[author]Zhao Qian

[content]

On Regulatory Strategies for Generative Artificial Intelligence in Assisting Judicial Adjudication

Zhao Qian

Professor, Law School of Southwest University

Executive Committee Member, CCF Computational Law Branch

Abstract: The application of generative artificial intelligence in judicial adjudication represents a critical step in advancing smart court construction and deepening China,s judicial reforms. Regulation of this technology aims to leverage advanced artificial intelligence tools with enhanced technical rationality, aligning with the integration requirements between technology and justice. By examining the technical characteristics of generative artificial intelligence, this paper clarifies and standardizes the operational landscape of artificial intelligence in judicial applications. To explore regulatory strategies for generative artificial intelligence in boosting adjudication, we must first establish the functional positioning of generative artificial intelligence as a regulatory rationale and its role in judicial processes. Three regulatory paradigms and their core principles are elaborated based on port-based technical support, automated document processing, and standardized case adjudication. The first one is starting-point regulation for technical support under portbased positioning, which aims to clarify port-based integration of data processing and port-based matching of text generation, providing essential material support for optimizing judicial adjudication through hightech features. The second one is instrumental regulation for document processing under high-efficiency positioning, which involves automated decoding of judicial document elements and automated drafting of document content, seeking to convert computational language generated from adjudication data into legal text more efficiently. The third one is targeted regulation for case adjudication under high-quality positioning,which focuses on standardized factual identification of similar cases and standardized application of case precedents, promoting objective decision-making in the knowledge genealogy of adjudicative data and the standardization of case adjudication criteria.

Keywords: Generative Artificial Intelligence; Smart Courts; Starting-point-based regulation; Means-based regulation; Goal-based regulation

Introduction

The Communique of the Third Plenary Session of the 20th Central Committee of the Communist Party of China established “deepening and standardizing judicial openness, implementing and improving the judicial responsibility system” as an important measure for “improving the systems and mechanisms for fair law enforcement and judicature.” As a key aspect of “coordinated promotion of reforms in judicial links,” it constitutes a specific manifestation of practicing “improving the socialist rule of law system with Chinese characteristics.” The use of generative artificial intelligence to assist judicial adjudication is integral to advancing the construction of smart courts and deepening China's judicial reform. Regulation in this area aims to leverage more technologically rational strong AI methods, attempting to clarify and even standardize the operational landscape of AI judicial applications under the background of judicial openness and judicial responsibility system reform. The provisions regarding the “scope of AI judicial applications” in the 2022 Supreme People's Court's Opinions on Regulating and Strengthening the Judicial Application of Artificial Intelligence also point the direction for actively addressing potential risks such as dilemmas in smart judicial transformation and bottlenecks in law-technology convergence, thereby exploring regulatory approaches for generative artificial intelligence in assisting judicial adjudication.
1. Regulatory Rationale for Generative Artificial Intelligence Boosting Judicial Trials
The regulatory rationale serves as the clue guide for exploring the regulatory strategies of generative artificial intelligence boosting judicial trials. It usually avoids the possible disorder of human-machine interaction caused by vague clue positioning by sorting out the corresponding technical characteristics of generative artificial intelligence. In recent years, based on the judicial reform practices of “AI applications in judicature in terms of adjudication acceptability, information completeness, information screening, and fact evaluation”, some scholars have also attempted to position “judicial AI case allocation” and legal management such as “dispute point induction and sorting, similar-case deviation prompt, retrial case adjudication deviation warning, closed-case verification, automatic inspection of irregular judicial acts, and integrity judicial risk prevention and control” as the core functional fields of current AI applications in judicature. In terms of similar-case adjudication, judicial AI can also, based on the “technological supremacy” stance, exert the role of objective AI in promoting the unity of similar-case handling and try to independently generate formatted and modularized judgment documents. Although it helps unify review standards, compress the space for arbitrary adjudication, and realize similar-case similar-judgment, the “quality of judicial products provided by similar-case similar-judgment systems is limited”, and it mostly only plays the role of integrating referee data and extracting general experience. Based on the situation, to clarify the regulatory rationale of generative artificial intelligence boosting judicial trials, we may follow the guidance that “super-intelligent human-like bionics or human-machine coupling interaction is the future development direction” and carry out the discussion based on the functional positioning of generative artificial intelligence and the application dimensions of generative artificial intelligence in facilitating judicial trials.
1.1 Functional Positioning of Generative Artificial Intelligence
Generative Artificial Intelligence is a type of strong artificial intelligence technology “based on natural language processing (NLP) technology and large language models (LLMs)”. It often emphasizes focusing on the orientation of science and technology as the primary productive force, and exerts enabling effects from the perspective of being a new type of labor tool element to promote the leaping development of productivity. Based on this, following the goal of “developing productive forces characterized by high technology, high efficiency, and high quality”, we may try to clarify the functional essence of generative Artificial Intelligence in view of the differentiated positioning under the guidance of three types of characteristics.
1.1.1 Port-oriented Positioning under High-tech Characteristic Guidance
Port-oriented positioning is the functional expression of generative artificial intelligence adapting to the guidance of high-tech characteristics. It mainly focuses on the construction of digital-intelligent integration platforms, and emphasizes laying the foundation for generative artificial intelligence to efficiently empower relevant professional fields by relying on the continuous improvement of generative artificial intelligence LLMs equipped with various rich corpora. This type of functional positioning is often placed under the trend of digital and intelligent social development in the dimension of the digital ecosystem. Following the overall cognition of the three-stage technical connection path of value presupposition, language training, and human-machine interaction involved in generative artificial intelligence algorithm models, it emphasizes that the rational circulation and utilization of data elements is the material basis for improving the computing power of LLMs. It not only requires relevant norms to fully protect the security of personal information and data while maximizing the elimination of unreasonable regulatory barriers to data training, but also tries to guide the relevant technology development and application subjects to fully respect the general laws of LLM iteration. Based on this, generative artificial intelligence under the port-oriented functional positioning should first be placed under the premise of “ensuring that the collection and aggregation of training data are controllable and that the source and content of training data are legitimate”, and clarify the possibility and feasibility of its “deep learning ability” helping become a technology port for the interconnection of all things. Furthermore, this type of strong AI technology needs to rely on systematic mechanisms for deep learning using convolutional neural networks (CNNs), natural language processing, large language models, pre-trained large models, etc., and gradually clarify the data element connection path characterized by data processing adaptability, data output result accuracy, and text generation form coherence and content accuracy during the data training process. It aims to actively respond to various uncertainties and even potential risks caused by the accelerating evolution of LLMs, and tries to promote the interconnection and mechanism integration of various public services, minimizing the negative impact of emerging technologies on social development.
1.1.2 Automation-oriented Positioning under High-efficiency Characteristic Guidance
Automation-oriented positioning is the functional expression of generative artificial intelligence adapting to the guidance of high-efficiency characteristics. It mainly relies on high-efficiency human-AI collaboration to continuously strengthen the enabling effect of generative artificial intelligence in reducing time costs and improving labor efficiency. In the process of exerting the high-efficiency characteristics of generative artificial intelligence, this type of functional positioning often emphasizes adhering to the scientific and technological ethics bottom line of “upholding human dignity and rights and respecting human values” by actively responding to new technical governance issues around “the drawbacks of algorithmic governance”. Based on this, under the automation-oriented functional positioning, generative artificial intelligence should consciously guard against technical risks such as algorithmic black boxes, multi-center or even decentralized algorithmic control, and questionable legality and reliability of training data, while adopting a more inclusive stance to actively respond to the digital-intelligent trend that the versatility of basic models obtained through pre-training is showing an isomorphic logic with human intelligent behaviors. Furthermore, this type of strong AI technology should clarify its automated functional positioning of independently detecting, processing, analyzing, judging, controlling, and realizing the independent generation of target texts based on the consideration of “enabling such rights to be implemented as soon as possible and realizing the transformation from rights to interests”, and strengthen its ability to independently process and improve existing knowledge levels through generative pre-training conversion and other methods.
1.1.3 Standardization-oriented Positioning under High-quality Characteristic Guidance
Standardization-oriented positioning is the functional expression of generative artificial intelligence adapting to the guidance of high-quality characteristics. It mainly refers to continuously adjusting and optimizing LLMs through measures such as “fulfilling content review obligations, fulfilling special labeling obligations, and establishing mechanisms for preventing, promptly identifying, and stopping the generation and dissemination of harmful and inappropriate information”, and striving to improve data accuracy and generate high-quality text as the result of technical empowerment. This type of functional positioning is often based on the technical self-optimization stance of clarifying the governance of technical issues by relying on effective technical means and tools, and emphasizes following the general and standardized approach guidance, and consciously avoiding technological empowerment dilemmas such as information cocoons, big data price discrimination, and algorithmic discrimination by embedding interpretable data standards into the algorithm architecture of LLM training. Based on this, generative artificial intelligence under the standardization-oriented functional positioning should be placed in the context of the widespread use of relatively strong NLP capabilities, and fully respect the general laws of emerging technology development. Furthermore, this type of strong AI technology needs to rely on the positive actions of different responsible subjects in strengthening data verification and quality control, ensuring the accuracy and reliability of data elements, in accordance with the goal guidance of refined organizational scale, multiplied management performance, and refined governance measures, and effectively exert the technical boosting role of different types of data elements by unifying the relevant codes of conduct and norms.
1.2 Application Dimensions of Generative Artificial Intelligence Boosting Judicial Trials
Generative artificial intelligence boosting judicial trials is a specific field for generative artificial intelligence to promote the modernization of the national governance system and governance capacity in accordance with the guidance of “attaching importance to the use of modern information technology means such as artificial intelligence, the Internet, and big data to improve governance capacity and governance modernization level”. It is necessary to sort out the three application dimensions of generative artificial intelligence in technical support, document processing, and similar-case adjudication based on the systematic deployment of judicial system reform.
1.2.1 Technical Support as the Logical Starting Point
The technical support role is often based on the port-oriented functional positioning of generative artificial intelligence. It emphasizes promoting the broader and deeper universal impact of judicial activities on the social organism by relying on the technical empowerment of generative artificial intelligence, realizing the organic coupling between the port-oriented technical supply of generative artificial intelligence and the technical needs of the integrated construction of smart courts, thus becoming the logical starting point for boosting judicial trials. To meet the macro needs of China’s modernization drive for the advancement and upgrading of AI technology, generative artificial intelligence mainly tries to systematically identify, store, and even deduce interdisciplinary knowledge content through data processing in organizing and integrating information processing models required for generative learning, and realizes the innovative and generative application of integrated knowledge systems by relying on the development and iteration methods of knowledge systems carried out around universal data discourse. In the specific field of generative artificial intelligence boosting judicial trials, it is conducive to “carrying out domain technological innovation and promoting domain design justice”, and then effectively improving judicial efficiency by relying on digital intelligent technology. Based on this, the technical support dimension of generative artificial intelligence boosting judicial trials should follow the logical path of the organic combination of technical rationality and value judgment involved in AI rule of law in the modernization process, and carry out specific work around the construction requirements of AI judicial application systems such as “trial assistance systems based on a new generation of artificial intelligence”. It is necessary to fully release the technical potential of generative artificial intelligence in strengthening the reasoning of judgment documents, promoting information exchange, reducing cognitive biases, and enhancing trust and consensus, from the two aspects of port-based integration of data processing and port-based matching of text generation, based on the autonomous response technical requirements achieved by training AI systems with massive data.
1.2.2 Document Processing as the Auxiliary Means
The document processing role is often based on the automation-oriented functional positioning of generative artificial intelligence. It emphasizes realizing the organic coupling between the automated technical supply of Generative artificial intelligence and the efficiency needs of the integrated construction of smart courts through data automated analysis and task content automated filling promoted by independent learning and model training, thus becoming the auxiliary means for boosting judicial trials. Although generative artificial intelligence is mostly only based on formal reasoning of relevance rather than highlighting the causal considerations required by legal reasoning, and it is difficult to “fully respond to the field needs of richness, rigor, and creativity of legal knowledge”, it can still emphasize that LLMs constructed by deep neural networks, convolutional neural networks, recurrent neural networks, etc., while promoting the improvement of judicial efficiency, reasonably exert their positive roles in promoting automatic document generation, automatic file archiving, and intelligent similar-case pushing. Based on this, the document processing dimension of generative artificial intelligence boosting judicial trials needs to follow the consensus requirements of judicial reform such as fairness and efficiency, consciously avoid the security risks of AI systems exceeding their auxiliary positioning, and establish the basic principle that the auxiliary results of AI technology are only for work reference. It often clarifies the human-machine collaborative and automated technical advantages of generative artificial intelligence from the two aspects of automated decoding of document elements and automated drafting of document content, and further emphasizes the in-depth use of intelligent assistance and big data technology based on the organic and orderly planning of relevant informatization construction, trying to alleviate the chronic problems of “litigation explosion, litigation delay, and case backlog” and reduce the transactional work pressure of relevant staff.
1.2.3 Similar-case Adjudication as the Core Goal
The similar-case adjudication role is often based on the standardization-oriented functional positioning of generative artificial intelligence. It emphasizes promoting the generalization of similar-case adjudication by relying on similar-case analysis standard models constructed through deep learning, and then realizes the organic coupling between the standardized technical supply of generative artificial intelligence and the quality needs of the integrated construction of smart courts, thus becoming the core goal of boosting judicial trials. Smart courts and even judicial AI usually mean a logical shift from text to code. The boosting effect of generative artificial intelligence needs to clarify the technical quality and efficiency goals of judicial trial application aimed at effectively adjusting the algorithmic referee power structure on the premise of equal emphasis on algorithm security and judicial quality and efficiency, and further try to explore and establish the implementation path of standardized procedures, refined argumentation, automated reasoning, and predictable results for similar-case adjudication in accordance with the task guidance of “focusing on solving the deep-seated problems affecting judicial justice and restricting judicial capacity”. Based on this, the similar-case adjudication dimension of generative artificial intelligence boosting judicial trials aims to alleviate the contradiction between integrated judicature and mechanical judicature, and try to avoid the security risks involved in “human dependence on technology in specific judicial decision-making and judicial organs’ dependence on technology companies in system construction”. It is necessary to promote the construction of intelligent auxiliary case-handling systems that “improve functions such as similar-case pushing, result comparison, data analysis, and defect prompting” from the two aspects of standardized identification of similar-case facts and standardized application of similar-case legal bases, based on the similar structural characteristics of the same type of cases.
2. Starting-point-based Regulation of Generative Artificial Intelligence Boosting Technical Support under Port-oriented Positioning
Technical support is the logical starting point for generative artificial intelligence to boost judicial trials under the guidance of high-tech characteristics, aiming to highlight data and algorithm-driven modernization of the trial system and trial capacity. This type of starting-point-based regulation intends to clarify the regulatory strategies for port-based integration of data processing and port-based matching of text generation around the direction of generative artificial intelligence technology promoting the transformation of knowledge generation paradigms supported by “computing power”, thereby providing necessary material support for the digital reform of judicial trials.
2.1 Regulatory Strategies for Port-based Integration of Data Processing
Port-based integration of data processing, as the carrier of the data processing capability of generative artificial intelligence, is often determined by the relevant algorithm architecture and computing power performance of LLMs. With the continuous development of LLM technology, it has become possible to provide necessary technical support for the goal of judicial modernization through data sharing mechanisms that promote the free flow of data, thus driving breakthrough progress in judicial reform. The regulatory strategy for this type of port-based integration needs to be carried out around two types of regulatory elements: training data input and result output, based on the consideration that “algorithmic automated decision-making is a completely programming-based ‘data input - result output’ decision-making”.
On the one hand, in terms of expanding the Adaptability of Training Data Input. Adaptability is a procedural evaluation index for the performance of data processing capabilities. As the prerequisite for improving the performance of intelligent technology, it mainly refers to the matching degree between data and LLMs during the training process. The increasingly prominent phenomenon of decentralized social organization forms in the process of “scientific and technological progress driving judicial big data to subvert judicial small data” makes the digital and intelligent development of legal services often face various uncertainties and complex situations. Therefore, generative artificial intelligence boosting judicial trials often emphasizes the optimization of rational judicial justice. The involved regulatory elements of training data input need to highlight the expansion of the adaptability of relevant data training, especially strengthening the ability to analyze and process unstructured legal text language data, and further establish a precise, effective, and universally applicable standardized system for similar-case adjudication by relying on various referee case databases. Based on this, during the judicial application training of LLMs, algorithm optimization such as Bayesian optimization, unsupervised learning, and gradient descent can be attempted to efficiently process various structured and unstructured referee data and realize abstract generalization. It often emphasizes relying on behavioral adjustment capabilities such as learning, selection, deduction, induction, and correction based on technical characteristics such as domain limitation, self-adaptation, data pre-positioning, and algorithm dependence, and sets the data resources for LLM training by accurately extracting the corresponding plot descriptions and language expressions in judgment documents and describing various complex logical relationships, realizing the all-round decoding of legal text language element data. Furthermore, by using its excellent training data adaptability to process human natural language, a social production labor intermediary, LLMs can further expand the functional feasibility of generative artificial intelligence in boosting judicial trials in terms of the generalization of similar-case adjudication standards.
On the other hand, in terms of enhancing the accuracy of result outputs, accuracy serves as an outcome-based evaluation metric for data processing capabilities. As the ultimate goal of improving intelligent technology performance, it primarily pertains to the assessment results of intelligent algorithm performance and their practical application effectiveness. Alongside the technological revolution brought about by generative artificial intelligence in autonomous human-machine interaction, linguistic logic comprehension, and digital content production, it is progressively influencing and reshaping the production methods and application pathways of human knowledge. Consequently, generative artificial intelligence-assisted judicial adjudication is often entrusted with the mission of enhancing judicial efficiency, safeguarding judicial fairness, optimizing the adjudication system, and improving the quality and effectiveness of trials. The regulatory element for result outputs should therefore emphasize the exponential increase in the scale of data training for large models, supported by neural network architectures and big data technologies. Through the scientific control of data training content and formats, it aims to precisely analyze and extract various adjudicative elements of similar cases from big adjudication data. Accordingly, in the training process for judicial applications, large language models may attempt to base their efforts on the consideration that “the development level of judicial artificial intelligence can be regarded as the primary determinant of its collaboration scope and depth with judges.” By employing methods such as contextual learning and chain-of-thought reasoning, they can achieve the automatic extraction, annotation, and transformation of adjudication data. This approach often follows the reform strategy of “integrating the deepening of judicial system reforms with the application of modern technology.” It emphasizes conducting pre-training with diversified and multilingual text data during the preparation and debugging phases of large language model integration. By adjusting the length and frequency limits of relevant information inputs, it effectively enhances the comprehension and expressive capabilities of large language models in handling complex semantic information. Consequently, large language models can leverage their precise result output capabilities to overcome data training challenges caused by semantic ambiguity. While effectively improving the accuracy of adjudication data result outputs, they can also profoundly expand the technical reliability of generative artificial intelligence in assisting judicial adjudication by universalizing adjudicative standards for similar cases.
2.2. Regulatory Strategies for Port-based Matching in Text Generation
Port-based matching in text generation, as a carrier for the text generation capabilities of generative artificial intelligence, often serves as a critical benchmark for distinguishing between “generative” and “deterministic” intelligent technologies. It aims to ground itself in the typological data identification, deconstruction, analysis, and storage of large-scale precedent judgments, relying on a domain-specific large language model for adjudication comparison formed through extensive data training. By continuously training on data and regularly generating new data in a cyclical manner, it achieves the intelligent generation of target text content. Such regulatory strategies for port-based matching should be developed based on considerations of the quality and practicality of the generated texts in the context of generative artificial intelligence-assisted judicial adjudication, focusing specifically on two regulatory elements: the content and semantics of the generated texts.
On the one hand, regarding the improvement of content coherence in generated texts, content coherence serves as a formal evaluation metric for the performance level of text generation capabilities. It emphasizes that the generated texts must demonstrate consistency and continuity in structure, logic, and semantics, thereby ensuring the readability and comprehensibility of the corresponding textual information. With the continuous integration of legal language and machine language on judicial big data platforms, generative artificial intelligence often relies on advancements in large model text processing technology supported by recurrent neural networks to assist judicial adjudication. The regulatory element for the content of generated texts should emphasize improving the content coherence of the generated texts by optimizing the prediction and arrangement of word meanings through methods such as “identifying and learning from existing data and processing input conditional information as required.” Accordingly, in the process of generating judicial documents, large language models should focus on the gradual, iterative technical characteristics of the generated text words rooted in previously generated content. Through recurrent neural networks, they should attempt to process the structural, logical, and semantic elements of the generated texts, autonomously completing the prediction and arrangement of the next word or sentence segment. This continuous process aims to approach the technical goal of automatically generating judicial documents that are complete in format, standardized in structure, fluent in language, and fully reasoned. This often highlights the use of case-specific knowledge graphs to accumulate and store large-scale knowledge and typological information about the constituent elements of specific cases. By “selecting the most matching ‘word chain’ based on the extensive word associations established in the large language model,” judicial documents are generated, achieving the regular combination and formatted presentation of various elements in the generated texts. Consequently, leveraging its coherent text generation capabilities, large language models can deepen artificial intelligence’s cognition, understanding, and even proactive expansion of human knowledge, enhancing the content readability of generative artificial intelligence in assisting judicial adjudication through the intelligent generation of judicial documents.
On the other hand, regarding the enhancement of semantic accuracy in generated texts, semantic accuracy serves as a substantive evaluation metric for the performance level of text generation capabilities. It emphasizes that the generated texts must accurately convey the semantic content of the input intent or information, thereby ensuring the reliability and effectiveness of the corresponding textual information. Under the guidance of judicial justice, generative artificial intelligence should not only promote formal consistency in similar cases but also strengthen substantive justice in individual cases. This is particularly important in judicial documents, where it is necessary to provide sufficient reasoning and highlight the specificity of legal application in individual cases. However, generative artificial intelligence-assisted judicial adjudication is often limited by the semantic rigidity and templated nature of automatically generated formatted texts. The regulatory element for the semantics of generated texts should emphasize a context-specific, problem-oriented interpretation of semantic content that integrates emotion, reason, and law, fully meeting the intelligent and flexible generation needs of judicial documents. Accordingly, in the process of generating judicial documents, large language models should fully consider the practical effectiveness demonstrated by deep learning and generative adversarial networks in addressing cross-modal tasks. Through the cross-modal data conversion capabilities of transformers, they should facilitate the automatic and precise translation between computational language and legal text language. To effectively guard against the inevitable conflict between “the judicial scene’s pursuit of stability and the uncertainty of technology,” large language models must also rely on their “semantic feature extraction capabilities, long-distance feature capture capabilities, comprehensive task feature extraction capabilities, parallel computing capabilities, and operational efficiency” to enhance the semantic accuracy of the generated judicial document texts as much as possible. Consequently, leveraging its accurate text generation capabilities, large language models can profoundly expand the reliability of generative artificial intelligence in assisting judicial adjudication through the intelligent generation of judicial documents.
3. Means-Based Regulation of Generative Artificial Intelligence in Document Processing through Automated Positioning
Document processing serves as an auxiliary means for generative artificial intelligence to support judicial adjudication under the guidance of high-efficiency characteristics, aiming to “promote the dual advancement of judicial reform and intelligent, informatized development in people’s courts.” This means-based regulation focuses on the formatting and element-specific requirements of judicial documents, emphasizing the practical goal of generative artificial intelligence in processing information input and output, decoding and translating computer language into textual language, and clarifying its role in “assisting humans to better understand the world and establish close connections with us.” It is necessary to clarify regulatory strategies for the automated decoding of document elements and the automated drafting of document content, striving to more efficiently convert the computational language output from adjudication data processing into the required legal textual language.  
3.1 Regulatory Strategies for the Automated Decoding of Document Elements 
The automated decoding of document elements is grounded in the technical features of natural language processing technology, which “enables effective communication between humans and computers using natural language.” It aims to accurately extract corresponding semantic elements from the legal language of judicial documents by more precisely understanding the semantics and context of the involved natural language, automatically retrieving vast amounts of judicial document data, and autonomously parsing the textual element characteristics of judicial documents. This ensures the standardization and adaptability of the automated decoding of judicial document elements. Such automated decoding regulatory strategies should be developed from semantic and relational perspectives, focusing specifically on two regulatory elements: the semantics within judicial document elements and the relationships between these elements.
On the one hand, regarding the clarification of semantic understanding within judicial document elements, this often points to the consensus-based value pursuit of deepening comprehensive reforms of the judicial system and effectively meeting diverse judicial needs. Generative artificial intelligence clarifies the legal semantic information embedded in elements such as legal application, fact-finding, and reasoning during the decoding of judicial documents. While recognizing that intelligent digital content editing technology possesses preliminary capabilities in semantic understanding and attribute manipulation, it further emphasizes the new technologies of generative artificial intelligence in “deep intelligence” areas such as concept abstraction and reasoning decision-making, as well as in “knowledge appropriation.” The regulatory element for the internal semantics of judicial document elements should address the complexity of legal provisions, the diversity of case facts, and the insufficiency of reasoning in judicial practice, clarifying the potential responses of generative artificial intelligence in understanding the internal semantics of these elements. Accordingly, this type of decoding of judicial document elements by generative artificial intelligence should leverage the technical advantages of its neural convolutional models, which are more complex than traditional algorithmic models and enable deeper analysis of various data elements. By employing new natural language processing technologies to transform complex legal textual language into standardized data language, it achieves the automatic, accurate, and efficient identification, extraction, and analysis of the relevant judicial document elements, systematically interpreting the legal semantic information within these elements. Additionally, it emphasizes the further use of deep learning models such as recurrent neural networks and large language transformers to automatically annotate and structurally process the corresponding judicial document element data, achieving formatted and textual presentations of decoding conclusions. Consequently, this type of decoding by generative artificial intelligence helps bridge the inherent gap between machine code and legal text, eliminating the technical barriers and cognitive limitations that judges face in learning code translation, and effectively enhancing the verifiability and acceptability of the decoding conclusions for judicial document elements.
On the other hand, regarding the clarification of relationships between judicial document elements, this often points to the collaborative, parallel, restrictive, causal, and other relationships among various judicial document elements clarified by generative artificial intelligence during the decoding process, as well as the influence proportions of different elements in generating adjudication conclusions. The regulatory element for the relationships between judicial document elements should be based on the systematic argumentation requirements of generative artificial intelligence in the construction of smart courts, which emphasize autonomous interpretability. It should further highlight how generative artificial intelligence goes beyond traditional judicial artificial intelligence, characterized by various judicial document databases, to conduct in-depth and sufficient argumentative and guiding analysis of the relationships between judicial document elements. Accordingly, this type of decoding of judicial document elements by generative artificial intelligence should rely on the technical feature of “calculating the probability of collocation distribution between words with massive computational power.” For element-based and modularized judicial documents, it should construct a knowledge genealogy of the corresponding judicial document elements through the full and effective application of intelligent algorithms. Given that the professionalism and rigor required for the content of judicial documents far exceed the modeling standards of general-purpose large language models, it emphasizes the need to establish and improve necessary case screening mechanisms based on the requirements for providing reliable and orderly algorithmic data learning sets, thereby more objectively and comprehensively presenting the full picture of relevant judicial practice. Additionally, this type of decoding by generative artificial intelligence must provide vast and sufficient training data for the relational analysis of judicial document elements. Generative large language models can thereby quickly retrieve knowledge fragments based on fragmented information materials and engage in autonomous associative reasoning and even knowledge integration. This ensures the utilization of such functional advantages to assist judges in making more objective and rational judgments regarding various elements during the generation of judicial documents.  
3.2 Regulatory Strategies for the Automated Drafting of Document Content 
The automated drafting of document content aims to follow the goal that “generating text is the core task in the legal field.” It employs technical means such as natural language processing and machine learning to precisely extract and summarize relevant textual elements from the database of adjudication argumentation corpora, effectively improving the quality of judicial documents. This often “achieves interactive adjudication through deep learning and mining of data, as well as the establishment of human-machine collaboration models,” attempting to minimize “interference from ‘personal favoritism’ and ‘relationships.’” Such automated drafting regulatory strategies must go beyond the mechanical extraction of document elements, establish richer document formats, and focus specifically on two regulatory elements: the quality and format of the drafted content of judicial documents.  
On the one hand, regarding the improvement of the quality of drafted content in judicial documents, this often points to the continuous enhancement by generative artificial intelligence of the accuracy, readability, and coherence of the set textual content within specific adjudication contexts during the drafting process. From the perspective of judicial artificial intelligence as a stable, standardized, and predictable form of intelligent technology-driven justice, it emphasizes more the functional goal of stabilizing adjudication expectations and unifying adjudication standards. The regulatory element for the quality of drafted content in judicial documents should effectively address the complex structural, semantic, and logical presentation requirements of judicial documents by strengthening the accuracy and credibility of the generated text, highlighting the convenience and practicality of the judicial documents drafted by generative artificial intelligence. Accordingly, this type of drafting of judicial document content by generative artificial intelligence should align with the current technological approach of deep learning and generative artificial intelligence, which is well-suited to the practical needs of judicial artificial intelligence in assisting judicial document writing. It should be based on the intelligently decoded modular judicial document elements to achieve standardized and precise expression of document content. Additionally, in response to technical risks such as the weakening of control over artificial intelligence due to algorithmic black boxes, this type of drafting by generative artificial intelligence must emphasize effectively improving the alignment between the expression of judicial document content and case facts, thereby further highlighting the contextual and systematic applicability of the drafted judicial document content.  
On the other hand, regarding the enrichment of the format of drafted content in judicial documents, this often points to the increasingly flexible demonstration by generative artificial intelligence of the diversity and multidimensionality of the set textual content in terms of style, form, and expression techniques during the drafting process. Based on the fundamental stance of judicial people-centeredness, it emphasizes more the need to start from the different types of cases involved and the practical needs of judges. The regulatory element for the format of drafted content in judicial documents should follow multidimensional approaches, such as optimizing database operation modes, increasing non-formal legal sources in databases, and clarifying specific situations for extra-legal argumentation. It should attempt to provide personalized judicial document generation solutions while ensuring consistency in standardized formats. Accordingly, this type of drafting of judicial document content by generative artificial intelligence should be conducted under the premise of effectively preventing data security risks such as data loss of control and data leakage. It should focus on the technical feature that “large artificial intelligence models can already integrate multimodal data and build bridges between different data types,” relying on new artificial intelligence technologies such as variational autoencoders and transformer cross-modal data conversion to standardize and structurally decode the personalized textual content characteristics presented by various modular judicial document elements. Additionally, the intelligently generated content of judicial documents often extends beyond textual forms, encompassing diverse expressions such as graphics, audio and video, metaverse digital content, and digital humans. This type of drafting by generative artificial intelligence should emphasize promoting the drafted content of various judicial documents to achieve convenient and orderly format conversions based on the practical needs of adjudication work.
4. Goal-Based Regulation of Generative Artificial Intelligence in Promoting the Adjudication of Similar Cases under a Standardization Framework
The adjudication of similar cases is the core objective of generative artificial intelligence in supporting judicial adjudication under the guidance of high-quality characteristics, aimed at promoting the establishment of relatively unified standards for adjudicating similar cases and even for legal application. This goal-based regulation seeks to follow the logical guidance of the generality of law regarding treating similar cases alike. It is grounded in the ability of generative artificial intelligence to “formulate problems and their solutions into forms and thought processes amenable to effective information processing,” attempting to advance objective decision-making concerning the knowledge genealogy of adjudication data for similar cases and even the standards for adjudicating such cases. It is necessary to clarify regulatory strategies for the standardized determination of facts in similar cases and the standardized application of legal bases in similar cases, specifically targeting activities such as the intelligent identification by computers of factual and legal disputes in litigation documents, thereby highlighting the corresponding role in universalizing adjudication standards.
4.1 Regulatory Strategies for the Standardized Determination of Facts in Similar Cases
The standardized determination of facts in similar cases aims to be grounded in generative artificial intelligence’s analysis of data related to the fact-finding process in adjudication, achieved “through analogical activities, thereby establishing a multidimensional and objective cognitive foundation.” It attempts to extract universal application standards present in the judicial determination of facts. These universal application standards often serve as the logical prerequisite for the intelligent recommendation of similar cases. Following the value orientation of adjudicating similar cases alike, the “principle of universalizability and repeatability” for standards determining facts in similar cases should be “equivalent.” This standardized determination regulatory strategy must emphasize creating a relatively healthy and sustainable algorithmic learning environment based on high-quality data screening, focusing specifically on two regulatory elements: the determination of basic facts and the determination of derived facts in similar cases.
On the one hand, regarding establishing the technical pathway for determining basic facts in similar cases, basic facts refer to the primary case facts that exist independently without relying on other facts. They often form the basis for analysis, reasoning, or judgment in the judicial process. These facts are directly related to the legal nature, composition of liability, and degree of liability in corresponding cases, playing a substantive and guiding role in the trial of similar cases, and ultimately affecting the legal application and even the outcome of adjudication in such cases. However, in judicial practice, the description of basic adjudicative facts, especially when judges compare the specific circumstances of a case under trial with guiding similar cases, often employs colloquial expressions. The regulatory element for determining basic facts in similar cases must therefore break through the application bottlenecks of traditional artificial intelligence natural language processing technology. It should attempt to establish a technical response pathway between colloquial expression and professional expression of legal terminology to complete the analysis and processing of basic factual data more accurately and professionally. Accordingly, in this application scenario, generative artificial intelligence should rely on the technical environment where AI for similar cases and judgments can achieve deep imitation of individual case facts and rule systems, along with formal logic transfer. It should attempt to use new natural language processing technologies capable of annotating basic facts and automatically calculating basic fact points, promoting the standardized, precise, and even professional parsing of basic facts. This often emphasizes generating a standardized knowledge genealogy for determining basic facts in similar cases through the induction and refinement of the characteristics of such facts. It aims to “automatically extract and deeply mine the summaries and basic factual bases of cases from case files, intelligently identify and match the causes of action in individual cases,” thereby effectively enhancing the accuracy of judges in determining the basic facts of relevant similar cases.
On the other hand, regarding overcoming the technical bottlenecks in determining derived facts in similar cases, derived facts refer to conclusions deduced based on basic facts, combined with relevant experience and knowledge through reasoning, induction, or interpretation. They often “assist judges in constraining personal preferences through the process of making tacit knowledge explicit, justifying their legitimacy through the commonality of experiential knowledge.” They play an important role in deepening (free heart-proof), thus influencing the legal application and adjudication outcomes in similar cases to a certain extent. The regulatory element for determining derived facts in similar cases must break through the technical bottleneck of traditional computing power, which can only achieve simple factual comparisons. It should achieve complex induction and analysis of the interrelationships and modes of interaction between case facts, “based on the extraction and comparison of factual and normative elements in cases,” thereby maximizing the avoidance of risks or even errors induced by incomplete factual induction. Accordingly, in this application scenario, generative artificial intelligence should be grounded in the premise that “big data conditions provide a factual basis for anticipating outcomes.” It should attempt to use new natural language processing technologies capable of extracting and separating derived facts from case facts and performing feature annotation, promoting the “integration of rule refinement and factual comparison.” This often emphasizes relying on data systems for similar cases to depict the complete process of case application through qualitative and quantitative analysis of the interconnections and modes of interaction among various elements for determining derived facts, constructing universal standards for determining derived facts in similar cases. It also involves targeted training of the autonomous evaluation and comparative thinking of generative artificial intelligence, providing more reasonable and standardized interpretative path guidance for generative artificial intelligence in supporting judicial adjudication. This aims to promote the improvement of the quality and effectiveness of judicial reasoning while maximizing and optimizing the exertion of its social functions based on the overall lower operational costs of the court system.
4.2 Regulatory Strategies for the Standardized Application of Legal Bases in Similar Cases
The standardized application of legal bases in similar cases aims to be grounded in the “hypothesis of predicting the same outcome as the most similar case,” relying on the big data modeling capability of generative artificial intelligence for similar cases to extract universal application standards present in the application of judicial adjudication bases. It often highlights the formation of a guiding normative effect through the comparison and refinement of cases of the same type, applicable to future similar cases and exerting possible binding force. This standardized application regulatory strategy must emphasize the unified and accurate analysis and processing of relevant data during the application of corresponding adjudicative norms, focusing specifically on two regulatory elements: the uniformity and the accuracy in applying legal bases for similar cases.
On the one hand, regarding ensuring the uniformity in applying legal bases for similar cases, uniformity, as a formal evaluative requirement for applying legal bases in similar cases, is the declarative expression of adjudicating similar cases alike in terms of internal constitution. Based on affirming the paradigm for applying legal bases in similar cases, which holds that “similar cases are a collection of cases governed by the same adjudicative rule,” it specifically refers to the common characteristics of data from similar cases that judges summarize and extract through generative artificial intelligence algorithms. The regulatory element for uniformity in applying legal bases for similar cases must promote the achievement of uniform outcomes in adjudicating similar cases and maintain the credibility and authority of the law by ensuring uniformity in the formal standards for applying such bases. Accordingly, in this application scenario, generative artificial intelligence should rely on the technical environment where “algorithms will, to a certain extent, restrain the arbitrariness of judicial personnel, increase the uniformity of legal application, and reduce bias.” Through generative artificial intelligence algorithms, it should conduct correlation analysis and similarity judgments regarding the involved adjudication data, involving continuous cross-referencing between case facts and adjudicative rules. It should also qualitatively or quantitatively depict the interconnections and modes of interaction among the relevant application elements, attempting to construct uniform standards for applying legal bases in similar cases. Furthermore, it emphasizes inputting adjudication data fed back from judicial practice into the generative artificial intelligence system for cyclic training, further confirming and solidifying the corresponding application standards. This aims to strengthen judges’ consistent understanding of the application of adjudicative norms and consolidate their conviction regarding the value goal of legal certainty.
On the other hand, regarding emphasizing the accuracy in applying legal bases for similar cases, accuracy, as a substantive evaluative requirement for applying legal bases in similar cases, is the binding expression of adjudicating similar cases alike. It specifically refers to the degree of difference and the influence of data from similar cases that judges compare through generative artificial intelligence algorithms. Following technical paths such as natural language processing, graph neural networks, and deep convolutional neural networks, it promotes the unification of outcome standards for applying legal bases in similar cases. The regulatory element for accuracy in applying legal bases for similar cases must emphasize more precisely and appropriately “fixing and refining the adjudicative rules for similar cases,” thereby obtaining, summarizing, and even justifying the legality and legitimacy of applying the ought-to-be norms in judicial adjudication, guiding judges to clarify the scope and conditions for applying the corresponding adjudicative norms. Accordingly, in this application scenario, generative artificial intelligence should be grounded in its relatively coherent cognitive logic, relatively autonomous dialogic logic, and relatively self-consistent reflective logic to accurately depict the judge profiles and case profiles involved in the application of the bases. It should attempt to improve the accuracy in applying legal bases for similar cases through profile comparison. Furthermore, even though generative artificial intelligence supporting judicial adjudication gradually strengthens the objectivity of fact-finding and reduces emotional misjudgments caused by subjective fact-finding, reliability testing in judicial adjudication still needs to be specifically advanced by relying on various public content evaluations at the subjective level. Therefore, it is necessary to emphasize further preventing the universalized standards for applying legal bases in similar cases from generating a decisive dominant position at the methodological level, avoiding letting this mechanical subsumption model absolutely influence judicial adjudication. This aims to maximize the bridging of potential conflicts between formal justice and substantive justice involved in adjudicating similar cases alike.
V. Conclusion

Generative artificial intelligence supporting judicial adjudication is a key aspect of advancing judicial reform guided by the new development philosophy. Clarifying its regulatory rationale helps avoid the disorder in human-computer interaction that may be induced by vague positioning of clues. The functional positioning of portability, automation, and standardization respectively represents the functional expression of generative artificial intelligence adapting to the characteristics of high technology, high efficiency, and high quality, thereby defining the operational dimensions of generative artificial intelligence in supporting judicial adjudication. The technical support achieved through the port-based integration and matching of corresponding data processing and text generation is the logical starting point for supporting judicial adjudication under the portability framework. The document processing achieved through the automated decoding and drafting of corresponding elements and content of adjudication documents is the auxiliary means for supporting judicial adjudication under the automation framework. The adjudication of similar cases achieved through the standardized determination and application of corresponding facts and legal bases for similar cases is the core objective for supporting judicial adjudication under the standardization framework. These aim respectively to achieve the organic coupling between the portability, automation, and standardization technological supply of generative artificial intelligence and the technology, efficiency, and quality requirements of the integrated construction of smart courts. Based on this, by systematizing the regulatory strategies for generative artificial intelligence supporting judicial adjudication, a framework of behavioral constraint guidance serving as an examination vehicle can be provided for subsequent research on the application patterns of generative artificial intelligence in supporting judicial adjudication. This attempts to clarify necessary practical samples for research on the new functions of digital administrative law as characterized by the effectiveness of emerging digital technology intervening in the construction of smart courts.

Originally published in Administrative Law Studies, Issue 5, 2025. Reprinted with permission from the WeChat public account “Administrative Law Studies Editorial Department”.