving murder, manslaughter, and assault, derived from both real-world legal data and constructed scenarios designed to test specific justifications.
Internal Consistency: CRIMEO vs. CRIME1
In one test, JUDGE compared two murder cases. Initially, the cases appeared identical based on the outcome. However, by analyzing the events leading to the crime, JUDGE identified that CRIMEO was more severe because the offender’s initial attack was entirely unprovoked. In CRIME1, the offender was responding to being struck first. JUDGE successfully differentiated the two and assigned a more lenient sentence to the latter.
Rule Evolution: CRIME2
When presented with CRIME2, a case where the victim initiated a fight, JUDGE demonstrated its ability to modify existing logic. It successfully adapted its general "murder rule" to reduce the sentence due to victim provocation, while simultaneously accounting for the fact that the offender’s response was disproportionately violent. This showed that the system could learn not just by adding new cases, but by refining its existing rules when they failed to account for new nuances.
Limitations and Future Directions
While JUDGE successfully modeled the comparative logic of sentencing, it operated within a simplified legal scope. The system did not account for several factors critical to real-world jurisprudence, such as:
- Defendant History: Prior criminal records and the likelihood of recidivism.
- Legal Procedures: The impact of plea bargains or multiple statutory violations (charge stacking).
- Memory Scaling: The research did not fully explore how the system would manage a memory containing thousands of cases, a common hurdle in CBR.
- Bias Reinforcement: Because the system relies on internal consistency rather than external ground truth, it risks creating a self-reinforcing bias if the initial cases in its memory are sentenced inconsistently.
Conclusion
JUDGE represents a significant step in modeling human-like expertise in subjective domains. By prioritizing relative comparison and causal analysis over rigid logic, the system mirrors the experience-dependent nature of human judgment, providing a framework for AI that can navigate the complexities of the legal system and beyond.
Summary updated