1. Introduction: The Era Opportunity of Generative AI in Judicial Decision-Making

Generative AI refers to technologies that generate relevant, precise, and creative content such as text, images, sounds, videos, and code based on algorithms, models, and rules, thereby achieving a level of human-like creativity. From the first AI that passed the “Turing Test” in 2014, Eugene Goostman, to November 2022, when OpenAI launched the generative AI ChatGPT, which completed an iterative improvement to GPT-4 just four months later, its remarkable performance and multimodal generation capabilities have attracted over 100 million users in a short time[1]. The success of ChatGPT has also triggered a new round of AI technology competition, following ChatGPT, Google Bard, and Microsoft Copilot, with China’s Baidu and Huawei also launching generative AI tools like “Wenyan Yixin” and “Kunpeng.” Overall, the rise of generative AI, represented by large language models, benefits from the combination of three factors: massive data advantages, reinforcement learning algorithm advantages, and continuous super computing power advantages[2], which can be simply summarized as “massive data feeding + deep learning + simulated training + fine-tuning + output processing”[3].

As digital judicial reform gradually advances, generative AI, represented by ChatGPT, with its significant technological innovations in deep synthesis and powerful computing power, has brought new development opportunities and risk challenges to the concepts, structures, and operations of traditional judicial decision-making, creating historic opportunities for adaptive adjustments in judicial practices. On January 30, 2023, a judge in Colombia made a court ruling for the first time using ChatGPT, hailed as the “world’s first AI trial.” The judge stated that ChatGPT can generate easily understandable judgment documents and predict court rulings based on past similar cases, significantly shortening judicial response times[4].

In fact, there have been precedents for the application of AI in the judicial field. In the late 1960s, West Germany began researching the potential application of computers in legal practice, and in 1973, built a judicial data system named JURIS[5]. In China, the initial application of AI in the judicial field occurred in 1986, when Zhu Huarong and Xiao Kaiquan established a mathematical model for sentencing in theft cases. In 1993, Professor Zhao Yangguang developed a computer-assisted sentencing system, which was adopted by several courts and procuratorates due to its functions of retrieving, consulting legal norms, and reasoning judgments on individual cases[6]. In recent years, with the development of AI technology, digital judicial reform has gradually accelerated. In 2015, the Supreme People’s Court first proposed to build a “smart court.” On December 8, 2022, the Supreme People’s Court issued the “Opinions on Regulating and Strengthening the Application of AI in Judicial Practice,” stating: “We must establish a judicial AI technology application and theoretical system with rule guidance and application demonstration effects, providing high-level intelligent auxiliary support for justice and serving the people.” In practice, systems like Beijing High Court’s “Smart Judge,” Shanghai High Court’s “206” system, and Ningbo’s “AI Judge Assistant” have achieved positive results, reducing the clerical workload of judges by more than 30% and improving trial efficiency by over 20%[7].

At the same time, as the application scenarios and scope of AI technology in judicial decision-making continue to expand, while improving judicial efficiency and alleviating the imbalance between litigation cases and judges, it has also raised concerns among scholars about judicial fairness. Professor Allen expressed early concerns about the application of AI in judicial decision-making, believing that it involves considerable uncertainty, ambiguity, and openness[8]. Professor Wang Lusheng pointed out that the application of AI in the judicial field may lead to judicial reform falling into “technocracy,” where the complexity of judicial reform is reduced to simple calculations, and the difficulties of judicial reform are avoided through technical means, ultimately leading to the risk of “goal substitution” in judicial reform[9]. Professor Sun Haibo also believes: “Generative AI can only play an effective auxiliary role in judicial decision-making; exaggerating or even deifying its role is not advisable”[10].

This article argues that, from a developmental perspective, generative AI assisting judicial decision-making is an unavoidable objective fact. Therefore, when discussing its application in the judicial field, it is necessary to thoroughly analyze its risks, find a balance between the technical rationality of AI and the value rationality of judicial decision-making, control the instrumentalist flaws of technical governance, and establish necessary limits and value boundaries for its assistance in judicial decisions, to achieve trustworthy intelligent judicial decision-making.

2. Mechanism Analysis: The Application Prospects of Generative AI in Judicial Decision-Making

According to the statistical data on judicial trial work from the people’s courts, in the first half of 2023, courts nationwide received 16.96 million new cases, an increase of 11.01% year-on-year; and concluded 15.262 million cases, an increase of 9.65% year-on-year[11]. With the advancement of the case registration system and the reform of the judge quota system, the surge in the number of cases and the disproportionate number of judges, along with the misallocation of judicial resources, has resulted in the problem of “more cases than judges,” leading to a situation where the quality and quantity of cases cannot be balanced. Therefore, utilizing generative AI tools to assist in similar case recommendations, evidence identification, and deviation warnings can effectively enhance judicial efficiency, liberate judges from mechanical labor, and alleviate current issues such as inconsistent legal applications and the shortage of judges.

(1) Intelligent Similar Case Recommendations to Achieve Similar Case Treatment

“Statutory law itself implies the risk of generating conflicts in judicial decisions.”[12] The uniform application of law is an important practical constraint and goal state of judicial decisions, as well as the logic behind the consistency and universality of legal validity. However, with the rapid development and dramatic transformation of the social economy, increasingly complex social contradictions and the “amplification effect” of judicial documents through the internet have led to frequent disputes over judicial decisions. Therefore, inconsistent legal applications have become a prominent problem troubling the judiciary in China[13]. “Similar cases should be treated similarly” is “the simplest form of justice.” Allowing the public to feel fairness and justice in every case is the ultimate value goal of uniform legal application, as well as the constitutional provision that “everyone is equal before the law.” Utilizing generative AI technology to assist in achieving similar case treatment and ensuring the consistency of sentencing discretion, as well as realizing the universalization of similar case judgment standards, is one of the directions of current digital judicial reform.

The technical path of generative AI assisting intelligent similar case recommendations to achieve similar case treatment is to access key information, case facts, etc., and then the system automatically matches and retrieves from the case topic database, intelligently tagging similar cases for comparison, thereby achieving precise similar case recommendations; and using massive similar case judgment data as a coordinate system, mimicking judges to predict decisions on the cases being decided, providing relatively standardized and consistent sentencing recommendations to assist judges in sentencing. Intelligent similar case recommendations can not only liberate judges from repetitive mechanical labor, saving judges’ efforts during case retrieval but also improve trial efficiency and alleviate the “more cases than judges” problem, while limiting judges’ discretionary power during the judicial decision-making process, avoiding inconsistent judgments in similar cases, judicial corruption, and judicial injustice.

(2) Intelligent Deviation Warning Mechanism to Reduce Judges’ Professional Risks

The “Decisions on Several Major Issues Concerning the Promotion of the Rule of Law in China” by the Central Committee of the Communist Party of China points out that the implementation of judicial accountability requires that “those who adjudicate should decide, and those who decide should be accountable.” Currently, China implements a lifelong accountability system for wrongful judgments, which must ensure both quality and quantity, imposing tremendous psychological pressure on judges. The emergence of intelligent judicial decision-making technology can provide judges with visual references for similar case rulings, issuing timely warnings for judicial decisions that deviate significantly from “median standards” or “average values,” allowing judges to correct their judicial decisions, alleviating their trial pressure, and reducing their professional risks.

(3) Intelligent Evidence Review to Assist Evidence Identification

The neutrality, objectivity, clarity, and empirical nature (big data analysis) of generative AI technology can, to some extent, compensate for or even replace the subjective and relatively uncertain judgments of experiential rules under the criminal proof model, reducing the evidentiary pressure on judges[14]. As a scientific evidence, generative AI’s review of evidentiary capacity and probative force is mainly influenced by the three attributes of authenticity, relevance, and legality, as well as the relevant rules of evidence stipulated in the Criminal Procedure Law of the People’s Republic of China. Currently, the intelligent review of evidence by generative AI goes through four stages: Stage One, collecting, storing, retrieving, mining, analyzing, and aggregating information on case facts, personnel, and evidence, establishing corresponding business topic databases, including case file databases, evidence standard databases, and case evidence model databases; Stage Two, using generative AI’s deep machine learning and powerful computing power for scene text recognition, manual marking of evidence flaws, and other analysis and verification; Stage Three, generating several verification rules for evidentiary documents such as litigation documents and evidence materials, constructing an evidence rule verification engine through rule model generation, key element analysis, and logical reasoning, achieving automatic verification of the aforementioned rules, analyzing and inferring the authenticity, legality, and relevance of evidence; Stage Four, intelligently constructing a complete evidence chain and knowledge graph of the case that is visualized, systematic, and data-driven, assisting judges in evidence identification.

The advantage of intelligent evidence review lies in constructing the cognitive process of judicial proof based on the mathematical language of scientific evidence, expressing the probabilistic relationships in mathematical terms and the subjective belief levels regarding the facts of the case as a set of numbers; by unifying the standards for evidence identification, assisting judges to accurately and efficiently utilize evidence to gain confidence in the facts of the case, overcoming inappropriate thinking such as “truth implies legality,” “confirmation implies legality,” and “stability implies legality.” However, due to insufficient algorithmic interpretability, strictly pursuing formal justice through the evidence standards of generative AI may also fall into the metaphysical trap of the “statutory evidence system.” Therefore, this cognitive path has its shortcomings, requiring judges to play the role of “gatekeepers” when conducting substantial reviews of the evidentiary capacity of AI evidence.

3. Rational Reflection: Risk Assessment of Generative AI in Judicial Decision-Making

When discussing the technological transformation brought by generative AI to the field of judicial decision-making, it is essential to address the “means.” The evaluation of facts and prediction of decisions by generative AI’s computer models is a technical activity, and the technical risks such as algorithmic black boxes, algorithmic discrimination, and algorithmic bias are caused by the improper use of technology; it is also essential to address the “means.” It is precisely because of the infinite creativity and exponential evolutionary possibilities of generative AI that we witness the power of algorithms while easily overlooking the essence of judicial decisions, thus falling into the crisis of Kant’s “Critique of Pure Reason.”

(1) Algorithmic Black Boxes Raise Doubts About Judicial Credibility

“Justice cannot hide; it cannot be more easily concealed where it is shown.”[15] Judicial decisions involve judges resolving disputes between parties through legal reasoning and interpretation, thus restoring social relationships and maintaining the stability of law. From this perspective, the function of judicial decisions is to reach solutions regarding specific justice through reasoning or argumentation, achieving resolution of disputes[10]. Inflation can lead to devaluation, and the same applies to judicial credibility; judicial decisions that cannot be fully understood will not be fully complied with, resulting in significantly increased social instability. Therefore, judges need to reasonably explain the effectiveness and reliability of the technology, processes, and results of generative AI-assisted judicial decisions. However, constrained by the complexity, technicality, and proprietary ownership of algorithms, there are secrets in the “input-output” decision-making process of algorithms that we cannot glimpse, thus intelligent adjudication faces risks of transparency and interpretability in judicial decision-making. For example, in the 2016 case of “Loomis v. Wisconsin”[16], the defendant Eric Loomis defended that the COMPAS sentencing algorithm violated his due process rights, but the algorithm’s parent company, Northpointe, successfully refused to disclose how the algorithm’s scoring calculations were made, citing proprietary protection as a commercial secret. This landmark case has sparked thoughts on how algorithmic black boxes affect the transparency of judicial decisions.

The existence of algorithmic black boxes not only fails to guarantee the right to know and defend of parties and defense lawyers but also prevents judges from adequately weighing and explaining the fairness and accuracy of their decisions, making the legitimacy, effectiveness, and reliability of generative AI-assisted judicial decisions more challenging, thus raising doubts about judicial credibility.

(2) Algorithmic Discrimination and Bias Trigger a Crisis of Judicial Fairness

Big data and deep machine learning algorithms provide the feasibility of objectivity, neutrality, and empirical references or decision-making bases for generative AI technology in assisting judicial decisions. However, over time, critical scholars’ research has proven that machine learning algorithms reproduce forms of “algorithmic oppression”[17]. Currently, the algorithmic discrimination and bias of generative AI stem from two levels. One is acquired algorithmic bias, where the pre-training data input contains biases or discrimination that trigger algorithmic discrimination and bias. For example, when calculating and predicting the sentencing of criminals in criminal cases, the algorithm considers statutory circumstances and discretionary circumstances, and when applying probation, it often overlooks that the “exposure risk” of criminals in society is often the result of excessive discrimination and disadvantage. This data is often related to risk variables, leading to “higher” risk scores, especially for marginalized groups in society. The second is inherent algorithmic bias, where the algorithm itself possesses discrimination and bias. Currently, intelligent adjudication tools are usually outsourced to other companies by the courts; if companies intentionally embed biases and discrimination in the underlying logic of algorithms through technical means, judges may not easily detect these errors, which will undoubtedly further trigger a crisis of judicial fairness. Such algorithmic discrimination and bias may not matter much in other areas of life, but for judicial decisions that substantively affect the rights and obligations of parties, such discrimination and bias are extremely harmful; they objectively enhance the likelihood of “intelligent adjudication” errors, leading to a more unfavorable situation for parties, and further creating inequalities in litigation status between legal subjects[18].

(3) Intelligent Judicial Decisions Cannot Balance Substantive Justice

In the context of law and justice, the concept of justice has transcended the mathematical calculations or definitions typically used in computer science and statistics[19]. In fact, justice is not only an objective recognition but also a subjective feeling. Simply declaring justice in judicial decisions is insufficient; it is also necessary for the public to feel justice. How to achieve substantive justice in individual cases? This cannot be separated from the application and understanding of value judgments and legal principles. In fact, the application of law involves not only factual judgment processes but also value judgment processes, encompassing both the application of legal rules and the application of legal principles. Although to ensure the stability of law, specific legal rules should be prioritized over abstract legal principles, when applying specific legal rules leads to injustice in individual cases, it is necessary to prioritize the application of legal principles to achieve individual justice. For example, in the U.S. case of “Riggs v. Palmer”[20], the judge invoked the legal principle that “no one should benefit from their wrongdoing” to deprive Palmer of his inheritance rights. In recent years, there have been frequent controversial cases in China’s judicial practice, such as the Shenzhen Parrot case, Xu Ting case, Zhao Chunhua case, and Wang Lijun case. The reason these cases’ outcomes are difficult for the public to recognize and accept, resulting in less than ideal judicial and social effects, is largely due to judges exercising too little value judgment in judicial decisions, while excessively mechanically applying legal rules, neglecting the application of legal principles, leading to structural imbalance, namely legalism. Unfortunately, in current judicial practice, the application of legal rules remains the primary mode of judicial decision-making. Therefore, at this stage, intelligent judicial decisions still cannot genuinely balance substantive justice, as reflected in the following aspects.

Firstly, from the perspective of legal texts, compared to legal rules that logically possess complete hypothetical conditions, behavioral patterns, and legal consequences, legal principles, which logically lack clear hypothetical conditions, behavioral patterns, and legal consequences, are more challenging for generative AI to understand and apply. Additionally, due to the diversity of human language, the meanings and usages of legislative language can vary significantly in different contexts, resulting in some ambiguity and uncertainty in the semantics of legislative language. Overall, the language used in legislation inevitably has a “semantic space”[21]. This easily leads to generative AI misunderstanding the legislative purposes and intentions behind legal norms.

Secondly, regarding value judgments, generative AI is incapable of understanding and making value judgments. As Heidegger pointed out: “The essence of truth and the interpretation of existence are determined by the human being as the true subject.”[22] As a legal system that regulates human behavior, law must exist in human nature. However, for generative AI, “law is a string of arrangements of ‘0’ and ‘1,’ a set of axioms, and so-called legal reasoning is a purely mathematical reasoning formula”[24]. From this perspective, generative AI cannot perceive or understand; its role is limited to factual judgments, merely distinguishing between “true” and “false,” without involving value judgments of “good” and “beautiful.” Therefore, when confronted with value judgments, generative AI either chooses to evade or transforms them into factual judgments. However, as Kant pointed out, one cannot derive the ought from the is. In other words, one cannot obtain a basis for what is substantively just or unjust from a single fact. In summary, while intelligent judicial decisions can pursue formal justice, excessive reliance on generative AI to achieve similar case treatment inevitably leads back to the old path of “mechanical jurisprudence,” moving toward the opposite of substantive justice, severely deviating from the goals of justice and the rule of law.

Thirdly, regarding the identification of case facts, generative AI struggles to understand new, difficult, and complex cases. Undeniably, generative AI can play a significant role in simple cases, such as theft, fraud, and private lending. However, with the rapid development of the social economy, various new, difficult, and complex cases are emerging. In the face of these new and complex cases, the law itself faces issues of legislative gaps, incomplete content, and ambiguous provisions, leading to inevitable conflicts between norms. For instance, the case of Qiu Mouming infringing on the reputation and honor of heroic martyrs. Both factors lead to interruptions in legal guidance in individual cases, and in the face of new, difficult, and complex cases, this guidance can be exhausted[10]. “The richness and diversity of life do not allow the law to provide abstract, clear, and appropriately structured regulations for all individual situations.”[26] This “open structure” is a defect from the perspective of the stability of law (i.e., the computability of law), but from the perspective of the adaptability of law (Schmiegsamkeit), it is an advantage. It gives general legislative sentences a certain adaptability to the structured living situations. However, this “open structure” is fatal for AI. Jack Rae, a core member of Open AI’s large language model, pointed out in “Compression for AGI” that the “intelligence” of generative AI is a form of compression (i.e., simplification)[27]. In the face of new, difficult, and complex cases with no applicable legal norms, the computational process of generative AI involves two necessary steps: first, compressing and extracting case elements based on language “statistical probabilities” and converting them into corresponding numerical codes; second, forcibly matching relatively applicable legal norms based on probability from the same language model to predict the probability distribution of judicial outcomes. This “forced” correction, compression, and probability judgment inevitably distort facts, mechanically deriving unreasonable judicial outcomes. Essentially, this is a form of “random justice,” akin to flipping a coin or rolling dice. In other words, for generative AI, the correctness of the output is not important; what matters is maintaining statistical consistency with the prompts. Kelsen foresaw that specific norms could not be logically or mathematically deduced from fundamental norms[29]. Postman also critically observed the phenomenon of statistical abuse[30].

Fourthly, regarding the judges’ discretionary power, generative AI inevitably encroaches upon judges’ discretionary space. Judicial decisions require judges to navigate between facts and norms, correctly addressing the correspondence between what exists and what ought to be, seeking specific, real laws, which relies on judges’ discretionary power. As previously mentioned, the core of intelligent adjudication is to present a quantitative analysis of all data, thereby unifying judicial discretion standards, suppressing power rent-seeking and judicial corruption, as well as overcoming the arbitrariness of discretion. While this formally meets the universal requirement of “equality before the law,” it inevitably leads to judicial rigidity, a lack of innovation, and substantive inequality, falling into the trap of “vending machine judges.” As Pound pointed out, “No legal system can achieve justice solely through rules without relying on discretion; no matter how meticulous or specific the system of rules is, all processes of implementing justice involve both discretion and rules”[32]. Furthermore, when faced with scenarios lacking applicable norms, under the “principle of prohibition against refusing to adjudicate,” judges’ discretion can also carry out “legal reconstruction,” filling legislative gaps and addressing the shortcomings of outdated laws. However, the “hallucinations” of generative AI make it fundamentally incapable of undertaking such significant responsibilities. Additionally, judicial decisions must not only focus on quantifiable information such as case facts and legal norms but also transcend norms and facts. Considering non-quantifiable information such as moral ethics, good customs, and conventions, all of which rely on judges’ discretionary power, is an unattainable aspiration for AI.

(4) Algorithmic Power’s Impact on Judicial Accountability

There is an inevitable causal relationship between risk and responsibility. When society moves towards an artificially induced uncertain risk and the unknown realm governed by interest games, discussions about responsibility inevitably arise[33]. In the irreconcilable contradiction between the reality of risk and governance rules, artificial intelligence becomes an object for judicial personnel to shift and share responsibility, facilitating the abdication of judicial responsibility.

In the overall framework of judicial accountability, judges exercising judicial decision-making power independently according to the law is the premise and foundation for ensuring accountability. However, as artificial intelligence is increasingly used in judicial practice, under the pressures of litigation explosions, year-end assessments, and rigid regulations of trial deadlines, there will inevitably be a tendency for judges to overly rely on generative AI. Once judges develop a technological reliance in judicial decision-making, the risks of “algorithmic power” become unavoidable, leading to the collapse of the power configuration and operational mechanism of judicial independence. Moreover, generative AI-assisted judicial decisions will inevitably lead to the alienation of the qualifications of trial subjects and the collapse of trial structures, resulting in a situation where the algorithm’s designers and judges jointly make judicial decisions. When the adjudicative power is transferred to others or objects, and the trial subjects cannot be determined, the distribution of the responsibility chain will create accountability challenges, significantly increasing the likelihood of “agentic shift.”

4. Towards Trustworthiness: Risk Mitigation of Generative AI in Judicial Decision-Making

While intelligent adjudication improves judicial efficiency, achieves similar case treatment, and enhances predictability of judgments, it inevitably raises public doubts about judicial credibility and fairness. To mitigate the technical risks of generative AI tools, such as algorithmic black boxes, discrimination, and bias, governance can be approached from three dimensions: data governance, algorithm governance, and control governance.

(1) Data Governance

A massive, accurate, consistent, and standardized judicial dataset is the “treasure” for pre-training and fine-tuning generative AI, and it is also a prerequisite for generative AI to assist judicial decision-making more intelligently. Existing research has proven that the generative content of foundational models of generative AI is related to training data[34].

1. Open Source Increment: Expanding Sources of Judicial Data

Currently, databases such as the China Judgments Online, Peking University Law Database, and Westlaw provide comprehensive and systematic judicial data support for deep learning training of generative AI. However, it still falls significantly short of the requirements for comprehensive data availability. Some scholars have pointed out that the current public access to judicial documents in China is “pseudo-public”[35]. For example, cases involving personal privacy, mediation settlements, and divorce litigation cannot be fully published online. In addition to the issue of “pseudo-publicity,” the number of publicly available judicial documents has also shown a sharp decline, with only about half of the concluded cases being published[36]. According to my statistics, regarding civil cases alone, the number of first-instance civil judgment documents published in 2022 was 4,388,998, a decrease of 46.18% compared to 8,155,478 in 2021. Therefore, there is an urgent need to comprehensively and timely increase the volume of publicly available judgment documents and expand the sources of judicial data. First, for specific judgment documents that are not suitable for public disclosure, anonymization measures can be taken to publish them after removing or deleting relevant information. Second, relevant implementation rules for public disclosure of judgment documents should be introduced, detailing the differences between judgments, rulings, and decisions, listing specific categories, situations, and standards to enhance the operability of publishing judgment documents online. Third, in terms of judicial data circulation, a mechanism for timely sharing of case data information among courts of different regions and levels should be improved, allowing generative AI to respond to and process data information in real-time.

2. Quality Improvement: Enhancing Judicial Data Quality

Ensuring data quality is a requirement of justice; safeguarding justice can prevent extreme polarization caused by artificial intelligence[37]. The standardization and formatting of judgment documents, as well as comprehensive reasoning, can help generative AI tools analyze, learn, and understand judicial behavior, ensuring that intelligent adjudication results are more precise, fair, and just. The quality of judicial data can be improved through both formal and substantive approaches. First, by unifying the expression format of judgment document content, generative AI’s misunderstandings can be avoided. By continuously enhancing the formatting and standardization of judgment documents, the incidence of omissions and errors can be reduced from the source. For example, errors in words, punctuation, expressions, and grammar. Second, specific requirements for the elaboration of judgment document reasoning should be clarified, deepening generative AI’s understanding of judicial patterns. Judgment documents must be based on facts and adhere to legal standards. For specific subjects and issues, clarity of thought and distinct levels should be maintained, aligning with the objective facts of the case, thoroughly justifying and illustrating the legitimacy and legality of the process of determining legal norms and case evidence, strengthening the legal, logical nature of judgment document reasoning, achieving consistency among reasons, facts, and judgment results.

(2) Algorithm Governance

1. Meta-Regulation: Embedding Algorithm Ethics

“Ethics first is the primary requirement for applying intelligent algorithms in social life, and it is also an important guarantee for the healthy development of science and technology.”[38] Embedding ethical norms in algorithm design helps to determine whether algorithmic actions meet ethical standards, thus categorizing algorithmic risks such as discrimination and bias as prohibited behaviors for correction and regulation, achieving “meta-regulation.” Specific requirements should encompass several levels. First, at the development level, although algorithms are technically value-neutral, the values of the algorithm designers are inherently mixed from the outset, possessing clear subjective attributes. Therefore, there is a need to strengthen constraint mechanisms on activities related to AI development, enhancing the self-management and self-restraint awareness of developers, regulating them against engaging in algorithm development that violates ethical standards. Second, at the theoretical level, regulatory agencies or industry associations should introduce algorithmic ethical standards that are forward-looking, comprehensive, and flexible, revealing the actual interests that may be infringed to address the ever-evolving algorithmic risks, compensating for the shortcomings of lagging algorithmic ethical standards and strengthening practical applications to guide algorithms toward positive outcomes. Third, at the institutional level, specific systems such as algorithm registration, algorithm safety assessments, algorithm accountability, algorithm governance, and third-party supervision should be established to clarify the boundaries of obligations and responsibilities for algorithm designers, achieving comprehensive scrutiny of algorithmic ethics from within to outside, thereby preventing algorithm abuse.

2. Core Optimization: From Algorithmic Black Boxes to Algorithmic Interpretability

Some scholars have pointed out: “Compared to the transparency of algorithms, the interpretability of algorithms is the goal.”[19] It is understandable that while algorithmic transparency is important, considering that parties involved in litigation face complex codes akin to “heavenly books,” the effects of algorithmic transparency are minimal and cannot dispel public doubts about the credibility of intelligent adjudication. However, through appropriate algorithmic explanation methods, it is at least possible to leave traces of information in most intelligent judicial decision-making processes, reproducing the processes, results, and applications of intelligent judicial decisions to the greatest extent, thus unveiling the veil of discrimination and bias within algorithms, achieving “de-blackboxing” of algorithms. Therefore, enhancing algorithmic interpretability is key to promoting trustworthy intelligent adjudication and gaining public “algorithm trust.” Specific explanation methods include the following.

First, localized explanations targeting specific questions. Generally, dissent regarding judicial decisions often focuses on specific points, and it is rare to completely deny the entire ruling. Therefore, compared to comprehensive global explanations detailing the complex logic and systematic working mechanisms behind the entire algorithm, localized explanations targeting specific issues are more intuitive, concise, and labor-saving, enhancing public understanding. The basic idea is to generate an explanation alongside the predicted results of the generative model. Common localized explanation methods include sensitivity analysis explanations, local approximate explanations, gradient direction propagation explanations, and activation mapping explanations, etc.[39].

Second, differentiated explanations for different audiences. Professional knowledge, public expectations, and work backgrounds can all affect understanding and trust in algorithms, so emphasizing different explanation methods for different subjects is a necessary choice to enhance the effectiveness of algorithmic explanations[40]. For example, for parties lacking relevant technical knowledge, model distillation explanation methods and visual image explanation methods can be employed, which, while technically incomplete or inadequate, may actually yield better results.

(3) Control Governance

Currently, both academia and practice have reached a consensus: judicial efficiency must never come at the cost of judicial fairness. No matter how artificial intelligence develops, it can never replace judges. In this contest for judicial decision-making power between machines and judges, humanity must ultimately prevail. Therefore, to reasonably define the boundaries and limits of generative AI tools in assisting judicial decision-making, ensuring the healthy development of intelligent adjudication should always start with highlighting and reinforcing judges’ primary status, coordinating the layered application of cases based on their complexity, and strictly implementing the judicial accountability system.

1. Human-Centered: Highlighting and Reinforcing Judges’ Primary Status

The essence of judicial activities is the concrete application of human practical rationality. From this perspective, judicial decisions cannot and must not become purely technical; excessive blind adherence, reliance, and unrestrained use of AI technology will inevitably undermine judges’ status and further compress the space for legal discourse. In fact, judicial decisions, based on legal principles, ethical considerations, and careful insights into case facts, formed through judges’ free conviction and “legal sense,” are ultimately a profound “art,” comparable to writing, painting, and performing arts. Even though generative AI possesses powerful deep learning capabilities, it cannot fully replace judges in making judicial decisions that simultaneously consider legal, social, and political effects; moreover, the original intention of digital judicial reform is not to let generative AI independently adjudicate but to play an auxiliary role in judicial decision-making, assisting judges in lawfully, comprehensively, and normatively collecting and reviewing evidence, unifying legal application standards, ensuring judicial fairness, and achieving equality before the law. Particularly at this time, generative AI still faces many technical challenges, and judicial reform is steadily advancing. Therefore, it is crucial to always highlight and reinforce judges’ primary status, grasp the scale of intelligent adjudication, and establish judges’ ultimate discourse power over judicial decisions.

2. Procedural Diversification: Coordinating Layered Applications Based on Case Complexity

It is evident that generative AI-assisted judicial decisions demonstrate significant efficiency advantages in simple cases with clear legal relationships and facts. However, as previously mentioned, it encounters difficulties in recognizing and understanding complex cases. Therefore, under the current mechanism of case complexity differentiation, applying AI-assisted judicial decisions in a layered manner based on the nature of simple and complex cases can better balance the relationship between judicial fairness and efficiency. First, emphasizing efficiency requirements for simple cases. Simple cases typically involve relatively straightforward legal relationships and clear facts, and the litigation procedures are generally expedited, making it possible for generative AI to leverage its exponential algorithmic processing capabilities and response speed to assist judges in element-based adjudication, including evidence review, case analysis, and predicting judgments, achieving rapid trial processes while minimizing errors. Second, reducing standardized control in complex cases. The legal relationships, case facts, and related evidence materials in complex cases are more intricate compared to simple cases, making it difficult for generative AI to accurately identify similar cases, leading to challenges in precise recommendations and predicting judgments. Therefore, in complex cases, the standardized control of generative AI in judicial decisions should be weakened, and intelligent applications should be used moderately.

3. Unveiling Artificial Intelligence: Implementing Judicial Accountability

Weber noted: “A person or many people can implement their will in a collective action even when faced with resistance from other participants.”[42] At this stage, generative AI does not possess the realistic foundation for accountability and cannot achieve legal “personification.” From the perspective of trial status, generative AI cannot challenge judges’ primary status; its use or non-use, acceptance or rejection, entirely depends on judges’ will and choice, with judges still controlling the entire judicial process. Therefore, it is necessary to unveil artificial intelligence and strictly implement the judicial accountability system, “let those who adjudicate decide and those who decide be accountable.” Moreover, strictly implementing the judicial accountability system will also encourage judges to rely on AI more cautiously, “thoughtfully obeying” AI, enhancing judges’ self-examination and review of judgment results, and increasing their sense of responsibility for ensuring judicial fairness.

After clarifying the need to strictly implement the judicial accountability system, it is also necessary to structurally analyze the states and causes of damage caused by intelligent adjudication, further delineating the boundaries between judicial responsibility and judicial immunity, and clarifying the scope of judges’ responsibilities, constructing a mechanism for sharing responsibilities that aligns with rights and duties. If a decision error is caused by logical errors in generative AI’s calculations, judges will not bear responsibility if they have fulfilled their reasonable duty of review; if a decision error is caused by judges’ own negligence or dereliction of duty leading to generative AI’s erroneous decisions, judges will bear primary responsibility; if the decision error is caused by the designer of generative AI, then the designer will bear primary responsibility, and judges will not bear responsibility.

5. Conclusion

From a development perspective, with the arrival of the “singularity” of generative AI, the integration of generative AI and the field of judicial decision-making is an objective fact that cannot be avoided, but we should always maintain a “technological humility” attitude, having a global awareness, neither making definitive judgments nor hindering the specific functions of technology. On one hand, generative AI-assisted judicial decision-making can indeed visualize formal justice and enhance judicial efficiency through similar case recommendations, auxiliary evidence identification, and deviation warnings, which has positive aspects; on the other hand, generative AI still faces many technical risks that need to be addressed, and its algorithmic black boxes, discrimination, and bias are still unavoidable objective realities, and its “alluring justice” inevitably leads to judges overly relying on generative AI, resulting in humans being “enslaved” by machines.

As Weizberg pointed out: “Human intelligence cannot be transplanted.”[43] Reflecting on the judicial decision-making career, which is based on human emotions, rationality, and experience, it is essential that in the open field of judicial decision-making, generative AI should not be allowed to operate independently; the primary status of judges in the field of judicial decision-making must never be shaken, and careful responses to the co-construction of technological power and judicial power must be ensured, avoiding the emergence of situations where “humans are in the loop.”

References:

[1] Xinjing News. Data Analysis | Over 100 million active users within 2 months of launch; What is ChatGPT? [EB/OL]. (2023-02-13) [2023-09-15]. https://baijiahao.baidu.com/s?id=1757705586200707936&wfr=spider&for=pc.

[2] Zhang Xinxin, Ding Jingjia. The Technical Principles and Process Innovations of Generative Intelligent Publishing [J]. Library and Information Knowledge, 2023, 40(5): 68-76.

[3] Long Keyu. Research on the Legal Regulation of Generative AI Application Anomalies: Perspectives from ChatGPT and Social Robots [J]. Eastern Law, 2023(4): 44-45.

[4] JANUS ROSE. A Judge Just Used ChatGPT to Make a Court Decision [EB/OL]. (2023-02-04) [2023-09-16]. https://www.vice.com/en/article/k7bdmv/judge-used-chatgpt-to-make-court-decision.

[5] Gong Xiangrui, Li Keqiang. The Computerization of Legal Work [J]. Law Science Magazine, 1983(3): 16-20.

[6] Democracy and Law Network. The Dream of “Computer Sentencing”: Professor Zhao Tingguang and His Sentencing Assistant System [EB/OL]. (2007-12-22) [2023-09-15]. http://www.mzyfz.com/news/mag/a/20071222/110234.shtml.

[7] Tencent News. The Supreme Court: Intelligent trial assistance systems reduce judges’ clerical work by over 30% [EB/OL]. (2022-10-13) [2023-09-17]. https://new.qq.com/rain/a/20221013A03SLD00.

[8] Ronald J. Allen, Wang Zhuhai. Artificial Intelligence and the Judicial Proof Process: Challenges from Formalism and Computation [J]. Evidence Science, 2020, 28(5): 588-599.

[9] Wang Lusheng. The Risks and Ethical Regulations of Judicial Big Data and AI Technology Applications [J]. Law and Business Research, 2019, 36(2): 101-112.

[10] Sun Haibo. Reflecting on the Possibilities and Limits of Intelligent Adjudication [J]. Journal of the National Prosecutors College, 2020, 28(5): 80-99.

[11] Supreme People’s Court of the People’s Republic of China. Normalization! The Supreme Court publishes major judicial trial data quarterly [EB/OL]. (2023-08-07) [2023-09-20]. https://www.court.gov.cn/zixun/xiangqing/408422.html.

[12] Chen Hangping. On the Generation and Identification of “Different Judgments for Similar Cases” [J]. Contemporary Law, 2012, 26(5): 26-31.

[13] Liu Lei. Achieving “Similar Treatment for Similar Cases” through Similar Case Comparison: A Perspective from Judicial Field Analysis [J]. Local Legislative Research, 2022, 7(2): 95-111.

[14] Long Zongzhi. Issues with the Application of Experiential Rules in Criminal Proof [J]. Chinese Criminal Law Magazine, 2021(5): 55-70.

[15] Zhang Anbang. Shaping the Authority of Judicial Power in the Context of Judicial Reform [J]. Zhejiang Social Sciences, 2021(9): 67-75,157.

[16] Zhu Tizheng. The Uncertainty Risks of AI-Assisted Criminal Judgment and Its Prevention: Insights from the Loomis Case in Wisconsin, USA [J]. Zhejiang Social Sciences, 2018(6): 76-85,157.

[17] Mark Schullenberg, Rick Peters. The Algorithmic Society: Technology, Power, and Knowledge [M]. Wang Yanchuan, Li Pengfei, Trans. Beijing: Commercial Press, 2023: 142.

[18] Liu Dongliang. Technical Justice: The Dual Variations of Procedural Law and Algorithms in the Age of AI [J]. Comparative Law Research, 2020(5): 64-79.

[19] Jiang Tao. Value Judgment in Criminal Law Interpretation [J]. Chinese Social Sciences, 2023(7): 38-160,207.

[20] Lin Laifan, Zhang Zhuoming. On the Judicial Application of Legal Principles: An Analysis from the Perspective of Normative Legal Methodology [J]. Chinese Law, 2006(2): 122-132.

[21] TIMOTHY A.O. ENDICOTT. Vagueness in Law [M]. Oxford: Oxford University Press, 2000: 1-2.

[22] Heidegger. Nietzsche: Volume II [M]. Sun Zhouxing, Trans. Beijing: Commercial Press, 2002: 824.

[23] Li Wei. The Human Basis in the Value Orientation of Family Law [J]. Law Science Magazine, 2017, 38(9): 63-72.

[24] ANNE VON DER LIETH GARDNER. An Artificial Intelligence Approach to Legal Reasoning [M]. Boston: The MIT Press, 1987: 19.

[25] Martin Heidegger. Speeches and Papers [M]. Sun Zhouxing, Trans. Beijing: Commercial Press, 2018: 7.

[26] Reinhard Chipelius. Philosophy of Law [M]. Jin Zhenbao, Trans. Beijing: Peking University Press, 2013: 15.

[27] JACK RAE (OpenAI). Compression for AGI [EB/OL]. (2023-04-22) [2023-08-05]. https://www.bilibili.com/video/BV1js4y1D7gz/.

[28] Arnaud R. Lord. Dialogues on Law: A Dialogue Model of Legal Constitution and Argumentation [M]. Wei Bin, Trans. Beijing: China University of Political Science and Law Press, 2016: 4.

[29] Ji Weidong. The Territory of Computational Jurisprudence [J]. Journal of Social Sciences, 2021(3): 116.

[30] Neil Postman. Technopoly: The Surrender of Culture to Technology [M]. He Daokuan, Trans. Beijing: CITIC Press, 2019: 129.

[31] Wang Tiansi. Causal Relationships in Big Data and Their Philosophical Implications [J]. Chinese Social Sciences, 2016(5): 22-42,204-205.

[32] POUND R. Jurisprudence [M]. Cambridge: Harvard University Press, 1959: 335.

[33] Wang Yaqi. Analyzing the Uncertainties of Global Governance from the Perspective of Risk Society [J]. Northeast Asia Forum, 2022, 31(3): 44-59,127.

[34] Chen Changfeng, Zhang Meng. Determined by Data? The Value and Ethical Issues of AIGC [J]. Journalism and Writing, 2023(4): 15-23.

[35] Lü Shengyun, Shui Juan. Preventing the “Pseudo-Publicity” Phenomenon of Judicial Documents Online [J]. Online Legal Review, 2013, 17(2): 97-113.

[36] Ma Chao, Yu Xiaohong, He Haibo. Big Data Analysis: A Report on the Public Disclosure of Judicial Judgment Documents in China [J]. Chinese Legal Review, 2016(4): 195-246.

[37] Li Xiaohui. How to Face the “Useless Class”: Artificial Intelligence and Future Justice [J]. Journal of East China University of Political Science and Law, 2019, 22(6): 68-77.

[38] Li Lin. Improving Strategies for Algorithmic Ethical Review Approaches [J]. Academic Exchange, 2023(4): 74-86.

[39] Ji Shouling, Li Jinfeng, Du Tianyu, et al. A Review of Machine Learning Model Interpretability Methods, Applications, and Safety Research [J]. Computer Research and Development, 2019, 56(10): 2071-2096.

[40] Zhou Xiang. Algorithm Interpretability: A Normative Research Value of a Technical Concept [J]. Comparative Law Research, 2023(3): 188-200.

[41] Shuai Yinan. The Realistic Possibilities and Necessary Limits of AI-Assisted Judicial Decision-Making [J]. Journal of Shandong University (Philosophy and Social Sciences Edition), 2020(4): 101-110.

[42] Max Weber. Economy and Society: Volume I [M]. Lin Rongyuan, Trans. Beijing: Commercial Press, 1997: 246.

[43] Rosemary Luckin, Li Haoyang. The Future of Intelligent Learning [M]. Xu Yehua, Trans. Hangzhou: Zhejiang Education Press, 2020: 93.

(Editor: LI Hong)