The Ethics of Automating Legal Actors

Josef Valvoda, Alec Thompson, Ryan Cotterell, Simone Teufel

Abstract

This position paper critiques the current trajectory of legal Natural Language Processing (NLP), which has largely focused on automating judicial decision-making through Legal Outcome Prediction (LOP). We argue that the automation of judges in common law systems presents insurmountable technical and ethical challenges, including the lack of moral reasoning in current models and the risk of systemic brittleness. Instead, we propose a strategic shift toward automating the adversarial role of lawyers. By reframing legal NLP tasks to support "AI lawyer-assistants," the field can better align with the principles of the Rule of Law while leveraging the technical strengths of large language models.

The field of legal NLP is experiencing a significant "renaissance," driven by the rapid advancement of Large Language Models (LLMs). However, much of this research has converged on a single, problematic task: Legal Outcome Prediction (LOP). In LOP, machine learning models are trained to predict the final ruling of a judge based on a set of case "facts."

While this task is popular in benchmarks like LexGLUE and COLIEE, it rests on a fundamental misunderstanding of the legal process. This paper argues that the field is focusing on the wrong "legal actor." By attempting to automate the judge—the ultimate arbiter of law and morality—we face severe ethical and structural risks. A more productive and ethically safer path lies in automating the adversarial role of the lawyer.

The Technical and Ethical Limits of AI Judges

The push for "AI judges" faces two primary gaps: technical and ethical.

The Technical Gap

In common law systems, judges do not merely apply static rules to facts. They are responsible for shaping the law through moral reasoning, social intuition, and an understanding of evolving societal norms—faculties that current LLMs lack. Furthermore, judicial decisions often rely on multi-modal inputs, such as assessing witness credibility through vision and sound during a trial. Text-only models are fundamentally ill-equipped for this level of social intuition.

The Ethical Gap

Automating the judiciary poses a systemic risk to the Rule of Law. A centralized AI model used across a court system creates a "brittle" legal infrastructure. A single bias or "hallucination" within the model could propagate through the entire system instantly, without the check-and-balance provided by a diverse pool of independent human judges. Moreover, the accountability essential to justice is undermined when decisions are made by opaque algorithms rather than human actors who can be held politically and morally responsible.

The "Judge's Voice" Bias in Datasets

A core methodological critique of current legal NLP is the reliance on judge-centric data. Popular datasets, such as those derived from the European Court of Human Rights (ECtHR), typically use "facts" as summarized by the judge in their final opinion.

This creates a significant bias. Because a judge writes their opinion after reaching a decision, they naturally highlight the facts that justify their conclusion. When an NLP model "predicts" an outcome from these facts, it is not predicting the law; it is merely recognizing the judge's post-hoc justification. This makes the prediction task artificially easy and fundamentally unrepresentative of the actual legal process, where facts are contested and uncertain until the final ruling.

A New Paradigm: AI Lawyer-Assistants

We propose a shift from Replacement (AI taking over the judge's role) to Augmentation (AI assisting the lawyer). Lawyers are inherently "easier" to automate for several reasons:

  1. Clear Metrics: A lawyer's success is measured by persuasion and the strength of their arguments, rather than the state-mandated burden of being a final moral arbiter.
  2. Lower Ethical Stakes: An AI that generates a strong legal brief is a tool for a human advocate. It does not replace the final decision-maker, thus preserving the human-centric nature of the trial.
  3. Adversarial Balance: In an adversarial system, human lawyers act as a filter for AI-generated content, maintaining the "checks and balances" of the legal process.

To support this shift, we suggest reframing NLP tasks. Instead of binary classification (violation vs. no violation), we should move toward a three-way classification (negative outcome, null outcome, and successful claim) to better model the strategic perspective of a lawyer.

Strategic Recommendations

To move the field forward, we offer four key recommendations for the NLP community:

  • Incorporate Multi-modality: Legal NLP must move beyond text to include vision and sound to capture the reality of courtroom interactions.
  • Engage Domain Experts: Lawyers and philosophers must be involved in developing the "moral sense" and reasoning capabilities of legal models.
  • Prioritize Explainability: Models must be transparent so that the public can exercise moral and political scrutiny over legal decisions.
  • Maintain Diversity: We must avoid the concentration of power in a single model, ensuring a diversity of adjudicators to protect against systemic failure.

Conclusion and Future Work

The primary hurdle for this new direction is the data gap. While judicial opinions are public and plentiful, lawyer-generated texts like legal briefs are harder to find. Future work must focus on developing models that can find precedents relevant to a desired outcome (lawyer-centric search) and building datasets that include the "full facts" of a case before they are filtered through a judge's opinion. By pivoting toward "AI lawyer-assistants," we can develop tools that enhance the legal system without undermining the foundations of justice.

Summary updated

BibTeX
@article{valvoda2024ethics,
  author = {Josef Valvoda and Alec Thompson and Ryan Cotterell and Simone Teufel},
  title = {{The Ethics of Automating Legal Actors}},
  journal = {Trans. Assoc. Comput. Linguistics},
  year = {2024},
  doi = {10.1162/tacl\_a\_00668},
}
Figure