Given the current global trend of actively exploring the integration of AI into dispute resolution, we contend that, under existing theoretical and normative frameworks, there is as yet no sound basis for incorporating AI—least of all for replacing judges—into these processes. This is not to dismiss AI’s contribution to dispute resolution; rather, we seek to clarify how AI can be responsibly strengthened in this field. Rather than adopting simplistic technological skepticism, we develop a comprehensive theoretical framework that integrates perspectives from computer science and jurisprudence. By tracing the logical sequence of dispute resolution—goal setting, data input, subject identification, algorithmic processing, and the output and attribution of responsibility—we identify five fundamental issues. First, at the goal-setting stage, a fundamental conflict arises between algorithms, which aim to optimize efficiency, and the judiciary, which pursues multiple values (such as a just resolution and a swift one). Second, at the data-input stage, the principle of “garbage in, garbage out” manifests as the intractable and dynamically interactive problem of “algorithmic bias.” Third, at the subject-identification stage, allowing AI to replace human adjudicators—whether fully or partially—alters the judicial proceedings and undermines procedural justice. Fourth, at the algorithmic-processing stage, the black box and the hallucinations of deep learning are in sharp tension with the judiciary’s exacting demands for certainty and reasoned explanation. Finally, at the accountability stage, outsourcing judicial authority to private developers creates a supervisory vacuum and weakens the state’s liability for compensation. We therefore propose that the legal responses and rule-making needed to address these foundational issues be put in place before, not after, technological implementation.
Ma et al. (Wed,) studied this question.