An Algorithm-centric Governance Model for Artificial Intelligence Judicial Security Risks

*Author   Yanhong Liu

Dean of the School of Criminal Justice, China University of Political Science and Law

Professor of Qian Duansheng Lecture at China University of Political Science and Law

Abstract: The essence of Al in the judicial fields is algorithmic, the retreat of judges is a practical feature of Al in the judicial fields. Algorithmic justice not only impacts on justice, but also fits in with the formal rationality value inherent in traditional judicial justice. Based on this, the safety risks of Al in the judicial fields stem from procedural risks, substantive risks, and inherent technological risks under the guidance of algorithmic justice. Therefore, an algorithm-centric governance model is the ideal and core approach for managing safety risks of Al in the judicial fields. For the future practice of Al in the judicial fields, governance can be structured around three dimensions: algorithmic training data, algorithmic computing models, and the results produced by the algorithms. This can be supported by specific measures such as data security and protection, the fine-tuning and optimization of models, and result correction and validation, thus constructing a three-dimensional governance model of “data-models-results”.

Key words: formal rationality; algorithm in the judicial fields; Al in the judicial fields; risk governance; generative Al; algorithm-centric

The technological revolution in the era of artificial intelligence has triggered the third industrial revolution, profoundly changing the allocation of production factors and the social structure order, and putting forward urgent demands for the modernization and transformation of the national governance system and governance capacity. Since the 18th National Congress of the Communist Party of China (CPC), the construction of informatization and digital intelligence in the judicial field has achieved remarkable results. Through the technological empowerment of modern technical means such as big data, cloud computing, blockchain, and artificial intelligence, the wheel of smart justice has rolled forward, which not only brings about changes in judicial procedures and the way of litigation participation, but also has a deeper impact on judicial logic and the concept of the rule of law. In general, the application scenarios of artificial intelligence in judicial affairs mainly include litigation services, online litigation, business assistance, decision-making assistance, judicial management, etc., showing a trend of gradually strengthening the breadth and depth of application, and the weight ratio of artificial to intelligent in the entire judicial process is gradually reduced. For example, the initial construction results of the judicial system informatization are centered on the digitalization of specific behaviors such as case filing and payment, case file review, document delivery, signature and seal delivery, etc., which is essentially the use of Internet technology to get rid of the constraints of physical space, especially paper media. In this process, although litigation convenience and judicial efficiency have been enhanced, the degree of intelligence is generally low, and there is no substantial impact on litigation rules and judicial models. The inefficiency of China's existing judicial model needs to be improved urgently, and there are also deficiencies in eliminating judicial corruption and judicial abuse of power. Therefore, the need to enhance artificial intelligence to improve judicial efficiency is imminent. In view of this, in the era of smart justice 4.0, the actual application scenarios of artificial intelligence have expanded to the entire judicial process and scenarios such as investigation, prosecution, trial, and execution, especially the deployment of smart push of similar cases, sentencing prediction, deviation warning, document generation, and automatic control systems, which indicates that artificial intelligence technology has been deeply involved in the core areas of judicial operation such as procedural control, substantive decision-making, and supervision and inspection.

Overall, driven by the smart judicial brain and judicial data middle platform, analytical and generative artificial intelligence technologies have been widely used. The new era of artificial intelligence justice has made digital courts, cloud courts, and robot judges a reality. The assertions that “code is law” and “algorithm is justice” are no longer fantasy, but a picture of smart rule of law that will inevitably appear within a foreseeable time. President Xi Jinping pointed out that security is the foundation of development, and we must better coordinate development and security, and use a new security pattern to safeguard the new development pattern. With the development of artificial intelligence judicial technology equipment and the deepening of application scenarios, the corresponding artificial intelligence judicial security risks will also become a realistic problem that judicial practice must face and respond to. The Supreme People's Court issued the “The Opinions on Regulating and Strengthening the Applications of Artificial Intelligence in the Judicial Fields” (hereinafter referred to as the “Artificial Intelligence Judicial Opinions”) in 2022, taking the principle of safety and legality as the primary basic principle, and stipulating that “Pursuing a holistic approach to national security, forbidding the usage of illegal Al technologies and products. Judicial Al products and services shall be legally developed, deployed and operated, and shall not be detrimental to national security and legal interests. It shall secure state secrets, network security, data security and personal information free from infringement, protect personal privacy, promote a harmonious and friendly interface between the user and Al, and provide safe, legitimate, and efficient intelligent judicial services.” The essence of effective security governance is effective risk governance, and “risk-oriented” is its core requirement. Judicial security governance of artificial intelligence needs to establish a typological, structured, and systematic governance strategy based on deconstructing the core of risk and the causes of risk. Therefore, the prevention of judicial security risks of artificial intelligence requires first clarifying the essence of artificial intelligence justice, and on this basis revealing the substantial risks of artificial intelligence justice, and then proposing corresponding governance strategies.

1. The essence of AI justice: Judges’ withdrawal and algorithmic justice

Fairness and efficiency are the eternal themes of judicial trials, and are also the basic values followed by the judicial application of artificial intelligence. Between fairness and efficiency, the latter is the original driving force for the development of artificial intelligence justice. With the advancement of artificial intelligence technology, the judicial application of artificial intelligence has gradually moved from automated office to automated decision-making. High-intelligence application scenarios such as the same judgment for similar cases, sentencing prediction, and document generation have begun to have an intrusive impact on the exercise of judicial powers such as conviction authority, sentencing authority, control over procedure, and judicial supervision authority. As a result, algorithms will penetrate the veil of efficiency and surface. Algorithmic justice characterized by the judge's retreat constitutes the essence of artificial intelligence justice.

1.1 Judicial efficiency and algorithm pre-positioning

For a long time, China’s “judicial function structure and role arrangement depend largely on the actual needs and choices of system designers rather than the self-positioning of the judiciary”. The problem caused by this system design is that the configuration of judicial judges may be out of touch with the actual needs of the judiciary, and cause the judges in the same court to be “uneven distribution of workload”, which in turn affects the improvement of judicial efficiency. More importantly, facing the arrival of the “litigation explosion” era and the new possibility of efficiency improvement enabled by the progress of digital intelligent technology, how to use digital intelligent technology to improve judicial efficiency and then promote judicial system reform has become an issue of the times that must be responded to. For this reason, since the 18th National Congress of the CPC, the Supreme People's Court has regarded judicial reform and judicial informatization as the “two wheels of a car and two wings of a bird” to promote the development of the work of the people's courts. On this basis, optimizing the configuration of judges with the help of digital intelligent technology and improving judicial efficiency have become the logical starting point and practical end point of digital intelligent technology empowering judicial reform. Facts have also proved that “a comprehensive litigation service system and court office system realize the automatic generation of documents of the people's courts, the intelligent push of similar cases, and the automatic filing of electronic files, freeing the case-handling judges from trivial and repetitive affairs.” Updating trial equipment with the help of digital technologies such as computers and the Internet will not only help to optimize the management of trial personnel, but also enable case handlers to directly review and understand the case through electronic case files throughout the second instance and retrial, which will have a significant effect on improving trial efficiency.

At present, with the progress of generative artificial intelligence, advanced algorithms represented by large language models have gradually moved from “behind the scenes” to “forefront” while promoting judicial efficiency, and have objectively involved core judicial business such as conviction and sentencing, application of law, and document writing. For example, legal logical reasoning has always been considered the core of judicial business and the key content to promote the determination of case facts and the application of law. However, generative artificial intelligence has mastered certain logical reasoning capabilities with the help of the chain of thought algorithm. Taking Chat-GPT as an example, the researchers not only achieved a high degree of accuracy in the benchmark test by calling the chain of thought algorithm, but also exceeded the GPT-3 model fine-tuned with the verifier, significantly improving the ability of large language models to perform complex reasoning. Specifically, with the support of the chain of thought, the model decomposes the problem into a series of step-by-step reasoning, and infers the next step based on the results of the previous step and the requirements of the current problem. Through this step-by-step reasoning method, the model can gradually obtain more information and accumulate correct inferences throughout the reasoning process, thereby greatly improving the accuracy of the model in complex reasoning. This means that advanced algorithms, mainly generative AI, have broken through the limitations of data-based correlations and can produce expected results beyond the algorithm's preset expectations, with the ability to express intentions and other features. In fact, with the significant improvement in the algorithm's logical reasoning ability, generative AI has reached or even exceeded human levels in terms of human-computer question-answering, text summarization, and suggestion assistance, and has the ability to analyze case research and complex legal facts. Take another example, “the writing of judgment documents is the last important procedure in the trial of a case and the most time-consuming task. The quality of judgment documents is positively correlated with the parties’ acceptance of the judgment results. Therefore, after deciding the judgment conclusion of the case, the main task of the judge is to write a high-quality judgment document.” Therefore, whether a judgment document can be written fluently, accurately and with high quality has also become an important indicator of the judge's ability. In practice, the leap in word processing capabilities of algorithms represented by large language models can not only realize the drafting of legal documents with specific content, but also realize the understanding of simple case facts and correctly cite relevant legal provisions of the litigation request. For example, when facing the core area of value judgment in judicial adjudication, by using a cyclic reinforcement learning algorithm to link the feedback model and the original model, the original model can be continuously iterated and upgraded until it is “reborn” and thoroughly grasps human preferences, thereby achieving “alignment” between the expression and intrinsic value of the algorithmic model and human common sense and values.

In short, although some people have raised criticisms that the current application of intelligent systems is not perfect due to factors such as subjective understanding of judicial personnel, technical maturity, and narrow applicability, these criticisms do not negate the efficiency function of artificial intelligence justice from the perspective of value theory. On the contrary, with the advancement of intelligent technology, algorithms are being put in front by enabling judicial efficiency, entering the core areas of justice, and possessing certain judicial decision-making capabilities.

1.2 Judges’ Retreat and Algorithmic Justice

With the increasing prominence of algorithm advancement, the withdrawal of judges as judicial subjects will become a basic feature of AI justice. Therefore, the essence of AI justice is algorithmic justice caused by the withdrawal of judges.

Algorithms refer to the solutions to achieve the optimal model. Algorithmic justice can completely replicate the “appearance” of human judges’ justice, and to a certain extent assist or replace humans in judicial judgment. Particularly, in the judicial field where the rules for the exercise of power are clear and the predictable results are clear, it is technically feasible to use algorithms to solve judicial problems, connect the condition input end and the conclusion output end through algorithms, and realize algorithm-enhanced automated justice. Although different types of artificial intelligence justice have differentiated input and output content, the fundamental difference does not lie in the difference between input and output, but in the difference in algorithmic solutions. In the field of artificial intelligence judicial adjudication, the process of constructing a machine adjudication model is the process of combining algorithmic knowledge with legal data. Among them, reflecting the laws of judicial adjudication through data, mining data correlation through algorithms, and testing model performance in the test set are the three main processes of machine adjudication. In these processes, although the source of knowledge for algorithmic justice is traditional justice, after the machine learns the basic laws of judicial adjudication, it can achieve “behaving like a judge”. As for whether it can ultimately achieve “adjudicating like a judge”, it depends on whether the algorithm level is sufficient to support its “thinking like a judge”. In other words, the judicial application of algorithms will inevitably lead to the retreat of judges. That is to say, compared with the adjudication activities that are completely monopolized by judges, the intervention of algorithms will inevitably lead to the decomposition of judges’ powers, and this process of decomposition is also the process of judges’ retreat.

Of course, some scholars have pointed out that artificial intelligence helps to achieve formal justice in the judiciary, but it is still far from achieving substantive justice that requires the wisdom of judges. But the problem is that this assertion is not a denial of the judge's retreat, but a questioning of the degree of the judge's withdrawal. In other words, the application of algorithms in the judiciary will only decompose the adjudication functions of human judges to a certain extent, but will not completely replace them. The withdrawal of judges is not the departure of judges. The essence of the withdrawal of judges is the partial replacement of the judicial power by algorithms. For example, most adjudications in judicial practice are the application of the monotonous reasoning logic of “legal norms → case facts → adjudication conclusions”. On this basis, the first step of judicial adjudication is to accurately find the legal norms corresponding to the facts of the case and accurately apply them to the facts of the case. In this process, applying legal norms to the facts of the case is not only a key step in completing judicial adjudication activities, but also a prerequisite for promoting the smooth progress of judicial adjudication. Ensuring the accuracy of the application of legal norms is precisely the advantage of algorithms. Compared with judicial adjudications conducted entirely by human judges, the degree of algorithm retrieval of legal norms will only affect the degree of the judge's withdrawal, rather than an absolute denial of the decomposition of the adjudication functions of human judges. For another example, the smart sentencing system, as an artificial intelligence machine that can complete criminal sentencing decisions, generates sentencing conclusions through algorithms. When it occupies a dominant position in the judiciary, it becomes a sentencing judge. When it is in an auxiliary judicial position, it can provide sentencing references, thereby achieving the goals of balancing the criminal circle and mitigating penalties during sentencing, avoiding harsh sentencing results, and when assuming judicial supervision tasks, it manifests itself as an early warning mechanism for different sentencings for the same case.

In short, no matter what its status is, the algorithms in AI justice have not undergone fundamental changes, and they always play the alternative function of “sentencing like a judge”. The withdrawal of judges is the basic feature of AI justice, and algorithmic justice is the essence of AI justice.

1.3 Algorithmic Justice and Formal Rationality

Once algorithms enter the judicial field, the legitimacy of AI justice will inevitably be questioned and criticized. These doubts and criticisms touch upon the value foundation of AI justice and point directly to judicial justice itself, that is, the justice risks of algorithmic justice.

The value foundation of justice lies in justice. The value criticism of AI justice also comes from judicial justice. After the litigation process has been digitally transformed, the academic community generally believes that algorithmic decision-making can optimize the litigation process and handle simple cases through procedural and formulaic calculations, but it is difficult to handle major, complex and difficult cases; it can promote formal justice, but it is difficult to achieve substantive justice. Although AI justice has incomparable advantages over human judges, it also has insurmountable defects. It cannot deal with uncertainty, does not have human common sense, and cannot make value judgments. Justice Zhang Jun also made an insightful analysis of computer sentencing, believing that “Every case is different. Inputting various factors into software and finally obtaining a sentencing situation is difficult to meet the specific reality of the case.” AI justice does have limited corpus and data acquisition capabilities, and cannot exhaust judicial scenarios and their consideration. The case facts, legal rules, individual experience, free evaluation of evidence, and emotional considerations that judges rely on in judicial judgments are difficult to express through structured data so that machine algorithms can learn. For example, in the criminal trial reference case involving a certain Mr. Wang's illegal business case, the defendant sold fake and inferior cigarettes with counterfeit registered trademarks, which constituted the crime of illegal business operation. According to the law, he should be sentenced within the statutory range of imprisonment of more than five years. However, considering that his daughter urgently needs a bone marrow transplant, Mr. Wang is the preferred donor and the only source of family income. The Supreme People’s Court approved the sentencing of Mr. Wang below the statutory sentence and announced a suspended sentence. Obviously, AI justice cannot make such a “leniency outside the law” decision. Because it cannot achieve the integration and balance of emotion, reason and law, AI justice is considered to undermine substantive justice.

The problem is that this questioning does not touch the essence of algorithmic justice, nor does it correctly understand the relationship between formal justice and substantive justice. In fact, judicial justice includes three logical structures: formal justice, substantive justice and procedural justice. Algorithmic justice not only does not impact judicial justice, but also conforms to the consistent formal rational value foundation of traditional judicial justice. Artificial intelligence justice is not only able to achieve formal justice but cannot meet the requirements of substantive justice, because there is a fundamental difference between the concepts of formal rationality and formal justice and they cannot be confused. Formal justice is also called distributive justice, which means equal treatment of equal situations and different treatment of different situations. Formal justice leads to the same judgment for the same case, but does not ask whether the judgment result meets the substantive requirement of punishment commensurate with the crime; substantive justice incorporates the substantive judgment of result fairness and requires a punishment judgment that complies with the principle of proportionality between crime and punishment and the principle of proportionality. Algorithmic justice is by no means a simple formal justice. The algorithm-led automated decision-making mechanism can effectively exclude interference from external factors to strictly achieve normative compliance, achieve formal justice and predictable justice, and at the same time, it can also integrate the social harm and proportionality of punishment results required by substantive justice into the algorithm through factorization and structured data annotation, thereby realizing the pursuit of substantive justice. For example, an AI system with deep learning capabilities can even automatically generate new element annotations without being limited by past adjudication experience. As mentioned earlier, with the help of large language model algorithms, generative AI has been able to achieve logical reasoning and value alignment, reflecting the inherent requirements for substantive justice. Therefore, the true connotation of algorithmic justice is a substantive justice dominated by formal rationality. The reason is that between formal rationality and substantive rationality, formal rationality has a superior principle and is the priority, while substantive rationality is the second priority, and substantive rationality can only be pursued under the premise of adhering to formal rationality. On this basis, formal rationality that regards compliance with legal norms as the primary value pursuit and makes substantive corrections based on value norms can obviously encompass formal justice and substantive justice, and can accurately summarize the substantive connotation of algorithmic justice.

In general, the essence of AI justice is algorithmic justice. When algorithms can achieve the distribution of rights, obligations and responsibilities, doubts and criticisms of algorithms will follow. But the problem is that the denial of algorithms based on the distinction between formal justice and substantive justice does not touch the key to algorithmic justice. In fact, given that fair justice is the last line of defense in maintaining social fairness and justice, it is necessary to continue to deepen the comprehensive and coordinated reform of the judicial system. The core of algorithmic justice is to use algorithms to achieve formal rationality in following legal norms.

2. Risks of AI Justice: The Triple Crisis of Algorithmic Justice

Algorithmic justice contains the connotations of formal justice and substantive justice, and has the formal rationality characteristics of traditional judicial justice. However, there are many differences between artificial intelligence justice and traditional judicial models. It is these differences that have caused a crisis in algorithmic justice and become the root cause of the judicial security risks of artificial intelligence. Therefore, it is necessary to analyze security risks based on application scenarios. In previous academic research, the portrait of the judicial security risks of artificial intelligence has been initially formed. For example, in the technical dimension, there are risks such as deep involvement of personal information, high uncertainty of scenario impact, high-intensity impact on the existing order, and ubiquitous threats to the public. These risks have brought about the elimination of the inherent attributes of justice, the weakening of the subject status of judges, the replacement of judicial reform goals, and the loss of control of judicial reform results. However, to truly examine the substantive crisis of algorithmic justice, it is necessary to conduct research from the basic dimensions of judicial justice.

2.1 Procedural Risks of Algorithmic Justice

It is generally believed that the participation, neutrality, and equality of the trial process are the basic components of the intrinsic value of the process. However, algorithmic justice that follows legal norms has, to a certain extent, challenged the intrinsic elements of procedural justice, thereby triggering procedural risks of algorithmic justice.

Firstly, algorithmic justice may eliminate the participation of judicial procedures. The participation of procedures is reflected in two aspects: on the one hand, both parties to the litigation have ample opportunities to participate in the process of making judicial judgments in terms of time and space, and can freely express their views and present evidence. On the other hand, it is possible to achieve substantive feedback on the results of judicial judgments and participate in the production of judicial judgments through debate. For example, Article 182, Paragraph 3 of the current Criminal Procedure Law of China stipulates that the People’s Procuratorates, parties, defenders, litigation agents, witnesses, appraisers and translators shall be informed of the cause of the case, the time and place of the court session, which is a guarantee for the time and space for the participation of both parties to the litigation. Article 186 of the Criminal Procedure Law stipulates that the prosecutor reads the indictment and the defenders and litigation agents ask questions, which is a substantial guarantee for procedural participation. However, the extreme pursuit of efficiency by artificial intelligence justice has, to a certain extent, compressed the due process and ignored the bridge of communication between the people and the judiciary. The algorithm-based automated judgment outputs the corresponding results the moment relevant information is input. The automated decision-making process makes the judicial process an automatic “vending machine”. The two parties are likely to have a subjective sense of distrust when confronting the cold machine. The subjective procedural justice value of “letting the people feel fairness and justice in every judicial case” is difficult to achieve.

Secondly, algorithmic justice eliminates the neutrality of the judicial process. In comparison, artificial intelligence is more neutral in results because it is not interfered by avoidance, fatigue, emotion and other issues. But this does not mean that algorithmic justice has no impact on the neutrality of the judicial process in any dimension. On the contrary, although the algorithm model has been continuously improved and optimized, phenomena such as algorithmic bias and algorithmic discrimination still exist, and thus eliminate the neutrality of the judicial process to a certain extent. In fact, “bias in, bias out” seems to have existed since the moment the algorithm was born. The training data of the algorithm, the training model of the algorithm and the output results of the algorithm may all lead to the phenomenon of algorithmic bias. In practice, even the most advanced generative artificial intelligence cannot achieve absolute freedom from algorithmic bias. While the large language model improves the intelligence and accuracy of ChatGPT, it also causes the legal risk of algorithmic bias to increase exponentially. Based on the output results of GPT-3, there are already a lot of biases based on gender, race and religion. It can be seen that while the algorithm strengthens judicial neutrality, it will also eliminate the neutrality of the judicial process due to the influence of algorithmic bias, thereby causing procedural risks of algorithmic justice.

Thirdly, algorithmic justice eliminates the equality of judicial procedures. Procedural equality requires that both parties to the litigation are evenly matched, that is, the prosecution and the defense are equal, in order to prevent the situation of “the strong preying on the weak”, which will have an adverse impact on the final decision of the case dispute. In terms of system, the setting of rights such as the right to remain silent and the right to cross-examination are concrete manifestations of procedural equality. However, the practical logic of algorithmic justice may have a certain negative impact on procedural equality. Taking the protection of the right to cross-examination as an example, whether it is the current intelligent evidence guidance system or the intelligent evidence analysis system, their judicial application is, in fact, based on massive data and intelligent algorithms. For example, the intelligent evidence analysis system launched by the Hangzhou Court of the Internet can compare evidence such as text works, pictures, and videos, find similarities, and realize the review of evidence. In this process, the application of intelligent technology in the field of evidence is actually to realize the comparison and verification of evidence through algorithms. The algorithmic comparison and verification are essentially a replacement for the inquiry and confrontation in the principle of cross-examination.

In short, although algorithms can ensure the efficiency of litigation procedures and improve the convenience of litigation, automated procedural settings will also affect the participation, neutrality, and equality of the procedures, causing procedural risks of algorithmic justice.

2.2 Physical Risks of Algorithmic Justice

There have been relatively solid studies that prove that AI justice is more conducive to accuracy and fairness in terms of results, has higher acceptability, can provide more comprehensive information, and is more effective in evaluating social effects, reflecting the universal justice function of algorithmic justice under the guidance of formal rationality. As for the realization of justice in special cases such as sentencing below the statutory penalty, it can be completely calibrated manually without affecting the basic value of algorithmic justice. However, algorithmic justice requires a scientific coding process of adjudication rules. Once the algorithm and related parameters are not reasonably designed, systematic deviations may occur, turning universal justice into universal injustice, which will have a huge impact on judicial justice.

Whether it is analytical artificial intelligence such as big data, cloud computing, and blockchain, or generative artificial intelligence, artificial intelligence justice uses algorithms to achieve the identification of case facts and the application of legal norms. In this process, it is necessary to digitize evidence, perform computations and integration on data, and output conclusions that people can understand. In other words, the digital transformation of legal fact identification and the application of legal provisions, and then the output of algorithmic calculations, is the basic process of algorithmic justice. For example, in the identification of legal facts, the application of big data evidence and big data investigation methods has become increasingly mature. Unlike traditional physical evidence, in the face of evidence with non-perceptible digital attributes, the importance of human senses in fact identification has begun to decline, and the identification of legal facts has begun to shift from sensory experience to the correlation between data. In this case, the identification of case facts is gradually changing from people-centered to digital-centered, showing a digital trend in the identification of legal facts. Similarly, the application of law has also shown a significant digital trend under the embedding of digital technology. The ambiguity of the concept of legal provisions requires that the law must be interpreted in order to be accurately applied. However, in the process of applying specific legal rules, there is not only legal interpretation, but also a factual interpretation that is independent of legal methods. Therefore, the application of law in the trial process is often a time-consuming and laborious process. At present, artificial intelligence is gradually developing into explainable artificial intelligence, and digital-driven legal application can simplify this process. For example, management systems represented by “smart case management” can automatically analyze and warn of deviations in legal application by integrating functions such as similar case retrieval and sentencing norms, ensuring the same judgment for similar cases. It is important that the digitization of fact finding and legal application not only provides the possibility of coding and automation for fact finding and legal application, but also enables the trial results to be presented in a more accurate and objective digital form, reflecting the universal justice of algorithmic justice under the guidance of formal rationality.

But the problem is that the fact finding and legal application based on algorithms are completed by finding the correlation between numbers in the legal sample data through the algorithm model. In essence, this correlation model is just a probabilistic and computational mathematical statistics and fitting. Even the most advanced generative artificial intelligence only applies probability theory more, and calculates the probability of the distribution of collocations between words with huge amounts of calculations through training on a large amount of data. Therefore, the accuracy of the data and the scientific nature of the model setting will affect the final output results. For example, due to the high reliance on historical data, artificial intelligence crime risk assessment will inevitably lead to the system reproducing the discrimination pattern and historical bias pattern in the data. In fact, “when we question human arbitrariness, we have more reasons to suspect that these system developers will write the bias of legal scholars, the arbitrariness of scientists, the economic interests of enterprises, etc. into the black box of the algorithm. They still cannot escape the manipulation of commercial, political, and strong values.” In other words, as long as the data used for training is cleaned to a certain extent and the model for running the algorithm is set up scientifically, the fact-finding and legal application based on the algorithm can focus on individual cases. The substantive risk of algorithmic justice does not lie in only focusing on formal justice and ignoring the “extralegal leniency” of the judgment results, but in the rationality and scientific nature of the algorithm's own data and model.

In short, while algorithms embody substantive justice under the guidance of formal rationality, they can also be affected by factors such as source data and relevant parameters, which in turn affect the accuracy of the actual results and trigger actual risks under algorithmic justice.

2.3 Technical Risks of Algorithmic Justice

Digital and intelligent characteristics of artificial intelligence justice determine that the realization of algorithmic justice faces inherent security risks that are different from the realization of traditional judicial justice. The main risks are the operation and maintenance security risks of artificial intelligence systems, which specifically involve the various processes of artificial intelligence programs to collect, store, transmit, exchange and process data. According to the relevant statements in the “Artificial Intelligence Judicial Opinion”, data security, personal information security and artificial intelligence system security are two important components of the inherent security risks of algorithms, and the intensity of security protection affects the public's trust in the judicial norms system.

The smooth operation of AI justice cannot be separated from high-quality data support. To a certain extent, the larger the amount of data in the field and the higher the quality, the better the effect can be achieved. For example, the key reason why generative AI represented by ChatGPT can achieve such remarkable results is that it relies on the training of large language models with massive parameters. But for now, the main technical framework of large language models comes from foreign countries and is based on Western values and thinking orientation. Therefore, the answers in them usually cater to Western positions and preferences, which may lead to ideological infiltration. Justice is not only related to the distribution of individual rights and obligations and the resolution of social contradictions, but also closely related to the core values and ideology of the country. Therefore, on the one hand, it is necessary to regulate the data processing of AI justice and strengthen the supervision of data application. On the other hand, it is also necessary to strengthen the management of cross-border flow of judicial data and prevent data that endangers national security from being used abroad. To this end, China's Data Security Law specifically stipulates data security. Personal information security is another manifestation of algorithmic risk. In the digital age, all behaviors based on individual information data will become “traceable”. In order to obtain more adaptive and more accurate algorithm services, it is inevitable to involve the collection and processing of personal information. In practice, many countries such as Germany, the United Kingdom, and Canada have expressed their intention to investigate OpenAI on the grounds of personal information rights protection and data security. Therefore, personal information security has also become one of the inherent risks of artificial intelligence technology. In addition, affected by the bottleneck of existing technology, artificial intelligence systems themselves also have unavoidable security risks. For example, government departments and related organizations in various countries have realized that the development of artificial intelligence technology requires the formulation of a standardized and reliable framework to protect humans from harm. However, as far as current technology is concerned, there is still a certain gap in achieving complete algorithm transparency and algorithm explainability.

In short, AI justice, which is essentially algorithmic justice, is also facing the triple crisis of procedural risk, entity risk and inherent technical risk on the basis of reflecting formal rationality. But in the final analysis, the real reasons for the above crises are factors such as the data for algorithm training, the model of algorithm operation, and the results of algorithm output. Algorithms are the root cause of the crisis.

3. Improvement of AI Judicial System: Algorithm-centered Governance Model

Focusing on algorithmic justice, algorithmic justice, and the risks caused by algorithms, the improvement of AI justice should be centered on algorithms, and the in-depth promotion of AI justice can be achieved through the construction of an algorithm-centered governance model. At present, the three basic stages of AI operation are the preparatory stage of pre-learning training and manual annotation-assisted algorithm upgrades, the operation stage of processing input data by the algorithm itself and obtaining processed data outputs, and the generation stage of data outputs flowing into society and affecting all walks of life in society. Therefore, training data, operation models, and output results are three important components of the algorithm-centered governance model.

3.1 Strengthening the data of algorithm training

Data is an important symbol of knowledge formation and civilization progress, and is the product of human understanding of the objective world. Data not only concerns the protection of digital rights and interests of the country, society, and the public, but is also closely related to the objectivity and neutrality of algorithm output results. Therefore, the enhancement of algorithm training data should be carried out from the two aspects of digital rights and interests and digital sources.

Specifically, in terms of digital rights, the enhancement of algorithm training data should be carried out in strict accordance with the legal framework of informed consent to enhance the legitimacy of data use. At the legal level, China has successively formulated a series of laws such as the Data Security Law and the Personal Information Protection Law, which have respectively regulated the collection and processing of data. At the same time, in November 2022, the “Provisions on the Management of Deep Synthesis Internet Information Services” jointly issued by the Cyberspace Administration of China, the Ministry of Industry and Information Technology, and the Ministry of Public Security stipulated that deep synthesis service providers and technical supporters should take necessary measures to ensure the security of training data and comply with relevant regulations on personal information protection. If the function of editing biometric information such as face and voice is provided, the user of deep synthesis service should be prompted to inform the edited individual in accordance with the law and obtain their separate consent. The “Interim Measures for the Administration of Generative Artificial Intelligence Services” implemented on August 15, 2023 also reiterated that generative artificial intelligence service providers should use data and basic models with legal sources, information with personal consent, etc. to carry out pre-training, optimization training and other training data processing activities in accordance with the law. It can be seen that although it is relatively scattered in the system, according to relevant laws and regulations, personal data processors must have a legal basis before processing personal data. On this basis, the enhancement of algorithm training data can be achieved through the construction of rules at three levels: personal information collection, provision and sharing, to prevent security risks caused by the application of intelligent technology. In terms of digital sources, the key to the enhancement of algorithm training data lies in the construction of a unified judicial digital resource library. According to the planning of the “Artificial Intelligence Judicial Opinion”, in the future, the promotion of artificial intelligence judicial applications needs to accelerate the construction and integration of judicial databases, data service platforms, judicial knowledge bases, artificial intelligence engines and other systems, create a physical judicial data middle platform and a smart court brain, and provide a core driver for artificial intelligence judicial applications for various businesses. On this basis, the digital resource library will be an important part of the construction of artificial intelligence judicial and the basic construction of artificial intelligence judicial. To this end, the construction of the judicial digital resource library needs to be carried out from the following two aspects: First, through the construction of the judicial digital library, the unification of existing judicial resources is realized, forming a multimodal digital resource library including text digital such as laws and regulations, judicial documents, legal journals, legal monographs, video digital such as court trial live broadcast and court trial video, and information digital such as litigation process and judicial management, so as to provide sufficient digital resources for intelligent algorithm training. Second, through the construction of a judicial digital database, key numbers can be cleaned and screened, so that the data for algorithm training has achieved the anonymization of personal information, the cleaning of sensitive information, and the screening of distorted information, thereby enhancing the high quality of the data.

In short, as the basis and prerequisite for algorithm governance, data must not only promote the compliance and legality of judicial data collection and processing and strengthen data protection, but also improve the quantity and quality through the construction of judicial data resource libraries to provide sufficient fuel for algorithm training.

3.2 Model of optimization algorithm operation

Optimizing the model can significantly enhance the algorithm's comprehension and interaction capabilities, thereby promoting the perfection of artificial intelligence in the judiciary. Facing the future practice of generative AI, the new generation of legal intelligence systems shows a trend of integrating large language models with legal expert knowledge. Therefore, the optimization of the algorithm model can be carried out from the following two aspects.

On the one hand, with the help of fine-tuning instructions in the large language model, the algorithm model can be optimized and the improvement of artificial intelligence justice can be promoted. Practice has proved that with the help of fine-tuning instructions, generative artificial intelligence represented by ChatGPT can not only handle natural language core tasks such as text classification, semantic analysis, information extraction, text generation, automatic abstracts, machine translation, dialogue systems, and information retrieval, but also realize the transformation of instruction expression from “human adaptation to machine” to “machine adaptation to human”. As long as humans open their mouths to express their demands, artificial intelligence can understand them and help humans to answer. Therefore, a possible path to optimize the algorithm model is to strengthen the fine-tuning instructions in the judicial field. Fine-tuning instructions for the judicial field are mainly concentrated on the following three points: First, construct interactive fine-tuning instructions with legal language as the main application scenario, improve the interactive ability of artificial intelligence technology in the judicial scenario, and then improve the participation of the parties in the algorithm decision-making process. Second, strengthen the fine-tuning instructions of value alignment to correct algorithm bias and promote the objectivity of algorithm results. Third, explore the interpretability of algorithms in the technical dimension, promote the transparency of algorithm models, and eliminate the inherent technical risks of algorithms. On the other hand, an algorithmic model with the participation of legal experts is constructed to strengthen the interpretability of AI justice through the operation of legal knowledge in the algorithm. Algorithmic justice does not naturally isolate procedural justice from substantive justice. On the contrary, procedural justice elements such as participation, neutrality, and equality can be fully guaranteed through human-computer interaction design. For example, the personal experience of judges in the digital intelligence era requires not only the physical presence of judges in the decision-making process, but also the digital presence of judges in the intelligent litigation process. The input control and result inspection of key elements of pending cases can be achieved through the adjustment of the correlation coefficient of the intelligent litigation system, and then the final judgment can be made. Similarly, the legal knowledge system of legal experts is used to realize the construction of the legal knowledge map, and then the algorithmic model of AI justice is improved.

In short, with the development of artificial intelligence technology, fine-tuning instructions and optimization algorithms for intelligent technology are also making significant progress. The composite algorithm of “legal knowledge graph + large language model” has also become possible. In this context, the security risks caused by algorithms can be solved to a certain extent by optimizing the algorithm model.

3.3 Results of the correction algorithm output

Whether it is procedural risk, entity risk or inherent technical risk, risk will always be manifested in the form of algorithm output results. Although the essence of AI justice is algorithmic justice, which presents the practical characteristics of judges withdrawing, as mentioned above, the judge's withdrawal does not mean the judge's departure, but a change from “front-stage leadership” to “behind-the-scenes supervision”. Substantive justice under the dominance of formal rationality still requires data annotation to achieve the alignment of algorithmic values. Therefore, in the field of algorithmic justice, judicial personnel still need to correct the results of algorithm output. In practice, based on the reinforcement learning mechanism of human feedback, generative artificial intelligence represented by ChatGPT has been able to answer questions in a way that is consistent with human intentions, knowledge and values, and achieve value alignment. On this basis, the correction of algorithm output results can be achieved through the construction of a feedback mechanism, thereby achieving the improvement of AI justice.

The correction of algorithm output results can be carried out from the following three points. First, construct a feedback mechanism for AI justice from the institutional level to realize the correction of algorithm output results. According to the plan of the “Artificial Intelligence Judicial Opinion”, in the future stage, through mechanisms such as the Judicial AI Ethics Committee, comprehensive ethical review, compliance review, and security assessment will be adopted to prevent and resolve security risks in the application of AI. In 2022, the “Guidelines on Strengthening the Governance of Technological Ethics” issued by the General Office of the CPC Central Committee & State Council Office also provided guidance for strengthening the construction of the Science and Technology Ethics (Review) Committee. In October 2023, the “Measures for scientific and technological ethics review (for trial implementation)” jointly issued by the Ministry of Science and Technology, the Ministry of Education, the Ministry of Industry and Information Technology and other ministries and commissions made specific provisions on the subject, content and procedures of science and technology ethics review. Therefore, a judicial AI ethics committee can be constructed at the institutional level, and the committee can realize the review and feedback of the algorithm output results. The relevant technical departments will correct the algorithm based on the feedback from the committee, forming a closed-loop mechanism of “algorithm result output-review feedback by the review committee-algorithm optimization and re-output”, thereby promoting the improvement of AI justice. Second, from the technical level, an open correction interface with feedback and error reporting should be constructed to guide and encourage judicial personnel to supervise the output results of the algorithm. It has been proved that algorithm training based on human feedback can greatly improve the accuracy of intelligent technology. For example, by calling the reinforcement learning technology of human feedback, ChatGPT has unprecedentedly achieved a deep simulation of human cognitive mechanism, which can better realize the refinement of relationship connection, identification of individual elements and content generation. On this basis, the development of relevant algorithms for artificial intelligence justice should reserve an open feedback interface so that the majority of judicial personnel can timely and conveniently realize the feedback of the algorithm output results and form a result feedback database. The technical development department will adjust and train the algorithm based on the feedback database to achieve the improvement of artificial intelligence justice. Third, from the individual level, the digital verification ability of judicial personnel should be improved, and an individual mechanism for correcting the output results of the algorithm should be constructed. Faced with the trend of the times in which artificial intelligence justice is rolling forward, as the main body of justice, the majority of judicial workers cannot be completely separated from intelligent technology. In fact, in the digital age, the reconstruction of labor content by digital technology forces workers to acquire new skills to adapt to new positions. Mastering and proficiently applying digital technology has become an essential skill in the digital age. Specifically in the field of AI justice, whether the output of the algorithm can be identified and verified is an important aspect of the digital skills of judicial personnel. On this basis, to correct the output of the algorithm, it is also necessary to improve the digital skills of judicial personnel from an individual dimension.

In short, the core of AI judicial perfection lies in algorithms, and corresponding governance should be carried out around the three dimensions of algorithm training data, algorithm operation models, and algorithm output results. Construct a three-dimensional governance model of “data-model-result”. Under this model, data, as the premise and foundation of AI justice, should strictly follow the existing regulations of China and promote the collection and processing of judicial data under the framework of informed consent and data security. As the core of AI justice, the algorithm model can optimize the model with the help of the latest technological achievements such as fine-tuning instructions and legal knowledge graphs, and then promote the possible systematic deviations of the algorithm model. As the terminal of AI justice, the results of the algorithm output need to be carried out through the construction of the ethics review committee at the institutional level, the feedback mechanism interface at the technical level, and the digital skills at the individual level to realize the correction of the algorithm output results.

Conclusion

At present, with the progress of generative artificial intelligence, human beings are entering the era of general intelligence. Intelligent technology will inevitably bring unprecedented opportunities and challenges to human social life. As an important part of social life, judicial life cannot stay out of it. Although the judicial response of “stepping forward in small steps” adheres to the ancient principle of “judicial stability”, it also conflicts with the “agile governance” required in the intelligent era. To this end, we must face the new possibilities and potential risks brought by intelligent technology, penetrate the veil of artificial intelligence justice and get to the essence. Only in this way can we coordinate the judicial development and judicial security in the intelligent era and safeguard the new development pattern with a new security pattern. On this basis, the essence of artificial intelligence justice is algorithmic justice with the retreat of judges as the practical feature, that is, algorithmic decision-making replaces judges' decisions as much as possible. Although the Supreme People's Court has repeatedly emphasized the auxiliary nature of artificial intelligence justice, the necessity of practice and the inevitable development of technology will lead to the nature of algorithmic justice that cannot be changed. Algorithmic risk is the core of artificial intelligence judicial risk. Therefore, the core of the governance of artificial intelligence judicial risk is algorithmic governance, not judicial management governance and compliance risk governance. Only with algorithms as the core and specific governance solutions can the security risks of artificial intelligence justice be effectively prevented.

The original article was published in the 4th issue of “Oriental Law” in 2024 (Special Issue Featuring Distinguished Scholars). Thanks to the WeChat public account “Shanghai Law Society Eastern Law” for the authorization to repost.