Algorithmic Learning Foundations for Common Law

Jason D. Hartline, Daniel W. Linna Jr., Liren Shan, Alex Tang

Introduction

In common law jurisdictions, the legal system is not a static repository of rules but a dynamic, evolving process. Precedents are established through court rulings, which then guide future decisions. This process bears a striking resemblance to an online learning algorithm, where the judiciary acts as an agent attempting to learn an underlying "true" legal standard—a function $f(x)$—by observing data points generated through litigation.

However, this learning process is hindered by a significant economic barrier: the high cost of litigation. This paper provides a theoretical framework to analyze how these costs create a "failure of learning" and proposes algorithmic interventions to ensure the legal system remains accurate and adaptive.

The Problem: Litigation Costs and "Dark Data"

When two parties are in a dispute, they face a choice: go to court or settle privately. While a court ruling produces a public precedent that benefits society by clarifying the law, it is expensive for the individual litigants. Conversely, an out-of-court settlement is often more efficient for the parties involved but remains "dark data" to the legal system. Settlements do not produce public rulings or refine the existing legal model.

This creates a critical selection bias. If the legal system only learns from the rare cases that reach a courtroom, it may never develop accurate precedents for specific types of disputes where settlement is the rational economic choice. Using the tools of computational learning theory, the authors formalize this as a failure of the system to converge on the correct legal standards, leading to persistent legal inaccuracy.

The authors model the judiciary using an online learning framework characterized by the following components:

  1. The Model: Cases arrive sequentially. For each case $x$, the system must decide whether to compel a court appearance—generating a noisy data point $y = f(x) + \eta$—or allow a settlement based on the current best estimate of the law.
  2. The Loss Function: The system's performance is measured by "regret," defined as the squared error between the applied legal award and the "societally preferred" award, plus the physical costs of litigation.
  3. Informational Externality: A central insight of the paper is that litigants provide a public good (a precedent) at a private cost. Because individuals do not factor in this "informational externality," they tend to under-litigate from a societal perspective.

Proposed Algorithmic Interventions

To address the learning failure inherent in the status quo, the authors analyze three alternative selection algorithms designed to optimize the legal system's evolution:

  • Compelled Appearances (Explore-then-Commit): The system forces the first $m$ cases to go to court regardless of cost. This "exploration" phase builds a robust initial model before the system transitions to allowing settlements.
  • Incentivized Appearances (Subsidies): Instead of force, the system offers random subsidies to litigants to defray court costs. This effectively allows the state to "buy" data points to improve the global legal model.
  • Individual Accuracy (KWIK Learning): Utilizing the "Knows-What-It-Knows" (KWIK) framework, the authors design an algorithm that only permits a settlement if the system is highly confident in its prediction. If uncertainty is high, the case is compelled to court to ensure individual fairness.

Key Results and Insights

The paper's mathematical proofs yield several significant findings regarding the efficiency of legal systems:

  • Failure of the Status Quo: The authors prove that if litigation costs are non-zero and no intervention occurs, the system's regret is $\Omega(1)$. This means the law stops learning, and inaccurate outcomes persist indefinitely.
  • Efficiency of Intervention: By compelling a strategic number of cases (approximately $O(\sqrt{T})$), the system can achieve a vanishing per-case loss, reaching an optimal regret bound of $O(\alpha \sqrt{c_{max}/T})$.
  • The Cost of Individual Fairness: Ensuring accuracy for every specific case (uniform legal certainty) is significantly more "expensive" in terms of data than ensuring average accuracy. To guarantee individual accuracy with high probability, the system must compel up to $O(n^3/\epsilon^4)$ cases, representing a much higher sample complexity.
  • Subsidies vs. Compulsion: The research suggests that carefully designed subsidy programs can achieve optimal regret bounds while maintaining deterrent constraints, ensuring that individuals do not engage in unlawful behavior simply to receive a subsidy.

Limitations and Future Directions

While the model provides a rigorous foundation for understanding common law evolution, it relies on several simplifying assumptions. The current framework assumes that cases are "independent and identically distributed" (i.i.d.), whereas real-world legal trends often shift due to technological or social changes. Furthermore, the model assumes a discoverable "true" function $f(x)$, a concept that legal realists might contest as overly objective.

Future work could expand this model to account for non-i.i.d. case distributions and more complex, non-linear feature spaces. Additionally, exploring strategic behaviors—such as lawyers intentionally obfuscating facts to influence precedent—would bring the model closer to the complexities of modern litigation.

Summary updated

BibTeX
@inproceedings{hartline2022algorithmic,
  author = {Jason D. Hartline and Daniel W. Linna Jr. and Liren Shan and Alex Tang},
  title = {{Algorithmic Learning Foundations for Common Law}},
  booktitle = {CSLAW},
  year = {2022},
  doi = {10.1145/3511265.3550438},
}

Loading related content...

Figure