[author]Xu Shuhao

[content]

The Reason Model of Judicial Artificial Intelligence and Its Functional Limits

Xu Shuhao

Assistant Professor, KoGuan School of Law, Shanghai Jiao Tong University

Abstract: Al learns to simulate judicial reasoning by analyzing key factual elements in case corpora. This learning process can be described through a reason model framework, which involves identifying reasons, assessing weight, and consolidating reasons to establish a comprehensive “priority ordering of reasons. This priority ordering helps judicial models reduce errors. However, since this reason model relies solely on case corpora rather than legal theories, including jurisprudential interpretation, it funda-mentally operates as trial-and-error exploration rather than theory-guided learning. Consequently, two structural limitations emerge. Firstly, when legal theories undergo breakthroughs that aren’t promptly applied to judicial practice, judicial models fail to update accordingly. Secondly, when laws undergo complex changes through amendments or repeals, Al struggles to automatically recognize systemic implications. These structural flaws demonstrate that the primary approach to Al adaptation remains human-led development of legal theories, followed by the generation of new priority ordering through case corpora.

Keywords: Artificial Intelligence; Digital Justice; Deep Learning; Reason Model; Automatic Decision-Making

Expanding the judicial application of artificial intelligence is a long-term strategic plan for China. As early as 2017, the New Generation Artificial Intelligence Development Plan (Guo Fa [2017] No. 35) issued by the State Council explicitly advocated deeply integrating AI technology into judicial scenarios to empower case-handling activities. In recent years, with the widespread application of intelligent case-handling assistance models nationwide, controversies regarding their functional boundaries and the persuasiveness of their conclusions have become increasingly prominent in both theoretical and practical circles. It is difficult for people to reach a consensus on what case-handling models can and cannot do. The root of this controversy lies in the methodological dilemma regarding the explanation of the internal process of AI decision-making. For instance, some views hold that judicial decision-making is a reasoning process based on reasons, whereas machine decision-making relies on probability calculation rather than reasons, which fundamentally questions the explainability of the AI decision-making process.
Whether AI decision-making relies on reasons is directly related to whether it can be explained and trusted, making the construction of an explanation system for judicial AI based on reasons an important theoretical task. To this end, this article attempts to clarify the internal mechanism by which AI simulates the identification and application of judgment reasons by microscopically examining deployed mature case-handling models and utilizing the analytical framework of the "Reason Model."
AI completes the discovery of reasons, the determination of reason weights, and the consolidation of reasons by learning key factual elements in the case corpus, forming a "priority order of reasons" to reduce the decision-making error rate. This clarification not only provides theoretical support for the explainability of AI judicial decisions but also helps to clarify its structural defects, thereby providing directional guidance for improving judicial AI.
1. The Decision-Making Mechanism of Judicial AI: Taking the Assisted Case-Handling Model as an Example
To deeply understand judicial AI, it is indispensable to conduct research around deployed mature models. Since 2022, Shanghai procuratorial organs have adopted a project-based operation mode of "one crime, one model," organizing prosecutorial personnel to dock with technology companies to systematically integrate similar cases within specific crime categories. They built a specialized case judicial database and, on this basis, established intelligent case-handling models for those crimes, typically exemplified by the "Dangerous Driving Case Full-Process Online Handling" model. Such models can perform various functions, including case card back-filling, comparison of the three major documents (police request, indictment, judgment), trial supervision, and document generation. Using the intelligent assisted case-handling model developed by the Shanghai People's Procuratorate as a sample for observation, and dissecting the specific process of AI embedding into judicial case-handling, helps to grasp the operational mechanism of judicial AI at a micro level. Due to space limitations, this article focuses its research on the document generation link, where automated conviction and sentencing are the main content.
1.1 Extraction and Classification of Case Fact Elements
As of February 2025, Shanghai procuratorial organs have completed the integration and deployment of 8 assisted case-handling models with existing procuratorial business application systems, covering 70% of the city's total case volume. Except for statutory exceptions, relevant legal documents for these cases must be produced through the assisted case-handling model. The operational flow for case handlers to generate documents relying on the intelligent assisted case-handling system can be roughly divided into four steps: selection of intelligent assisted case-handling mode; parsing of the prosecution opinion letter (police report); setting of fact/plot elements; and generation of assisted decisions.
After entering the system's built-in intelligent assistance module, case handlers can exercise the right to refuse automated decisions under statutory conditions. For example, if a case handler believes there are grounds for "non-prosecution due to insufficient evidence" (Article 175 of the Criminal Procedure Law), "absolute non-prosecution" (Article 177, Paragraph 1), or "conditional non-prosecution" (Article 282), they have the right to manually adjust the case handling settings and switch to the traditional mode. Once the intelligent assistance function is selected, the system will automatically perform full-element structured processing on the prosecution opinion letter created by the investigative organ and imported into the system, automatically identifying and extracting the fact/plot element data of the case.
Case fact/plot elements play a foundational role in the automated decision-making process. The term "element" specifically refers to fact plots or factual statements that can influence legal conclusions; it is the smallest semantic unit and meaning fragment that AI can identify. In the machine-assisted decision stage, whether to recommend prosecution, apply probation or imprisonment, how to set the sentence term, and which historical judicial data to push, all depend on the specific elements contained in the case. Case fact elements can be divided into three categories: first, crime constitution fact elements; second, general sentencing plot elements; and third, special plot elements. Among them, the crime constitution fact elements are set based on the constitutive facts of the crime as clarified by the Criminal Law and relevant judicial interpretations. General sentencing plots refer to plots that affect sentencing in any type of crime, including the suspect's age, type of criminal record, type of bad conduct, surrender, confession, plea of guilt and acceptance of punishment, and recidivism. Special sentencing plot elements are set specifically according to the unique attributes of different crime types. If the case handler has no objection to the plot elements parsed by the system, they can directly enter the assisted decision stage; naturally, the case handler also has the right to modify the conviction and sentencing elements parsed by the system, or manually enter new element content based on new circumstances obtained through self-investigation or supplementary investigation.
In the assisted decision stage, the decision model will automatically start the calculation program and generate "suggested conclusions" regarding whether to prosecute, whether to apply probation, and specific prison terms. At this point, the system presents a single operation interface composed of two functional sub-interfaces: one is the "Case Review Conclusion" sub-interface, containing two columns: "Prosecute/Do Not Prosecute" and "Imprisonment/Probation," where the system pre-fills recommended conclusions based on automated review results; the second is the "Sentencing Calculation" sub-interface, which displays various sentencing factors. Each factor is presented in the form of a magnitude bar, where the position of the bar intuitively characterizes the severity level of the corresponding sentencing factor. The system automatically generates the configuration ratio of the magnitude bars by capturing and analyzing the content of the prosecution opinion letter, and completes the quantitative calculation of the sentence based on this ratio.
1.2 Automated Decision-Making Based on Elements
The case-handling model is a functional hybrid, and not all its components rely on AI technology support. The core part of the procuratorial case-handling model, which is also the key operational mechanism of judicial AI, is the algorithm system applied from the setting of fact plot elements to the assisted decision. However, the sentencing calculation in this link relies on explicit calculation formulas. These formulas are built into the system as linear functions and do not fall within the scope of judicial AI in the narrow sense. Regarding how much a sentencing plot can increase or decrease a sentence, the calculation rules are usually explicitly stipulated in normative documents such as the Supreme People's Procuratorate's "Guiding Opinions on Sentencing" and various local "Detailed Rules for Implementation of Guiding Opinions on Sentencing." For example, the drug case sentencing model explicitly uses the quantity of various types of drugs converted into heroin multiplied by a certain coefficient to calculate the sentence; the sentencing model for crimes of aiding information network criminal activities calculates the sentence by multiplying the number of involved SIM cards and bank cards by corresponding coefficients. When two or more sentencing plots coexist, there are also explicit formulas understood by judicial personnel for calculating the sentence after the plots are superimposed. In principle, the system only needs to program these calculation formulas into the computer to achieve automated calculation, without any complex or unexplainable aspects.
The module that truly requires the application of AI technology is the binary classification model specifically for the "Prosecute/Do Not Prosecute" and "Imprisonment/Probation" decisions, after stripping away the sentencing calculation part. This model only outputs a binary conclusion of "Yes" or "No" for specific questions. In practice, except for absolute non-prosecution, non-prosecution due to insufficient evidence, and conditional non-prosecution cases to which this model does not apply, relative non-prosecution cases usually require the activation of the case-handling model for assisted decision-making; and the choice between relative non-prosecution and prosecution belongs to discretionary judgment, involving uncertainty. If relatively simple judgments are involved, the assisted case-handling model sometimes uses decision tree algorithms to classify possible decisions under uncertainty. For complex decision scenarios, the assisted case-handling model needs to use deep learning for global consideration to determine the optimal set of decision weights. This involves letting AI learn the complex functional relationship between similar past judicial decisions and the influencing factors behind them.
Looking at the entire process of the case-handling model's decision-making, the current technical path adopted is: first, perform structural processing on the prosecution opinion letter, then use key field automated extraction technology to automatically identify and extract key plot element data. On the interface operated by the handler, these plot elements are distinguished into crime constitution fact elements, general sentencing plot elements, and special sentencing plot elements. These three types of elements collectively influence the decision on whether to prosecute and whether to apply effective imprisonment. The working principle of the deep learning algorithm is: technical personnel mark the "decision" parts of a large number of indictments, non-prosecution decisions, plea affidavits, and judgments for a certain type of case in advance (e.g., marking them as "Prosecute/Do Not Prosecute" or "Guilty/Not Guilty"), and then compile these judicial documents into a training set; the AI neural network continuously learns through trial and error in the training set, eventually autonomously discovering the internal correlation between combinations of case plot elements and judicial decisions. Figuratively speaking, if the process of a case handler dealing with a case is understood as a process of making an intuitive judgment on whether to prosecute based on case information and their own legal and moral cognition, then the task of the neural network is to analyze the generation mechanism of this "intuition," that is, to identify with which case information it has established an implicit and robust correlation.
2. The Dimension of Reasons in the Explainability of Judicial AI
Different AI technology paths lead to differences in the degree of explainability. For example, if the "Expert System" path of "organizing trial rules of certain specific case types into codable logical sentences" is adopted, or the "Decision Tree" algorithm (excluding ensemble learning and random forests), the corresponding reflection and explanation process is like a mathematical verification, which can be completely restored by humans. Given that the aforementioned assisted decision models have adopted deep learning algorithms based on neural networks, the key to handling the explainability issue no longer depends on the choice of algorithm type, but on how to make the model's learning process explainable.
2.1 The Unexplainability of the Model Itself
Some scholars suggest that by disclosing the algorithm source code, architecture, parameters, and weights, applicants seeking explanation can autonomously understand and even run the algorithm based on the disclosed content, thereby achieving supervision of the model. The biggest problem with this approach is that mere disclosure has almost no substantive effect on "understanding the algorithm's operational logic." Scholars have already provided strong theoretical arguments for this. In addition, corresponding evidence can be raised from the technical principle level of deep learning.
The technical foundation of deep learning is the multi-layer neural network, the basic structure of which consists of an input layer, (multiple) hidden layers, and an output layer. Each neuron in the hidden layer learns a simple function, and these simple functions combine to form a complex function. The more "layers" in the hidden layer, the deeper the neural network, and the higher the complexity of the final constructed model. If the relationship between information collected in judicial activities and judicial decisions is understood as a non-monotonic, non-linear complex functional relationship, then under the premise that the "overfitting" problem does not occur, the more complex the function learned by the neural network, the better its performance (such as prediction accuracy) may be. The question is, how exactly does a multi-layer neural network improve accuracy by establishing complex functions? To answer this, one must first grasp the technical characteristics of the "hidden layer."
In a multi-layer neural network, each hidden layer is composed of a large number of units, and each unit contains specific weight parameters. After the input value of the unit is multiplied by the weight and summed, the activation value of that hidden layer is obtained; when the activation value meets a preset threshold, the neural network passes the summation result as an output value to the next hidden layer and repeats this weighted summation process until it reaches the output layer. This structure containing multi-layer neural networks can be viewed holistically as an activation function, which handles the non-linear mapping relationship between input data and output results. The reason why neural networks can learn useful matters from received data that are difficult to be completely summarized and regulated lies in the particularity of their learning process. Deep learning relies on two basic algorithms: the backpropagation algorithm and the gradient descent algorithm. Among them, the function of the backpropagation algorithm is to assign responsibility, while the gradient descent algorithm is used to clarify the learning rule for each weight. Training a neural network requires the simultaneous application of both algorithms.
Specifically, computer engineers randomly set the number of layers in the hidden layers and the initial value of each unit's weight at the initial stage of training. The reason why initial values can be set "arbitrarily" is that during the training process, the labeled dataset will help the neural network gradually adjust these weight parameters, and backpropagation plays a key role in this. At the beginning of training, the neural network will inevitably "make mistakes." The function of the backpropagation algorithm is to "remember" the mistake and assign responsibility in reverse. If the forward operation flow of the neural network starts from the input layer, passes through all hidden layers from left to right to complete weighted summation and activation, and passes the result to the output layer; then reverse responsibility propagation is to modify the weight values of each hidden layer layer-by-layer from right to left based on the prediction error, and then judge whether the error is reduced by observing the output result of the adjusted hidden layer activation function. This process iterates continuously until the neural network can eliminate the error in each training sample, thereby obtaining a set of optimal weight parameters. The entire adjustment process is completed autonomously by the machine, that is, the process of calculating the output error and adjusting the weights via backpropagation to reduce the error. It requires the computer to rely on its computing power to automatically perform repeated trial and error and improvements. The number of calculations and complexity far exceed the processing power of the human brain. Reducing or even eliminating errors and achieving the best modeling of the dataset relies on the operation of the gradient descent algorithm. Simply put, every time the neural network uses a set of weight parameters to adapt to the dataset, a certain error is generated. Aggregating these errors yields the Sum of Squared Errors (SSE). It is generally believed that the smaller the error sum, the better the modeling effect of the function on the dataset. Theoretically, the Sum of Squared Errors sum can be expressed in three-dimensional space as a canyon-shaped error surface, where the closer to the bottom of the error surface, the smaller the error and the higher the modeling accuracy. Therefore, the deep learning process can be understood as a process of searching for the lowest point on the error surface. The so-called "gradient" refers to the degree of slope between different points on the error surface, and gradient descent is the algorithmic process of gradually approaching and finally reaching that lowest point.
Every time the machine uses backpropagation to iteratively update the weight set of the hidden layer, it can be understood as a movement attempt by the neural network on the error surface. This movement may lead to an expansion of the error or a reduction of the error, and only the adjustment that reduces the error will be retained by the system. The machine will continue to iterate and optimize on this basis until it reaches the bottom of the error surface. Some scholars have made a vivid metaphor for this process: it is like a mountaineer trapped on a mountain peak due to sudden heavy fog, while his off-road vehicle is parked in the valley. Because visibility is only a meter or two underfoot, the mountaineer can only fumble and move based on immediate judgment of the slope, eventually reaching the valley. The fumbling process of the neural network is exactly the same as that of this mountaineer. This also shows that disclosing parameters and code provides almost no substantive help for understanding the working process of the neural network. Even the computer engineer controlling the learning activity only randomly sets the initial values at the initial stage of training and is almost unable to track the autonomous iteration process of the neural network throughout. More importantly, the neural network can only carry out exploratory learning through "trial and error." This unique learning method makes it difficult for humans to imagine the machine's learning logic by analogy with their own learning patterns.
The powerful performance of neural networks stems precisely from their "black box" attribute. Or rather, it becomes a black box precisely because the process of deriving an accurate weight set through backpropagation trial-and-error and gradient descent correction contains a massive amount of calculation that humans simply cannot verify, and it is this massive amount of calculation that ensures the minimization of error. Some scholars call this phenomenon the "tension between the performance of an algorithm and its explainability."
2.2 Explanation Centered on "Reasons"
What is truly meaningful is actually the second dimension of explainability, namely the explainability of the reasons for judgment. Its core presupposition is that judicial activities need to unfold around interpreting the law and reasoning. The decision-maker should explain the case facts upon which the judicial conclusion is based, the legal articles and legal interpretation methods applied, as well as factors such as human sentiments and public order and good morals that are combined. If these matters that should be explained are reduced to the category of "reasons," what the decision-maker needs to provide are legal reasons, moral reasons, and policy reasons that support their judicial decision. A reasonable explanation is essentially an explanation based on reasons. As one scholar pointed out: regarding the explanation of technical problems, i.e., how the intelligent system makes a decision, if the reasoning process cannot be explained in detail, "at least a partial explanation and a coarse-grained summary explanation should be given, rather than offering no reason at all. It is not difficult to understand that a legal decision attached with 'reasons' is a reasonable and acceptable decision."
Many scholars criticize judicial AI, and the core reason lies in their belief that judicial AI lacks explainability at the level of reasons for judgment and does not possess the ability to convert the neural network iteration process into justification and reasoning activities. This criticism can be disassembled into two levels: First, AI lacks the thinking of reasoning, and in the application process, there is no internal activity based on logic and rationality to prove the judicial conclusion; Second, AI cannot achieve persuasion of humans by "giving reasons." Of course, these two levels are relatively independent; even if judicial AI does not actually carry out reasoning activities, it may still possess the ability to engage in persuasive activities. For example, Large Language Models (LLMs) can already generate texts, charts, etc., containing reasoning content relying on the "predict next token" algorithm, providing auxiliary support for judgment reasoning. However, critics can still claim that "predicting the next token" relies on probability calculation rather than logical reasoning, and the inability to reason remains a fundamental limitation of AI. This fundamental limitation can be further elucidated: the operating principle of machine learning algorithms does not reproduce the logical process of legal reasoning, but seeks statistical correlations between case-handling behavior data, case information data, and judicial decisions through data mining. In other words, it essentially "replaces normative judgment with probability calculation." Precisely because of this, AI cannot truly "understand" the semantic connotation of the document fields and sentences it processes, lacks the common sense and reason possessed by humans, and certainly cannot sort out the internal process of the mental activity of subsuming facts under law in the way of human cognition; under existing technical conditions, AI cannot spontaneously generate "value judgments" with subjectivity, nor does it have the ability to conduct complex interest balancing. After all, "value weighing is not calculation; it cannot be quantified, nor can it be coded."
The above argument seems reasonable at first glance. It captures key differences between machine thinking and human thinking: machines cannot understand and respond to reasons, while humans are "animals of reasons." Human decision-making relies on reasoning. The core difference between human judgment and machine judgment lies in "whether one can identify and utilize judgment reasons." Even if a machine can respond to a user's request for explanation and generate explanatory or conversational text that the audience can understand to introduce the reasons for a judicial prediction, it still cannot dissolve the essential difference between humans and machines: the reasoning process humans use when reaching a conclusion often has consistency and identity with the reasoning process they display when defending that conclusion; whereas the process by which a machine reaches a conclusion and the activity by which it defends the conclusion constitute two completely different activities. The essence of the latter is an LLM application with the target task of "how do I respond to judicial questions more like a human." If this inference holds, then it is undoubted that judicial AI lacks explainability in the sense of reasons.
Regrettably, the logic of the above inference seems to fall into the predicament of the re-description fallacy. If one merely states the technical details of the deep learning algorithm's learning target, architecture, parameters, number of hidden layers, and activation values exactly as they are, then facing the thousands of billions of calculation units in the hidden layers and its extremely complex backpropagation calculation process, evaluating whether they are similar to human rational reasoning ability has no substantive meaning, because these are mechanisms at two different levels. Just as we would not use the mechanism of billions of neurons in the human brain processing information to explain the macro thinking process of humans, similarly, we should not question the performance of the algorithm in the macro judicial scenario solely because the mathematical mechanism on which the algorithm itself relies is so unique. The key to the problem lies not in whether the machine uses code, natural language, probability-based calculation, or emotional factors to handle tasks proposed by humans, but in whether its performance in executing the task is sufficient for us to reasonably attribute a certain way of thinking or rational achievement of humans to AI, and whether it achieves practical effects identical to rational thinking.
From existing evidence, the deep learning algorithms of current intelligent assisted case-handling systems have found the functional equivalent of "reason weighing" in pattern recognition, thereby truly realizing the gradual improvement of machine judgment accuracy. Specifically, after the judicial technical team marks a large number of case fact plot elements, they systematically integrate them into a standardized corpus to train the machine's sensitivity to perceiving judicial reasons. Among them, crime constitution fact elements, general sentencing plot elements, and special sentencing plot elements are all plot "instances" extracted from legal documents. They themselves are similar to the specific reasons considered by humans during weighing. By learning the correlation between element combinations and existing judicial decisions, the machine grasps the basic cognitive structure of the weights and roles assigned to these case plots by case handlers. At this point, the case-handling model contains an extremely complex non-linear function capable of mapping the reasoning mode of judicial personnel. What this function simulates is: when numerous plots appear simultaneously in a case scenario, which plots or plot combinations will humans choose as the reasons for reaching a specific conclusion. This working mechanism of judicial AI can be defined as the "Reason Model."
Of course, the Reason Model does not claim that the decision function contained in the case-handling model is itself engaging in reason weighing, but rather believes that the decision function is responding to human demand for reason weighing. Elena believes that what deep learning algorithms achieve is not intelligence itself, but communication ability—more precisely, the ability to precisely respond to specific human communication needs obtained through training. For example, the reason why LLMs can give persuasive answers is not because they understand the essence of the questioner's question, but because they have pre-mastered the core needs of the questioner through high-intensity training. Therefore, to understand and explain judicial AI, one must first find out what unique communication needs it serves. Starting from this significance, this article does not believe that the working principle of the case-handling model is itself a Reason Model, but argues that the communication needs served by the case-handling model can be precisely characterized by the Reason Model. The core content of this communication need is: regarding the current case, what decision would decision-makers in past judicial judgments make? It is precisely due to the existence of this communication need that the communication form simulated by the case-handling model constitutes a kind of "Reason Model."
3. The Generation and Operation of the Judicial AI Reason Model
The analytical framework of the Reason Model helps to restore the process of AI using the case library to learn judgment reasons. This restoration does not stop at the "two skins" level (separation of form and substance) where the machine generates judgment reasoning texts, but substantively describes how AI "learns" to make judicial decisions. The symbolic framework used by John Horty when discussing AI models of precedent can be used exactly to explain how the case corpus provides support for judgment reasons. Describing how AI "learns" to make judicial decisions not only helps to objectively assess its actual ability in "mastering" judgment reasons but also helps to more precisely grasp the essential difference between AI's way of utilizing reasons and that of humans, thereby laying a theoretical foundation for precisely identifying defects in AI automated decision-making and carrying out targeted improvements.
3.1 Case Library Foundation and Prerequisites for Generating the Reason Model
The core working mechanism of AI, simply put, is to crack the extremely complex functional relationship between dataset input and output. For example, if the input of a certain dataset is [2,3] and the output is 6, one can easily infer that the functional relationship is multiplication. However, for datasets where the functional relationship is complex enough to exceed the human brain's direct parsing ability, AI application generally follows this working mode: first find a target dataset that may contain complex functional relationships, then use deep learning algorithms to "crack" the functional association between data items, and finally achieve prediction functionality. This working mode can also be migrated and applied to the judicial field.
Assuming that all past judicial practices—that is, the practice of all judicial officials ascertaining case facts and applying the law—can be completely collected and converted into a series of uploaded datasets, then based on the regular characteristics that judicial activities should have, it can be presumed that there is some reliable functional relationship between these datasets, only that effective cracking tools are currently lacking. AI scientists attempt to break down the problem as follows: there is some regular association between data representing case fact factors and data representing case results (such as prosecute or do not prosecute decisions, judgments, or rulings). Therefore, it is necessary to collect these two types of dataset information and encode their relationship into a specific function through machine learning. Once this function is obtained, scientists can use it to predict the legal document generation results corresponding to new cases. This is precisely the substantive connotation of the concepts of "Robot Judge" and "Robot Prosecutor" under existing technical conditions. This work relies on two fundamental premises: first, assuming that all judicial practices that have occurred can be converted into storable data; second, believing that a definite association can be formed between these data, that is, the laws contained therein are generally clear and identifiable.
Currently, the task of judicial AI is to establish exclusive datasets for a few simple case types with sufficient data volume and prioritize training, such as carrying out document structuring processing for case types like dangerous driving, intentional injury, aiding information network criminal activities, and picking quarrels and provoking trouble. To this end, a case library that meets requirements in terms of quantity needs to be built. The so-called case library can be understood as a database formed after documents are structured, and the scope of content of AI-generated documents will not exceed the information scope already present in that database. The deep learning process needs to be carried out gradually in the training set and the test set: first, technical personnel, judges, and prosecutors annotate the meaning of specific fields in the documents (automatic machine annotation can be realized later), then the machine autonomously learns the correlation between case elements and judgment results in the training set until it can match document elements and their weights with judgment results well. The function of the test set is to provide document samples that have not appeared in the training set to verify whether the machine can assign the highest confidence to the correct judgment result.
As mentioned earlier, this "element-based rather than rule-based" design scheme has become the mainstream technical path for current judicial AI. The intelligent assisted case-handling system forms decision conclusions precisely based on various identified and extracted plot elements. This technical reality means that a theoretical framework capable of accurately describing the machine learning process must satisfy two conditions: First, it should be a theory about elements rather than a theory about rules. Its core proposition is not that AI needs to achieve "direct" coding of legal rules, but that AI should use case plot elements to identify judgment reasons, and indirectly identify legal rules through judgment reasons; Second, given that the most important corpus for current machine learning is the case set, this theoretical framework must belong to the category of case theory and be able to explain how past case sets help AI identify judgment reasons. The precedent reason model proposed by John Horty fits these two conditions perfectly. This is a theoretical framework regarding how past case information constrains current judicial decision-making. According to this reason model, "the most important part of a precedent is the court's evaluation of the weights of the competing reasons presented in that case, manifested as a priority order among these reasons. Thereafter, other courts are not necessarily bound by the rules established by the precedent, or change these rules in some way, but merely follow the (reason) priority order already derived." Horty's descriptive framework centered on the priority order of reasons was initially used to explain the generation mechanism of precedent binding force, and was subsequently applied to the field of judicial AI research based on case-based reasoning (CBR), gaining extensive discussion in the cross-disciplinary research of law and AI in recent years.
In the application process of the intelligent assisted case-handling system, the internal laws contained in judicial data can be described exactly through the priority order of reasons, which aligns with the reason model theory proposed by Horty. Specific arguments will be unfolded below. It needs to be clarified that using the priority order of reasons to describe judicial data laws does not aim to elucidate the technical details of machine learning, but focuses on explaining the core issue of "how to help AI acquire judgment reasons by optimizing the case library." The goals of AI learning judgment reasons are two: one is to search for the optimal function relying on deep learning algorithms; the other is to provide the model training with a dataset that can more accurately present judicial laws by optimizing the case language material library. Although we cannot directly analyze the internal mechanisms of the model but demonstrating how the case corpus provides support for AI reasons remains an effective explanatory strategy.
3.2 Structural Classification of Case Plots and Reason Presentation Forms
Assume there is a case library with a sufficient number of documents for a certain case type, where each case is associated with one or more legal documents, and all cases contain a certain number of fact plots. It can be considered that the plots recorded in linguistic form by case handlers in all cases in the case library constitute a set F, where F is the total collection of plots for this type of case (e.g., the 102 key element data items involved in theft crimes constitute the possible total plots for that case). That is to say, for any single case appearing in the case library, the set X of all its plots is contained in F, i.e., X⊆F. It should be noted that X represents all plots of a single case, while F represents all plots covered by the case library. Both are manifested as field information extracted according to specific element classification standards, and ultimately correspond to specific case facts. Taking the crime of dangerous driving by drunk driving as an example, after establishing a case library for this type of case, to train the AI, technical personnel must let the AI know how the plots in F affect the decision of the case handler. The case handler's decision is essentially a balancing activity because the case handler will simultaneously grasp plots supporting prosecution and plots supporting non-prosecution, and needs to decide which side's plots are more dominant. Accordingly, the fact plots in any single case can be divided into two types: "supporting prosecution decision" and "supporting non-prosecution decision."
Let Fa denote the total plots supporting prosecution, i.e., Fa={f1a,f2a,f3a,...,fma}; let Fb denote the total plots not supporting prosecution, i.e., Fb={f1b,f2b,f3b,...,fmb}. Fa and Fb cover all possible fact plots for this type of case, including those facts that do not appear in the current case but belong to the potential scope of dangerous driving cases. Although neutral plots exist in practice that neither support prosecution nor support non-prosecution, based on the need for argumentative focus, they are ignored here for the time being. Ideally, the total plot set F of the case library appears as the union of Fa and Fb, i.e., F=Fa∪Fb. Extracting any single case X1 from the case library, this case typically contains both plots supporting prosecution and plots not supporting prosecution. For example, X1={f1a,f2a,f1b,f2b}. This means that in this drunk driving case X1, there are 4 plots significantly related to whether to prosecute, with fact plots supporting prosecution and not supporting prosecution each accounting for 2. For ease of understanding, "f1a,f2a" can be concretized as "blood alcohol content reaching 80mg/100ml or more" and "suspect intentionally drank alcohol before blood sample extraction"; "f1b,f2b" can be concretized as "has obtained victim's forgiveness and compensated for losses" and "moved the car a short distance in the parking lot."
When the prosecutor reviews case X1, they must judge which of the four plots is the main reason for their decision based on the Criminal Law and relevant judicial interpretations. If they decide not to prosecute mainly based on f2b, then it can be inferred that, in their view, the Criminal Law and its judicial interpretations require them to treat f2b as the main reason for non-prosecution. Regrettably, the machine cannot directly identify the substantive connotation of the reason for judgment. Facing case X1, humans can use Criminal Law articles and other regulations to determine that f2b meets the conditions for non-prosecution and use it as the reason for non-prosecution. The machine cannot understand the above content in the way of human cognition. Criminal Law articles cannot be directly embedded into its decision function, nor can the machine directly construct the subsumption relationship between constitutive elements and case facts. Therefore, when handling case X1, what the machine learns is merely: in the decision logic of human case handlers, the plot set f1bf2b holds an advantage in the balance against the plot set f1af2a, thereby becoming the basis for the non-prosecution decision. But it cannot identify which of f1b and f2b is the constitutive element or decisive reason for non-prosecution, because a single case X1 cannot provide enough information to support this identification; nor can it determine whether the reason for non-prosecution can still win in the weight balance when a new reason supporting prosecution f3a appears in a new case X2.
3.3 Formation of the Priority Order of Reasons
Further work involves incorporating diverse individual case documents into the case library, enabling the machine to gradually master the weighing logic of judgment reasons. To describe this learning process, the concept of "Priority Order of Reasons" needs to be introduced. Knowing that in case X1, if the case handler makes a non-prosecution decision based on the case plot set f1af2af1bf2b, then the machine will discern from it: when these four plots appear simultaneously, f1bf2b will hold the advantage in the balance. For the sake of discussion, this situation can be expressed as f1b,f2b}>X1{f1a,f2a. Here, ">X1" represents the priority order of reasons in that specific case, which is to say, simply regarding case X1, the two reasons supporting non-prosecution are prioritized in ranking over the two reasons supporting prosecution. That is, the plot set on the left carries heavier weight than the plot set on the right and can better support the decision conclusion it belongs to.
Accurately identifying this priority order is of great significance; it reveals the initial mechanism by which machines learn from humans to make correct judgments. It is not difficult to imagine that if the training set contains a large number of cases with the same plot combinations as X1, then through continuous trial and error, the machine will gradually learn that whenever these plots appear in a case, the probability of making a correct non-prosecution decision is higher. In other words, at this point, the reason order relationship between f1bf2b and f1af2a will expand from a single case to the entire case library. That is, it can be considered that a stable priority order of reasons has formed at the case library level. This process can be further described using symbols: First, mark the case library containing a large number of cases of a certain type as T. If in the cases of this case library, as long as the plot set f1af2af1bf2b appears, the case handler will make a non-prosecution decision, it can be expressed as f1b,f2b}>T{f1a,f2a, where ">T" represents the general priority order of reasons existing in case library T. This order expansion from individual cases to the whole provides the core basis for the machine to master the decision rules of similar cases.
It can be considered that in the case library composed of a certain type of case, there are numerous entries presented in the above form, constituting a set list of priority orders. However, it is not ruled out that there may be contradictory priority relationships, such as a case library simultaneously containing f1b,f2b}>T{f1a,f2a and f1a,f2a}>T{f1b,f2b. For such cases, when certain specific plot combinations appear, the judicial decision of prosecution vs. non-prosecution may present a situation of equal strength. In other words, these two entries are inconsistent in logic, which at least indicates that past case handlers have not yet reached a consensus on the judicial decision under this plot combination. Usually, it is believed that although such inconsistent situations occur in practice, as long as the consistent entries occupy a clear advantage in quantity, a model capable of reliably identifying the priority order of reasons can still be trained.
3.4 The Acquisition Mechanism of Judgment Reasons
The learning process of judgment reasons is realized through the continuous strengthening of the case library, which mainly includes three tasks: discovery of reasons, delineation of reason weights, and consolidation of reasons.
3.4.1 Discovery of Reasons
As mentioned before, the first case incorporated into the initial case library is X1, and its case fact plot set is f1af2af1bf2b. Assuming that the human prosecutor makes a non-prosecution decision in this case, the priority order of reasons learned by machine learning is f1b,f2b}>T1{f1a,f2a. This means that for the case library T1 at the initial learning stage, the only case X1 it contains provides a single reason priority. Of course, the machine cannot learn too much useful information from this single entry because it does not know what played the decisive role in the decision result—was it "already obtained victim's forgiveness and compensated for losses," or "moved the car a short distance in the parking lot," or their combination.
After the case information of X1 is stored, a case X2 containing new markers is incorporated into the case library. The updated case library is denoted as T2, and T2=T1∪{X2}. The marked case fact plots contained in X2 include f1af2af2b, and the prosecutor's handling result is non-prosecution, i.e., f2b}>X2{f1a,f2a. At this point, the reason order list contained in T2 has two entries, which are f1b,f2b}>T2{f1a,f2a and f2b}>T2{f1a,f2a. If the machine learns in the correct way, the information it should master at this time is: in case X1, the fact plot that truly influenced the prosecutor to make the non-prosecution decision is f2b, and this conclusion can be extended so that when the marked supporting prosecution case plots are only f1af2a, if the non-supporting prosecution plots contain f2b, then the highest confidence is assigned to the non-prosecution decision. In other words, after adding the new case, the machine has mastered a key judgment reason for the non-prosecution decision in the class of cases like X1, which is "moved the car a short distance in the parking lot" and all fields incorporated under this element. This process of identifying key plots through multi-case comparison can be briefly called "Discovery of Reasons."
3.4.2 Delineation of Reason Weights
Human prosecutors and judges can usually clarify the weight or strength of a judgment reason, but AI does not possess this initial ability. The case plot structure appearing in similar cases like X1 is f1af2af2b. If case facts other than these plots appear in a case, the machine cannot identify them. It cannot determine the relationship between the new plot and f1af2a, nor can it clarify whether this plot can defeat f2b. Once the machine cannot understand the meaning of the new plot, it can be considered that it does not know the boundaries of the judgment reason. At this time, to let the AI learn to recognize the difference in reason weights, new cases must be added to strengthen the case library.
Assume a case X3 containing new markers is incorporated into the case library, and X3={f1a,f2a,f3a,f2b}. The updated case library is denoted as T3, and T3=T2∪{X3}. The case document shows that the prosecutor decided to prosecute, and the corresponding priority order of reasons is expressed as f1a,f2a,f3a}>X3{f2b. This constitutes a major update to the original priority order of reasons. It indicates that in this case library T3 the reason priority list simultaneously possesses two entries: f1a,f2a,f2b}>T3{f2b (Note: This looks like a typo in the original text or OCR, likely meant f2b}>{f1a,f2a combined with the previous context, but strictly translating: the text says f1a,f2a,f3a}>T3{f2b and f2b}>T3{f1a,f2a). If f2b}>T3{f1a,f2a contains the judgment principle for a class of cases, then f1a,f2a,f3a}>T3{f2bconstitutes an exception to this judgment principle. This means that when the relevant case presents the plot f3a,f2b will be defeated; the weight of the reason "moved the car a short distance in the parking lot" is insufficient to support a non-prosecution decision, and the machine should assign the highest confidence to the prosecution conclusion. For ease of discussion, f3a can be concretized as "violently obstructing police from enforcing the law." It can be found that the AI did not infer this exception by itself, but identified this exception from the documents entered into the library. The boundary and strength of judgment reasons are gradually constructed exactly through the "new plots overturning precedents" appearing in newly added cases. As long as the AI can obtain a sufficient number of exception plot samples from the case library, it can gradually carve out the weight levels of a judgment reason, realizing defeasible reasoning or retractable reasoning. This process of calibrating reason weights through exception plots can be called "Delineation of Reason Weights."
3.4.3 Consolidation of Reasons
A more common situation is that although people have marked many new plots in the documents, these documents are still based on the plot f2b to make a non-prosecution decision. Do these newly added marked documents have significance for machine learning? The answer is affirmative; these new documents are by no means useless. They play a strengthening role for the existing priority order of reasons. Initially, when the prosecutor made a non-prosecution decision based on f2b, the machine could only learn that this plot could defeat a limited number of supporting-prosecution plots. But as documents with new markers for non-prosecution decisions are continuously added, the plots that can be defeated by f2b in the case library also expand significantly. This constitutes a strengthening of the case library, and the priority order of reasons is updated accordingly because f2b can support a non-prosecution decision in more diverse plot combinations. This process of strengthening reason efficacy through newly added similar decision documents can be called "Consolidation of Reasons." Some scholars have noted that if "similar cases appear constantly, and the same judgment pattern operates repeatedly, the judgment rule established by the prior case will become increasingly stable." This stabilization process is realized precisely by constantly expanding the range of plots defeated by the judgment reason.
The above is merely a simplified case restoring the mechanism of the case library in supervised learning. After a large number of marked documents enter the case set for neural network training, the entire learning process becomes so complex that it cannot be completely described in natural language. However, no matter how complex the process is, the above three tasks are the most basic; they advance simultaneously and intertwine with each other. Lacking the discovery of reasons, one cannot find the facts playing a key role for the judicial decision; lacking the delineation of reason weights, it is difficult to clarify the weight boundaries of reasons in decision-making; lacking the consolidation of reasons, AI will be unable to respond to new plot combinations. To make AI perform better in every training sample and reduce the error rate, it is necessary to complete the three basic tasks of reason discovery, reason weight delineation, and reason consolidation as fully as possible.
4. Limits and Improvements of Judicial AI under the Reason Model
After utilizing the Reason Model to restore the process by which the case corpus assists AI in learning judgment reasons, we can further examine whether the various defects of judicial AI pointed out previously truly exist, which features constitute the fundamental difference between humans and machines in identifying judgment reasons, and how to overcome the structural defects of judicial AI in situations where this difference cannot be eliminated in the short term.
4.1 Response and Clarification to Several Opinions
A most common criticism is that neural networks actually do not possess human consciousness; they act merely according to established procedures and cannot understand the meaning of the case and the causal relationships within it like humans do via documents. This statement can be summarized by John Searle's argument, namely that AI can master syntax, but the human mind uniquely masters semantics. Whether this idea that AI cannot understand semantics has been challenged by constantly developing Large Language Models is currently inconclusive, but it must be clarified that when we ask whether AI can understand and respond to reasons like humans, we actually only need to directly judge whether AI can make correct decisions based on reasons, without needing to consider whether the mental content of it engaging in such activities is sufficiently human-like. This is just like if a machine can diagnose cancer with extremely high accuracy, whether it truly understands the meaning of "cancer" is not a key issue. In fact, this criticism is associated with a deeper doubt, namely that AI cannot replicate the human worldview and meaning system. Therefore, scholars believe that it cannot make value judgments of right and wrong, good and evil like humans. However, this view is open to question. Inferences based on the Reason Model show that AI is not incapable of learning value judgments. Deep learning is reverse engineering of human thinking; its task is to crack the complex functional relationship between case fact plots and judicial decisions. If the case handler themselves makes decisions based on legal principles, sentiments, and reason, and even culture and national conditions, then when engaging in reverse engineering, what the AI learns is not just the law, because the priority order of reasons contains not only legal reasons but also moral reasons and prudential reasons. After all, the performance of AI depends entirely on the judicial data fed by humans. AI's judgment ability is ultimately parasitic on human judgment ability. As long as value judgments are implicit in historical judicial data, a correctly trained AI can certainly encode value judgments into the neural network to improve the accuracy of assisted decision-making.
Actually, with the improvement of assisted case-handling models and the continuous superimposition of case plot element data, the priority order of reasons contained in the case corpus will become richer, thereby more fully absorbing the value judgments of judicial personnel. For example, the earliest intentional injury case assisted case-handling model launched by Shanghai judicial organs contained 86 items of key case element data, while the theft case assisted case-handling model contained 102 items. In the subsequently launched assisted case-handling model for crimes of aiding information network criminal activities, the key element data reached 106 items, and for picking quarrels and provoking trouble, it reached 144 items. Among these element items, purely legal elements account for only a part. In addition, Shanghai judicial organs are promoting the full online migration of offline case-handling behaviors to eliminate the "dual-track case handling" mode of online/offline separation. After the promotion of the online single-track case handling mode, machines will be able to obtain judicial data directly from the "case-handling behaviors" of handlers and other judicial personnel without relying on fields presented in documents. This means that content previously unexpressed in documents, especially data carrying the handler's common-sense judgments and moral intuitions, will have the opportunity to be captured by the AI's neural network. In this situation, judicial AI will have the ability to make value judgments because the material for value judgments is already buried in historical judicial data.
If the Reason Model accurately describes the way AI learns judgment reasons, it is not difficult to infer that the iteration of judicial AI depends on whether the case library is sufficiently strengthened, and this strengthening depends on the accumulation of judicial data. The case-handling model cannot create data out of thin air, nor can it skip data to directly evaluate reasons. If data accumulation is insufficient, the case-handling model's sensitivity to perceiving reasons will be insufficient. After all, judicial decisions must form symmetry with historical data; insufficient data inevitably leads to a lack of decision precision. It can be seen that the core dilemma facing judicial AI is that it lacks a mechanism to enable it to achieve active decision-making and shape new practices in a state of information scarcity. A widely cited example is that AI cannot provide a legal answer that is both unexpected and rational for hard cases. To be fair, the inability to handle hard cases is a weakness that does not constitute a significant reason to deny the value of AI-assisted judicial decision-making. Humans are also often helpless in the face of hard cases. As long as the positioning of smart justice lies in improving judicial efficiency and reducing repetitive labor, temporarily giving up on solving complex and difficult cases seems excusable.
Some scholars worry that AI-assisted adjudication will force case handlers to lean towards historical average judgment results, ignoring attention to the specificities of individual cases, thereby deviating from substantive justice. In fact, current assisted case-handling models have set at least three thresholds to prevent the occurrence of injustice in individual cases: first, case handlers have the right to refuse to accept the system's automated decision suggestions when entering the model operation; second, when there is a deficiency or deviation in the system's automatic extraction of case elements, case handlers can manually input supplementary key information; third, even if the system generates an assisted decision conclusion, case handlers still have the right to adjust the result based on the actual situation of the individual case. It can be seen that the fallibility of AI regarding special individual cases is a defect that can be managed or remedied.
4.2 The True Defect of Judicial AI Under the Reason Model
The true dilemma facing judicial AI is not the inability to cope with hard individual cases, but the lack of innovation ability. The consequences it brings are structural, to the extent of shaking the promise of judicial automated decision-making regarding improving case-handling efficiency and saving judicial resources. There is a huge difference between AI's reason identification mode and the human way of identifying reasons. To put it more bluntly, machines learn reasons from data, while humans learn reasons from theory. Humans organize reasons with a set of theories, while machines cannot construct and apply theories; they can only identify judgment reasons in reverse and imitatively based on comparisons between plots in the case library (discovery of reasons, delineation of reason weights, consolidation of reasons). Reasons are not bred from a set of legal theories but are acquired from the reverse engineering of historical judicial data. If the formation process of judicial data is "generating 'data' from 'case-handling behavior'," then machine learning is essentially restoring case-handling behavior from data. Whether it is the discovery of reasons, the delineation of reason weights, or the consolidation of reasons, it is all about completing an overall task: finding the judgment reason that appears most "precise" from all dimensions within historical judicial data. "Precise" here means eliminating errors, i.e., decisions must be symmetrical with historical data, thereby maximizing the reduction of the unexpectedness of judicial decisions and ensuring that the handling of similar cases maintains stable and predictable consistency. The basic principles of deep learning in neural networks are the backpropagation algorithm and gradient descent algorithm, which serve a single goal: minimizing error. The example of "going down the mountain" mentioned earlier hints at this: facing the surrounding fog, the mountaineer needs to find the path to the valley through constant trial and error. Similarly, the machine needs to "tentatively" repeat trial and error to find the hidden layer weights and activation values that perform best in all training samples, avoiding decision results that are surprising or unexpected. This working mechanism inevitably pursues the safest correctness.
The machine's trial-and-error learning does not exist under theoretical guidance, whereas human learning is inevitably guided by theory. Learning from mistakes is essentially remembering a correct state reached by chance, remembering the correct steps as a "reward," and then continuing to grope in the fog of insufficient information. In contrast, according to Whitehead's classic "three-stage learning process" of education, human learning of theory undergoes a triple cycle of romance (discovery/exploration), precision (accuracy/mastery), and generalization (synthesis/clarity), thereby realizing the transfer and promotion of knowledge. Learning from theory is not blind but possesses necessity, sudden enlightenment, and undeniable nature. The essence of theory is a cluster of propositions; it contains a set of reasoning modes and causal cognitive modes for judging things. It aims to achieve a self-consistent explanation of all phenomena. Especially when new phenomena appear, the existing theoretical edifice needs to be re-examined and continued to maintain its own integrity. Just as T.S. Eliot's interpretation of art theory states, "The existing monuments form an ideal order among themselves, which is modified by the introduction of the new (the really new) work of art among them. The existing order is complete before the new work arrives; for order to persist after the supervention of novelty, the whole existing order must be if ever so slightly, altered; and so the relations, proportions, values of each work of art toward the whole are readjusted; and this is conformity between the old and the new."
This tendency and desire to maintain integrity constitute the motive force for subversive theoretical narratives. Therefore, theory does not always aim to reduce the degree of "surprise"; on the contrary, it may even deliberately create surprise and challenge past fixed opinions. in this sense, theoretical work is far ahead of machine learning, but it is precisely this head start that allows it to keep a distance from the solidified structure formed by long-term data accumulation, thus being able to adapt to "reform." When laws, policies, social concepts, and social practices undergo replacement and innovation, and judicial practice has not yet fully responded, so relevant data has not been captured by AI, humans can already use various theories to adapt to changes and quickly form countermeasures. This is a huge advantage of humans and something AI cannot do. Neural networks can almost only use the holistic reason order of old cases to understand the meaning of newly appearing cases. However, if we have a set of legal theories, it will use the new cases to adjust and transform the holistic order formed by old cases. The transformed order remains complete but cannot be captured by the neural network in advance.
AI can understand what old integrity is, but it cannot understand what new integrity is. In a case library that has been fully learned, when a unique new case is incorporated, for AI, the priority order of reasons is not changed at once. The change of order requires repeated discovery of reasons, delineation of reason weights, and consolidation of reasons—a "Ship of Theseus style renovation process." This is not a change that a solitary case can drive. According to the Reason Model, it requires enough samples, and these samples must cover enough diverse situations, to strengthen the case library and let the neural network undergo effective trial and error. In contrast, theory can transform the entire reason order by absorbing merely one case. The appearance of a "new trick" stimulates and agitates the theory, and the new theory, after transformation, can immediately reshape the meaning of all past cases in the library. This ability to reverse-transform the vast majority of old content with extremely little new information is unique to theory.
Someone may ask, can't theory be created from data? Probably not. The reasons why theory cannot be born from data are: First, data aims to reflect reality, while theory aims to change reality; theory is not a "mirror of the past." Second, theory is driven by concepts. Theory constructs itself top-down. When a novel theory is proposed, there is not necessarily corresponding practice, nor is there data formed through these practices. That is to say, in the absence of data, concepts and propositions are formed first and then applied to the world. This ability to change the world sometimes generates unprecedented data. Third, theory pursues integrity in explanation. The persuasiveness of theory is not implicated or corrected by past data, because past practices may have made holistic errors, while theory pursues logical or rational integrity. If past data conflicts with this integrity, what should be abandoned is not the theory but the data. Fourth, in competition, theory wins by its "uniqueness." The essence of competition between theories is the trade-off between several different question expansion and convergence schemes; and the strength of a theoretical scheme is determined by its persuasiveness relative to other theories. The persuasiveness of a new theoretical scheme is reflected in multiple dimensions: its special argumentative layout and design, the unique analytical entry point it focuses on, the opportune moment it allows new concepts to enter the stage, and ultimately, its individualized way of speaking—only a sufficiently individualized way of speaking can breed unprecedented persuasiveness. However, individualization itself means not being bound by established data. The above four points explain exactly why novelty is a characteristic unique to theory, because new theories cannot grow naturally out of old historical data.
According to the two basic algorithmic logics of deep learning, in the case where the training set does not undergo significant changes, the modeling method will remain unchanged. This means that in the cognitive framework of AI, the current order will be assumed to "continue forever." This structural disadvantage will produce the following two consequences in specific judicial practice.
First, when radical changes and innovations occur within legal dogmatics and other legal theories but are not timely transmitted to judicial practice, AI cannot detect the reform. There is a "1633 Metaphor" in the AI theory circle, imagining a Large Language Model trained only on all scientific publications and other written materials before 1633 (the year Galileo was sentenced to life imprisonment for promoting heliocentrism). At this time, if someone asks whether people should support "heliocentrism" or "geocentrism," the LLM will only reiterate and feed back the scientific consensus of the majority, i.e., "geocentrism." This thought experiment can be translated to the Chinese context. There might be a so-called "2020 Metaphor" in the criminal justice practice of justifiable defense in China. Before the Supreme People's Court, the Supreme People's Procuratorate, and the Ministry of Public Security jointly issued the Guiding Opinions on Application of the System of Justifiable Defense according to Law (Fa Fa [2020] No. 31), the priority order of reasons adopted by AI in justifiable defense cases must conform to "basic compatibility theory," meaning the interests protected by the defender and the interests infringed upon by the defense counterattack should be roughly balanced. The long-term "judicial inertia" fostered by this concept meant that in both prosecution and trial stages, factors like "defense can only be implemented against violent infringement," "defense can only be implemented at the moment of violent infringement," and "defense can only inflict damage equivalent to the violent infringement" were taken as the fundamental basis for judging whether defense was established or excessive. This led to an abnormally low recognition rate of justifiable defense. The Yu Huan case in 2016 and the Yu Haiming case in 2018 pushed the legal circle to fundamentally reconstruct the conditions for the cause, timing, object, and intent of justifiable defense, abandoning the basic compatibility theory under strict balance and acknowledging the dynamic nature of infringement, adopting an ex-ante standard from the defender's standpoint, and even affirming the defender's privilege of error within a certain range.
However, the above theoretical shift was not directly transmitted to the practical circle; there was a certain time lag. Although the second-instance judgment of the Yu Huan case and the decision to dismiss the Yu Haiming case were groundbreaking, at that time, the vast majority of historical judicial data was still deeply influenced by the basic compatibility theory and the theory of interest balance. In the vast sea of similar cases, a few isolated cases were powerless to turn the tide. Even in the thinking of AI, the Yu Huan case and Yu Haiming case were merely "deviant instances" during modeling, belonging to "statistically tolerable errors" in data. One can imagine that what the neural network learned from judicial data was inevitably the relatively harsh standard for recognizing justifiable defense. There is a transition period from conceptual reform to practical shift, and precisely in this stage of insufficient supply of new data, judicial AI may fail. Even if new data enters the AI training set during the transition period, due to the inconsistency between new and old data, the priority order of reasons finally learned by AI is likely to be incoherent, unable to support a definite judicial conclusion.
Second, when the law itself, which is complex enough, undergoes addition, modification, or abolition, AI cannot change its decision pattern accordingly. For simple legal amendments, AI usually only needs a few adjustments by the case handler on the operation interface to achieve adaptation, such as modifying checkbox options, adjusting the association logic between options, or resetting discretion magnitude bars. These operations will directly reflect as corresponding changes in decision tree branches and knowledge graph nodes. When facing major changes such as the promulgation of new laws or the abolition of old laws, the situation is more complicated. The re-coordination between laws, and between laws and interpretation methods, will produce complex systemic effects. Taking a legal text with a "General Provisions—Specific Provisions" structure as an example, if the norms in the General Provisions are modified, how do the changed General Provisions norms affect the Specific Provisions? If the norms in the Specific Provisions are modified, can the General Provisions be applied to the modified Specific Provisions, and if so, how? These questions all possess complexity, and this complexity stems from the fact that the "General—Specific" structure itself implies a set of departmental legal theories. For example, if legislators amend the Criminal Law text in the future to include the general provisions of non-genuine omission offenses into the General Provisions of the Criminal Law, then how non-genuine omission offenses apply to the specific crimes in the Specific Provisions of the Criminal Law becomes a theoretical question. People might ask, apart from the omission offenses explicitly stipulated in the Criminal Law, do all crimes already stipulated in the Specific Provisions of the Criminal Law, as well as future new crimes, have "omission" establishment situations? At this time, a general theory about omission offenses is needed to judge one by one whether specific crimes "are suitable" for omission situations, and "how to apply" them.
Regrettably, AI cannot "create something out of nothing" to form such a set of theories; it lacks sufficient data needed for modeling. Consequently, in the vacuum period where the law has been amended but theoretical work is not yet sound, judicial AI cannot directly apply the modified law but can only wait for a large number of judicial case handlers to reach a consensus on the systemic effects after the legal amendment with the help of omission theory. AI's re-understanding of the consensus of case handlers is actually a secondary understanding, or derivative understanding. In contrast, humans, who are good at building prospective theories, can handle the above dilemmas relatively easily. The effective operation of machine learning relies on the "symmetry between data and belief"; accurate judgments are produced only when data accumulates to a certain magnitude. However, humans can break through this symmetry constraint when applying theory. Even facing a small amount of data, or even a single typical case, they can discover cognitive breakthroughs through theoretical tools, thereby rapidly adapting to the new situation and completing the transformation of the decision model.
4.3 Overcoming the Limitations of Judicial AI Under the Reason Model
Under the Reason Model, although AI can learn the extraction method of judgment reasons from data, it does not master any legal theory justifying these judgment reasons. Under existing technical conditions, there is no effective way to resolve this dilemma, but it can be mitigated through certain means: Given that the core characteristic of AI lies in its ability to reverse engineer human thinking, this characteristic can be used to compensate for its defects. If learning judgment reasons from long-term judicial data is AI's first request for help from human intelligence, then when reform occurs, humans updating judgment reasons based on legal theory and feeding them to AI constitutes AI's "second request for help" from human intelligence.
4.3.1 Maintaining a Moderately Sized Case Library with a Flexible "Exit-Entry" Mechanism
When normative documents such as laws and judicial interpretations are substantially modified in terms of fact-finding, application of law, etc., and the systemic effects produced by the modification are assessed as unclear, the use of AI-assisted case handling should be cautious. Only after accumulating enough human judicial data should judicial AI be retrained and deployed. The priority order of reasons is accumulated and enriched through the strengthening of the case library. The more complete the reason ranking list, the more beneficial it is for the machine to accurately identify judgment reasons. In the phase of connecting new and old laws, judicial data will undergo a process of re-accumulation, and correspondingly, the reason ranking list must also undergo a renewal period. During the investigation, some grassroots case handlers reported to the author that sometimes for two cases with roughly the same plot, the decision content recommended by the intelligent system differs greatly. The reason lies in being in the adaptation period of replacing old norms with new ones. To avoid misleading judicial personnel, adjustments should be made to the intelligent assisted case-handling system at this stage. If the cost of stopping the system is too high, one can consider establishing a connecting "exit-entry" mechanism, clarifying the conditions and scale for old cases to "exit the library," especially needing to clear cases handled long ago that cannot be incorporated into the new rules from the case library. At the same time, ensure that cases handled under the new law are entered into the library in a timely manner to achieve dynamic updating of the case library, thereby reducing the error rate of assisted decisions. Perhaps, to enable AI to learn the judicial decision patterns after rule modification more quickly, a relearning mechanism can be considered. For example, in the already online assisted case-handling model, if the case review conclusion is manually modified and is inconsistent with the conclusion recommended by the system's assisted decision (e.g., changing "prosecute" to "do not prosecute," "imprisonment" to "probation," or changing the sentencing and fine range recommended by the system), the handler triggers a deviation prompt when submitting the review conclusion and needs to fill in the reasons for deviating from the recommendation for the document to take effect. This architecture can be used to train the neural network's adaptability to the modification of laws and judicial interpretations. Specifically, if case handlers begin to modify recommended review conclusions in large numbers after rule modification, the handlers' explanations for deviant decisions can be injected into the training set as new reasons for the AI to learn, thereby adjusting the priority order of reasons more quickly.
4.3.2 Establishing a "Case Library Adversarial System"
A case library adversarial system should be constructed, meaning different legal knowledge groups independently generate judicial data and form mutual checks and balances. Within the legal professional community, there exist sub-groups such as police, prosecutors, courts, lawyers, and scholars, as well as sub-groups divided by region and trial level. These sub-groups should be allowed to build their own case libraries and train AI independently, thereby forming a priority order of reasons unique to each group. The advantage of this approach is that if a certain legal professional sub-group makes significant progress in legal interpretation and application, this progress will not be suppressed because the group is a minority. This forms a mechanism for checks and balances and improvement. While avoiding systemic rigidity, it facilitates each sub-group contributing its understanding of the law to the entire legal professional community. In terms of system design, the case library adversarial system allows each digital project to form a team following the principle of "business personnel leading, technical personnel following up, business personnel proposing case-handling needs, and technical personnel providing personalized realization," choosing historical judicial data to train AI according to their collective values. Thus, even if assisted case-handling systems of different sizes are formed, if different models can form "incompatible" review conclusions, it will prompt handlers and adjudicators to further review the case and reflect on their collective historical experience. This helps various sub-groups in the legal professional community to achieve mutual checks and supervision, avoiding the trap of "human-machine collusion." Some scholars point out that the human-machine interaction in smart justice is essentially "not a relationship between human and machine, but a relationship between the group of judges and the individual judge," because AI is "the crystallization of the wisdom, thinking, and experience of the entire judge group and the logical mode established by humans on this basis." If this assertion holds, then the construction of the adversarial system provides an institutional tool for people to reflect on and adjust this collective experience and thinking mode.
Currently, it is not realistic for each People's Court and People's Procuratorate to establish case libraries to train AI individually. A more reasonable approach is to take the High People's Court or Provincial People's Procuratorate as the unit, deploying business personnel from three levels of courts/procuratorates to form a special task force with internal technical personnel or external technical personnel, responsible for training the assisted case-handling model. In the future, a working mechanism of "unified organizational level, dispersed training level" can be considered, i.e., the same team builds a unified algorithm architecture, while grassroots courts/procuratorates, intermediate courts, and branch courts/procuratorates can independently select different ranges of cases to train their models according to their needs. This move can ensure calculation quality while taking into account regional differences in society, economy, and culture in different areas.
Conclusion
Current practice shows that if judicial AI is to learn judgment from cases, it must simulate human identification and utilization of reasons. However, AI can only learn reasons from the case corpus, whereas humans can create, reflect on, and correct reasons through legal theory. This difference leads to AI's lack of ability to cope with change. The more fundamental reason behind this phenomenon is that only the advancement of theory itself can absorb laws, concepts, and cases with substantive innovation, thereby rapidly adjusting the priority order of reasons. To this day, human wisdom remains the sole driving force promoting theoretical advancement.
There is a tendency in current academic research: placing AI in the core protagonist position, with jurisprudence becoming one of the many disciplines cheering for its achievements. However, legal theory cannot and need not become a secondary, service-oriented theory attached to AI technology. Legal theory contains humans' profound understanding of normativity and causal mechanisms. Machines can only track theoretical trajectories from data but cannot construct theory out of thin air. Future human-machine collaboration perhaps should clarify a core division of labor: humans provide theoretical energy, leading conceptual innovation and practical progress; AI preserves progress results, lifting overall judicial efficacy to the baseline level and above.

Originally published in "Chinese Journal of Law" (Fa Xue Yan Jiu), Issue 5, 2025, pp. 54-73. Thanks to the WeChat public account "Fa Xue Yan Jiu" for authorization to reprint.