Legal AI Needs More Than Prediction: A Case for Reliability and Justification

Artificial intelligence has become remarkably adept at the mechanics of law. Modern systems can search millions of cases in seconds, extract legal knowledge from dense statutory text, and predict how courts are likely to rule on a given dispute. Yet a prediction, however accurate, is not the same thing as a legal justification. When an algorithm is deployed in legislation, law enforcement, or adjudication, the people affected must be able to understand whether its output actually aligns with legal reasoning and with the fundamental values that underpin the judicial system. That gap between statistical performance and normative accountability is the central concern of a new Perspective published in National Science Review.

In the article, Fang Wang reviews the two great paradigms of legal computability and argues that the field must move beyond mere technological empowerment toward a deeper integration of normative values with technological innovation. The argument arrives at a moment when courts and law firms around the world are adopting artificial intelligence at unprecedented speed, raising urgent questions about whether these tools can be trusted with decisions that shape liberty, property, and justice itself.

The dream of making law computable is far older than computing. The seventeenth-century philosopher and mathematician Leibniz envisioned legal rules expressed in symbolic form, so that conclusions could be derived through logical operations much like geometry. The advent of computers in the mid-twentieth century revived this vision and gave rise to rule-based reasoning, in which legal knowledge was encoded as formal rules and applied to facts through deduction. But law, unlike mathematics, does not rest on universally accepted axioms. Many legal concepts resist full formalization, and legal systems contain inherent inconsistencies and gaps. These structural limitations make it difficult, and often impossible, to derive sound legal conclusions through purely formal logical deduction alone.

A second, empirical paradigm emerged instead from real-world legal data, and it has developed along two key directions. Legal Judgment Prediction, or LJP, maps case texts to likely judicial outcomes, answering the question of what result is probable. Empirical Legal Studies, or ELS, examines causal and evaluative questions, helping to explain why particular outcomes occur and supplying evidence relevant to their assessment. Natural language processing has powered this transition: advances ranging from Word2Vec-based word representations and Transformer architectures to today’s large language models have dramatically improved the ability to process and analyze legal text, laying a technical foundation for downstream tasks such as judgment prediction and empirical legal analysis.

The practical consequences are already visible. In China, the national Smart Court has integrated more than 320 million pieces of legal data records, and the Shenzhen Intermediate People’s Court launched China’s first trial-specific large language model in 2024. In the United States, tools such as Harvey AI and CoCounsel are increasingly used by lawyers, and some state courts have begun introducing approved artificial intelligence products. In the United Kingdom, the Ministry of Justice launched an AI Action Plan for Justice in 2025. Legal computing is no longer a laboratory curiosity; it is reshaping the daily practice of justice systems worldwide.

Yet most artificial intelligence models, for all their power, reveal statistical correlations rather than legally sound reasons. The Perspective explains that existing research generally follows two approaches to interpretability: ex ante methods that embed legal logic into model design, and post hoc explanations that attempt to rationalize the behavior of already-trained models. Future systems, Wang argues, could go further by incorporating the logical relations among legal norms, factual elements, and practical reasoning in concrete legal scenarios, and by embedding judicial values directly into model and algorithm design rather than treating ethics as an afterthought.

To that end, the article proposes a value-guided framework situated within a society–law–ethics–algorithms–data system. In this framework, social values can guide model objectives and decision preferences; evolving legal norms can inform model knowledge and reasoning across temporal, spatial, and contextual conditions; and ethical norms can function as executable constraints within the system itself. Possible technical pathways include embedding these values and rules into loss functions, reward modeling, and constrained optimization, while robust optimization and careful constraint design may further strengthen model robustness. High-quality, well-structured datasets, built through bias mitigation mechanisms, are equally essential to ensure the quality of the inputs on which models are trained.

The Perspective also sounds a warning about a subtle psychological hazard: the anchoring effect. When artificial intelligence generates recommendations, those suggestions may bias subsequent human judgment, nudging judges toward algorithmic outputs and creating a path dependence that erodes independent reasoning. Reducing this risk requires a two-pronged response: algorithmic safeguards designed to prevent undue reliance on machine recommendations, and targeted training that strengthens judges’ capacity for independent, critical evaluation of what the algorithm tells them. Trust, in other words, must be earned through design and education, not assumed.

The framework is illustrated through three scenario-specific applications. In intelligent legislation, quantitative simulation and evaluation methods could assess the broader impacts and risks of legal rules, detect inconsistencies and gaps in the law, and support the recording-and-review mechanism for normative documents; quantitative frameworks for the proportionality principle could evaluate whether legislative measures are suitable, necessary, and properly balanced against the rights they restrict. In law enforcement, big-data supervision models in China have identified excessive or inconsistent penalties, while satellite remote sensing has helped procuratorial authorities detect environmental violations; these methods could extend to emerging fields such as data classification, tiered protection, and risk assessment under the Data Security Law, and federated learning offers a way to overcome data silos by enabling collaborative model training without centralizing sensitive data. In adjudication, judicial data could support transparent, explainable systems built around deep human–AI collaboration, in which human judgment retains central authority while algorithms and human expertise co-evolve: as algorithms process the “is” data they excel at, they can learn from how humans handle “ought” information, keeping computational results aligned with legal reasoning and judicial values.

The article closes with a caution and a note of optimism. Empirical findings are grounded in observed real-world data and should not be mistaken for universally valid truths; the patterns of past cases do not guarantee the justice of future ones. But the Leibnizian pursuit of computable law does not imply permanent infeasibility. Recent studies have begun exploring mathematical proof assistants, including Lean, to formalize and verify legal rules and procedures with machine-checked rigor. Combining empirical learning with formal reasoning, the Perspective suggests, may hold the greatest promise of all: a legal artificial intelligence that not only predicts what courts will do, but can show, in a form humans can verify and trust, why the law justifies it.

Subject of Research: Interpretability, reliability, and value-guided design of legal artificial intelligence systems

Article Title: Can legal artificial intelligence achieve reliability and justification?

Article References: Can legal artificial intelligence achieve reliability and justification?. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: legal artificial intelligence, legal judgment prediction, judicial values, interpretability, rule-based reasoning, empirical legal studies, large language models, smart courts, federated learning, formal verification, Lean proof assistant, human-AI collaboration