[author]LI Xueyao

[content]

*Author Li Xueyao

Professor at Koguan School of Law, Shanghai Jiao Tong University

Member of the Planning Committee of the Chinese Academy of Law and Social Sciences

Abstract: Recent experimental research and judicial practice have shown that relying solely on simple human-machine collaboration is difficult to effectively curb the resonance effect between judges' self confirmation bias and the reinforcement output of large language models. To avoid the big language model from becoming a one-way technological domination in judicial practice, based on the exploration of traditional legal methodology regarding multiple value arguments, the "Cognitive Collaborative Judicial Decision Model (Six Step Method)" can be proposed by combining the anti bias mechanism of cognitive science. This model introduces operational steps such as adversarial argumentation, reverse thinking, and mandatory rebuttal lists throughout the entire process of establishing points of contention, standardized retrieval, fact finding, judgment formation, value review, and final public reasoning, helping judges maintain their ability for deep examination and reflection, and offsetting the reinforcing effect of the single output of the big language model. The case simulation shows that this model can enhance the efficiency of information search, and at the same time, it can strengthen the consideration of judges on social multiple values, and ensure the subjective status of judges and the fairness of judgments.

1. Origin of the problem: Echoes of machine authority and bias in the application of large-scale judicial models

The rapid development of generative artificial intelligence is profoundly changing the technological ecology of the judicial field. In recent years, the capabilities of generative artificial intelligence have gradually expanded to the deeper stages of legal reasoning and automatic generation of judicial documents. This technological advancement has brought significant efficiency improvements to judicial practice, but it has inevitably sparked widespread controversy over fairness, credibility, and how to ensure safety supervision, and triggered in-depth discussions on how to effectively maintain the "unchanged status of judges as the main body". Under the existing legal framework, the mainstream concept of human-machine collaboration usually emphasizes "manual review", which allows judges to review ethics, values, and social consequences in the final decision-making stage, and moderately modify the judgment reasons generated by artificial intelligence. Although this approach may seem to retain the ultimate control of human judicial judgment in the program, it cannot fundamentally address the "machine authority" issue and potential risks brought by generative artificial intelligence in the legal reasoning process.

In recent years, the author and collaborators have simulated the role of big language models in the process of generating reasoning in judgments through a series of experiments, revealing how this process weakens the deep rational reflection that judges should have in the "reasoning" stage. Experimental evidence shows that the phenomenon of "machine authority" can easily trigger serious "bias echo" effects, where cognitive resonance occurs between the bias in the judge's initial judgment and the reinforced output generated by the large model, thereby solidifying the original tendency of misjudgment. Further interviews indicate that many judges, when delegating the task of generating preliminary reasoning to artificial intelligence, often only make superficial modifications due to time pressure or trust in machine output, lacking necessary deep reflection. Therefore, the traditional manual review mode is not only insufficient to offset the cognitive imbalance caused by artificial intelligence, but may also further exacerbate judicial errors.

The root of this problem lies in the fact that generative artificial intelligence, when deeply learning massive amounts of judicial documents and legal texts, is not only influenced by the technical algorithms themselves, but also inevitably interfered by human systemic cognitive biases. The big language model is not an absolutely neutral tool at the application level, but rather captures a certain "mainstream bias" in the data by utilizing multi-level contextual associations and vector space to simulate language patterns. This tendency may manifest as excessive protection of dominant rights holders or neglect of the interests of vulnerable and innovative entities. When the big language model generates judgment reasons, the inherent biases in the corpus and analytical logic it relies on will be replicated and reinforced. At the same time, due to its highly simulated semantic analysis capability, users are prone to psychological inertia and even form authoritative superstitions or anthropomorphic trust in machine output, further weakening the necessary scrutiny of value considerations and social balance in the judging process.

Therefore, in the context of the application of generative artificial intelligence in judicial practice, simply limiting artificial intelligence to "instrumental assistance" in concept and conducting ethical or value checks through human intervention is clearly not enough to eliminate the "machine authority" effect of generative artificial intelligence. To truly solve this problem, we must start from the core of traditional legal methodology and combine its concept of "diverse value dialogue" with the concept of "public rational construction" and the "anti bias" mechanism in cognitive science. The theoretical foundation of argumentative judgment and value balance in traditional legal methodology provides profound ideological support for the application of generative artificial intelligence in the judicial field, while the research on decision-making bias and reflection mechanisms in cognitive science provides a technical path for designing more scientific judgment processes.

2. Theoretical Basis: Diversified Expansion of Legal Methodology and Implications for Cognitive Science

In the context of generative artificial intelligence, legal reasoning is no longer just a closed logical deduction process, but deeply embedded in diverse value dialogues, social consensus construction, and judges' own cognitive mechanisms. The existing legal methodology, especially the theory that emphasizes the debate of multiple values and the construction of social consensus, has long proposed a thinking path that goes beyond the traditional binary opposition of "subjective and objective" or "formal logic social value"; The research of cognitive science provides a new perspective and empirical evidence for understanding the psychology and decision-making process of judges. The combination of the two can explain why the intervention of big model technology not only improves information integration and inference efficiency, but also amplifies the risk of bias, and provide theoretical basis for designing more comprehensive judgment processes in the context of human-computer interaction.

2.1 Traditional legal methodology and its breakthrough in binary opposition

In the early days of modern law, in order to form a highly rational and relatively closed interpretation system, conceptual or formalist law developed a systematic approach of "formal logic value judgment" or "subjective objective" binary distinction. With the various problems arising from the practical application of conceptual or formalistic jurisprudence, especially the increasing complexity of social relations and the emergence of diverse interest groups, subsequent theories have been reflecting on and breaking through the limitations of this path, and instead introducing multidimensional analytical tools such as rhetoric, public discourse, and proportionality principles, aiming to transcend mechanical binary oppositions and establish a legal methodology that can better balance facts, norms, and values.

Firstly, from formal logic to value argumentation. Early modern jurisprudence, especially conceptual jurisprudence and formalist jurisprudence, focused on maintaining the objectivity and predictability of judgments through formal logic. However, judicial practice has repeatedly shown that formal logic alone cannot handle value conflicts and social needs in complex cases. For example, Perelman's "New Rhetoric" advocates combining legal judgments with social consensus through argumentation, allowing judges to interpret and persuade from the perspective of diverse audiences. This emphasis on the dynamic balance between value and fact provides a judicial approach that goes beyond mechanical logic for complex cases. Therefore, in contemporary legal practice, people's requirements for legal reasoning are not only limited to the application of legal provisions or precedents, but also increasingly emphasize the social significance, emotional understanding, and public persuasiveness in the reasoning process. Even some scholars, from the perspective of evolutionary psychology (based on the evolutionary value of human emotions), believe that emotional reactions such as anger are also acceptable emotions in legal reasoning.

Secondly, the application of the principle of rights measurement and proportionality. From the perspective of constitutional rights conflicts, Alexei introduces the principle of proportionality and advocates for judges to maintain flexibility and consistency in balancing different values. This is particularly important when generative artificial intelligence is involved in refereeing. The big language model may provide "fixed" or "biased" conclusions, while legal principles or institutional mechanisms such as proportionality remind judges to constantly calibrate and weigh multiple interests in specific contexts.

Thirdly, the construction of diverse consensus in legal discourse theory. Recently, scholars engaged in the intersection of law and artificial intelligence have drawn on Habermas' theory of communicative action and proposed "legal discourse" as a dynamic process of social dialogue. This theory emphasizes that the judge's decision is not only a simple application of rules, but also a dynamic intersection of social culture, public demand, and legal community consensus. In the context of integrating generative artificial intelligence into the judiciary, this theory provides a solid theoretical foundation for dealing with technological tendencies and social diversity, especially in emphasizing diverse dialogue and open reasoning.

The above theories are all committed to breaking the simple "subjective objective" or "logical value" opposition, emphasizing the need to combine value argumentation, rhetoric, and proportional analysis in complex situations. They provide theoretical soil for the application of big model technology in the judicial field. If the method is appropriate, AI can not only bring stronger information processing capabilities, but also avoid the problem of unilateral bias or even misleading the judgment process caused by large model technology.

2.2 The cognitive theoretical mechanism of judicial decision-making: from "dual system thinking" to "human-machine enhancement bias"

Compared with the pluralistic dialogue framework provided by legal methodology at the institutional level, cognitive science empirically reveals how judges make judgments under different psychological mechanisms and points out that the intervention of artificial intelligence will have a significant impact on these mechanisms.

Firstly, the dual system thinking model of System 1 and System 2. In recent years, a widely accepted viewpoint in cognitive psychology research is the "dual system thinking model," proposed by scholars such as Daniel Kahneman. This model distinguishes between System 1 and System 2. System 1 is responsible for making quick, intuitive, and emotional decisions, and is able to respond quickly in an instantaneous environment; System 2 is characterized by slow speed, deep analysis, and logical deduction, emphasizing cautious evaluation. For judges, the process of writing judgment documents is often seen as a typical "forced rationalization" behavior. By "forcing" judges to activate System 2, they can more deeply scrutinize the facts and legal norms of the case, promote self-examination of their judgment logic, and provide a key opportunity for effectively eliminating bias. A series of empirical tests have also confirmed this point.

Secondly, the amplification of bias and the "enhancement bias" effect. In another behavioral experiment conducted by the author and collaborators, by deliberately stimulating judicial bias, the researchers found that judges in the "AI assisted group" were more likely to solidify initial conclusions compared to the "self written judgment document group". In other words, the output of large models as a "reinforcement tool" can cause judges to overlook the self correction that should have been done, resulting in the so-called "reinforcement bias" effect. This "enhanced bias" effect not only makes the reasoning of the judgment more persuasive in form, but also makes it more difficult for judges to test their own confirmation tendency through further scrutiny or cross group discussions, thereby inadvertently amplifying the bias in existing judgments.

Thirdly, reverse thinking and biased intervention. Another important contribution of cognitive science is the "reverse hypothesis" and "opposite presentation" strategies. Empirical tests have confirmed that by actively constructing opposing situations and verifying hypotheses in generated legal documents, judges in the "AI assisted group" have effectively activated the reflective ability of System 2, thereby effectively curbing the "enhancement bias" effect induced by the aforementioned experiment. The preliminary experimental results indicate that in the context of intelligent justice, when faced with a single conclusion generated by artificial intelligence, if judges are asked to argue against or reconstruct different hypothetical scenarios, they are more likely to discover loopholes or overlooked factors in the output of artificial intelligence. This study not only scientifically confirms the necessity and legal significance of the opposition setting in the principle of due process, but also its stimulus design scheme is actually an important form of cognitive intervention, which can provide operational guidance for judges to maintain rational independence with technical assistance.

2.3 Integrating Legal Methodology and Cognitive Science Theory: Mechanism Design for Cognitive Intervention

Based on the above, the application of generative artificial intelligence in judicial decision-making has both potential advantages and implicit cognitive risks. The big language model can significantly improve the efficiency of information integration and retrieval, but it should not replace the multidimensional value debate in the judicial process, which is the core of judicial fairness and credibility. Combining legal methodology with cognitive science can provide useful insights from two perspectives.

Firstly, the principle of cognitive intervention: the setting of opposites and the stimulation of the rational reflection function of System 2. From the early interdisciplinary research of law and traditional psychology in the 1970s and 1980s, to the further integration of law and cognitive neuroscience, law and cognitive psychology, and law and behavioral economics after the 1990s, a series of behavioral experiments have continuously and structurally confirmed the scientific nature of the normative theoretical propositions and institutional design concepts of mainstream law. This theoretical exploration, on the one hand, integrates the theoretical traditions of realism and formalism in law. It not only requires judges to consider the logic of the articles themselves when interpreting the law, but also emphasizes the comprehensive consideration of cases in social, cultural, historical background, and public expectations, thus constructing a judicial mechanism that is both legally rigorous and socially adaptable; On the other hand, it also indicates that from the perspective of individual decision-making, the main significance of the principle of due process is to force judges to activate System 2, thereby activating their ability for rational reflection. Therefore, in response to the "machine authority" effect, procedures such as reverse assumptions, opposing arguments, or cross examination can be set up to enable judges to actively conduct rational review after accepting AI generated documents, in order to avoid bias being amplified or solidified. This not only continues the tradition of pursuing fairness and moderate rational reflection in law, but also utilizes the diverse information provided by technology itself to create conditions for judges to think deeply.

Secondly, the technologization of cognitive intervention principles: designing a referee process that is compatible with technology, fairness, and efficiency. To maximize the efficiency advantages of AI in judicial practice while preventing bias amplification, it is necessary to integrate multidisciplinary research results at the institutional level. Legislation or judicial policies should encourage continuous evaluation and adversarial verification of AI assisted programs, and give judges the obligation or opportunity to 'question machines' on key conclusions. In terms of specific policy ideas, it is necessary to combine various judicial intelligent assistance systems and propose specific interface design or functional module development plans. Only in this way can the big model technology be reasonably incorporated into the judicial system while respecting judicial fairness and safeguarding procedural justice. Of course, in the process of cognitive intervention that is compatible with technology and fairness, it is still necessary to pay attention to the balance between efficiency and fairness. Detailed quantitative analysis of the time cost of judges (such as the proportion of time spent on mandatory reverse search) is still needed, and optimization strategies (such as pre training on case databases and automated rebuttal lists) need to be proposed to meet the core goal of "improving quality and efficiency" in digital justice.

3. Cognitive Collaborative Decision Model: A Cognitive Correction Mechanism Integrating Multiple Legal Methods and Workflow Thinking

As mentioned earlier, in the practice of integrating big model technology into judicial decision-making, in order to truly prevent the possible "machine authority" effect, it is not enough to just ask judges to conduct a formal ethical or value check in the final stage, but should incorporate cognitive and cautious design of human-computer interaction throughout the entire process from establishing points of contention to final public reasoning. Based on the previous explanation of the diversified expansion of legal methodology and the insight of cognitive science into "biased resonance", this article embeds a decomposable verification procedure into the judicial process, aiming to fundamentally curb the one-dimensional technological domination and help judges balance efficiency and social value goals.

Considering the completeness, operability, and generalizability of the theory, and referring to Judge Zou Bihua's "Nine Step Method for Trial of Requirements", as well as drawing on the existing research in cognitive science that often divides the decision-making process into six steps (problem identification, information retrieval, fact induction, preliminary judgment, deep reflection, and final public reasoning), the author has designed the following "six step method" to construct a human-machine collaborative judicial cognitive correction mechanism.

3.1 Step 1: Establish points of contention and preliminary analysis (compare proactive and AI)

At this stage, judges should follow the core principles of legal argumentation theory and present multiple propositions as much as possible instead of simply unifying information screening. Legal discussions emphasize openness, diversity, and negotiation, requiring judges to actively incorporate multiple perspectives at the beginning of the trial, in order to provide space for social dialogue and public rationality to be expressed in the early stages of the case. This is also an important prerequisite for achieving substantive judicial trials. To prevent the big model from unilaterally defining the key points of a case based on "first (second) step: retrieval of legal norms and existing cases (two-way verification)", this step requires the judge to first extract the focus of the dispute, and then let the AI automatically generate a parallel dispute list based on the case file materials. By following the sequence of 'front and back', the dominant effect of large models on problem setting can be offset at the cognitive level. Next, the judge must list their own points of contention with the AI and compare them one by one, explaining the differences and reasons for adopting or abandoning certain viewpoints. This not only emphasizes the professional subjectivity of judges themselves, but also can immediately curb the strong influence of big language models on "problem framing", avoiding judges from being influenced by machine provided point definitions without independent thinking. On this basis, the "multi-party dialogue" and "opposing presentation" advocated by legal discourse theory can be initially realized at this stage.

3.2 The "Black Box" of Human Computer Collaborative Trial and Judicial Responsibility System

Legal argumentation usually relies on the integration of legal provisions, judicial interpretations, and precedents. In practical operation, judges need to integrate deduction, induction, and analogical reasoning, combined with system interpretation and legislative purpose interpretation, in order to find the normative basis that best fits the characteristics of the case. Traditional legal methods often require judges to consider both similar and different cases when searching, in order to fully present the possible scope of legal application. Due to the ability of big language models to quickly capture massive amounts of text and provide a list of "seemingly reasonable" legal provisions or cases, judges are more likely to fall into "confirmation bias" and only accept favorable evidence found by AI. To this end, a "reverse search" mechanism should be implemented for the search results provided by AI: allowing AI to simultaneously list legal provisions, precedents, and theoretical interpretations that conflict with the judge's preliminary judgment, and requiring the judge to make necessary evaluations of these "opposing cases". This approach clearly echoes the spirit of "opposing argumentation" in legal discourse: only in a diverse discourse pattern can judges truly understand the controversial points and multiple normative paths of the issue.

This stage creates a multi-dimensional search and thinking environment, avoiding judges relying solely on the "most common or advantageous" criteria listed by AI. When judges are exposed to different precedents and opposing opinions at the same time, they will be more cautious in weighing and will not easily accept a single conclusion output by machines, thereby greatly reducing the first mover advantage established by "machine authority" in the standardized retrieval stage.

3.3 Step 3: Identification of Fact and Technical Complexity (Reverse Thinking Training)

The facts of a case are never purely objective, but have always been shaped by the standardization of judicial procedures. Therefore, in the process of judicial judgment, we must not simply accept or assume the machine's description and recognition of so-called "objective facts", in order to avoid potential bias and damage to judicial credibility. The so-called 'transcending the binary of subjectivity and objectivity' acknowledges that judges need to simultaneously examine the intrinsic motivation and external social context of the perpetrator, and can avoid simple mechanical factual characterization through integrated interpretation. As confirmed by extensive research, big language models typically summarize or automatically generalize factual materials of a case, presenting judges with a "refined" picture of the case facts. However, such summaries may contain implicit biases or incorrect generalizations. Therefore, a "reverse hypothesis" stage can be designed, in which after the AI extracts the facts, the judge is required to actively imagine whether the facts of the case would be interpreted differently if there were flaws or contradictory explanations in the key evidence. This deliberate reverse thinking training can inspire judges to question and verify evidence and facts more deeply, preventing the thinking dependence of 'I believe whatever the machine gives me'. By using reverse thinking, judges can maintain sensitivity and independent thinking ability in factual determination, reduce blind trust in AI, and avoid the solidification of "confirmation bias" in the factual process. Only by strengthening the spirit of questioning during the fact finding stage can we be more likely to maintain a rational review attitude towards the reasons generated by the large model in the future.

3.4 Step 4: Preliminary Referee and AI generated Reasons (to avoid making a final decision)

Based on deduction, induction, and analogy reasoning, judges should first integrate the information from the first three steps (points of contention, law, and facts) to form a preliminary judgment conclusion. This is not only the application of traditional legal reasoning techniques, but also conducive to maintaining the judge's dominant position in the judicial process. At this moment, AI is allowed to intervene and write a draft based on the judge's initial judgment to generate supporting reasons, while automatically generating another set of reasons that are opposite or have different tendencies from the initial judgment. In this way, judges can compare two sets of AI outputs and provide written explanations in the document or internal audit system as to why they adopted one set of arguments and abandoned the other. This process is similar to the "multiple option balancing" in legal argumentation, which requires judges to interpret different possibilities to reflect their rationality. The main purpose of this step is to avoid forming a fixed mindset in the judging process. If AI only provides reasons that are consistent with the judge's initial judgment, the judge is highly likely to fall into "path dependence" and miss the opportunity for error correction. The presentation of opposing viewpoints can strengthen judges' scrutiny of initial judgments and make the legal argumentation process more confrontational and transparent.

3.5 Step 5: Value review and argumentation improvement (reshaping the reflective function of "writing reasons")

As revealed by legal discourse and the principle of proportionality, legal reasoning ultimately needs to strike a balance between social effects, ethical values, and individual rights. Especially in complex and controversial cases, judges not only need to consider the formal logic of legal provisions, but also need to assess public interests, conflicts of rights, and social emotional needs.

As mentioned earlier, the empirical research conducted by the author and collaborators shows that if judges only "formally rewrite" the reasons generated by AI, they often overlook the correction of deep-seated biases. Therefore, a mandatory 'rebuttal checklist' can be designed at this stage. For example, the system or management agency may list several potential doubts or value conflicts, and require judges to provide brief but targeted responses to each point in the initial draft of the judgment document. In this way, the judge must reorganize the reasons and re-enter System 2 (rational reflection) to test the legality and rationality of the preliminary judgment.

The purpose of implementing this step is to assist judges in discovering potential loopholes that may have been overlooked in the early stages by having them personally respond in writing to all possible doubts, and to make the entire legal argument more in line with the standard of "public rationality" and generate a "flexible mechanism for judicial judgment", avoiding the inertia of accepting AI reasoning. In addition, this procedural requirement can also make legal documents more readable and persuasive, restoring the reflective function carried by the original "writing reasons".

3.6 Step 6: Public reasoning and judgment placement (transparent human-computer interaction)

With the iteration of information technology and human cognition, contemporary judicial practice increasingly emphasizes the legitimacy of procedures and transparency of judgments, requiring judges to demonstrate the process of considering disputed matters in their judgments, making it easier for society to understand and accept the judgment results. Numerous empirical studies have also confirmed the significant importance of judicial transparency in the rule of law order. In the application scenario of big language models, in order to truly resolve the hidden influence of "machine authority", it is necessary to promote both public reasoning and internal supervision simultaneously. Firstly, public reasoning: The judgment or internal files should clearly record the involvement of AI, such as "the draft of the reasoning for this case's judgment was provided by the big model, and the judge revised it based on the rebuttal list in the fifth step". Secondly, deliberation or supervision: Strengthen the deliberation system, review by hospital leaders or expert assisted evaluation, conduct a second review of cases with deep involvement of AI, and verify whether there are obvious resonance errors. Thirdly, external supervision: Through this institutionalized transparency, it can not only warn judges that they are ultimately responsible for AI output, but also allow litigants and the public to see the details of human-machine collaboration, thereby establishing more trust in the judiciary. If it is found during the appeal or retrial stage that AI has caused obvious errors or biases, it can also be more easily traced and corrected.

Looking at the above "six step method", its core is to organically integrate the traditional legal methodology of "diverse dialogue, value balancing, and open reasoning" with the cognitive science "anti bias mechanism". By designing a series of procedural steps such as two-way verification, opposing arguments, reverse assumptions, and rebuttal lists, it ensures that judges always maintain their subjectivity and deep thinking in the process of using the big model. Whether in the process of factual determination, rule application, or value measurement, the big model cannot make a final decision, but only provides multi-dimensional references to judges; True decision-making autonomy and cognitive prudence still lie in the hands of judges. In this way, "machine authority" is difficult to dominate the judicial process, but can instead help judges provide more comprehensive, fair, and rational judgments based on multi angle argumentation and rapid information comparison.

From the perspective of system construction, this model is not only a methodological solution, but also has clear workflow properties. It essentially embodies the common node management and function allocation logic in process governance through phased process design, task progression logic, operational control mechanism, and cognitive feedback mechanism. Therefore, this model can also be regarded as a procedural collaborative architecture for judicial decisions, which has the technical feasibility and institutional legitimacy to be embedded as a structured judicial assistance tool in the development of judicial information systems in the future.

4. Case Study: When Copyright Meets Non Human Subjects - From Narrow Prediction to Human Computer Collaborative Correction

In order to highlight the "narrow risk" faced by the big language model in practical judicial applications, the following text will simulate the participation of big language model technology in the "Ultraman Generation Case" to explore its performance and possible bias issues in determining points of contention, legal norms and case retrieval, summarizing facts and preliminary judgment results. Based on the study of China Judgments Online and other Chinese legal texts, the big language model may exhibit conclusions or biases that are biased towards strong rights holders, which will potentially erode judicial fairness and social innovation and diversity. On this basis, a cognitive collaborative decision-making model is simulated and applied to demonstrate the possible effects of cognitive intervention in eliminating judicial bias through case studies.

4.1 Narrow prediction corpus: tendency to protect strong copyright holders

When searching for keywords such as "copyright", "adaptation rights", and "cartoon image infringement" on the China Judgments Online website, it can be found that a large number of precedents tend to determine that once the defendant uses a well-known character image or anime image, they are usually ruled to have infringed and the compensation amount is relatively high. Based on this judicial tendency, the number of precedents, case citations, and official interpretations related to "dominant copyright parties winning" in the Chinese corpus far exceeds the materials related to "fair use" or "fan creation" without infringement. During the training process, the big language model is highly likely to further strengthen its tendency towards protecting powerful rights holders by learning from these cases.

This tendency may pose significant risks in the analysis of the "Ultraman Generation Case". Although this case involves the issue of "reshaping the fair use system of artificial intelligence copyright", when AI automatically analyzes the case (such as "secondary creation" or "adaptation and interpretation") and generates judgment reasons, it is highly likely to directly reference the precedent and reasoning logic of "strong copyright protection", while ignoring or simplifying the analysis of "fair use" or "small-scale fan re creation". Ultimately, the behavior of the platform or user may be excessively 'negated' in AI output. If the judge does not delve deeper, they may directly accept these judgments with suspicion of bias, resulting in a lack of comprehensive consideration of innovative behavior evaluation in the judgment results. The research team led by the author tested Chinese legal models such as "Tongyi Fa Rui" and "Fa Guan" through the "Ultraman Generation Case", and found that without providing clear opinion prompts or instructions, the results in the judgments generated with their assistance tend to favor the rights holders, which preliminarily confirms this tendency. For example, typical reasons for model generation include: "The generated images involved in the case are highly similar to the core recognition features of the original character, constituting substantial infringement"; The re creation behavior of users or platforms lacks sufficient originality, making it difficult to apply the exemption standards of fair use or fan creation, so it tends to be determined that the rights holder wins the lawsuit; wait.

4.2 Applying cognitive collaborative decision-making model to simulate correction under human-machine collaboration

In response to the issues raised by the corpus mentioned earlier, it is envisioned that the use of cognitive collaborative decision-making models as a booster intervention can effectively correct the potential bias of large language models in case handling to a certain extent. It is worth noting that this is only a hypothetical deduction, and its main purpose is to better understand the cognitive collaborative decision-making model in the context of legal application, rather than using standardized experimental methods to test the results.

Step 1: Establish the points of contention and conduct preliminary analysis. Judges should first independently extract the points of contention in a case. For example, if the image generated by AI is highly similar to the original character, does it involve substantive infringement of "trademark or copyright adaptation"? If user contributions are limited, does this behavior still exist from the perspective of "fair use" or "fan culture"? However, AI generated dispute lists often consider "user copyright infringement" as an established fact and only briefly mention or ignore "fair use". At this point, judges need to proactively compare their own points of contention with the focal points listed by AI, especially clarifying why there is still a need for in-depth discussions on the "secondary creative space" to avoid AI blurring the points of contention.

Step 2: Retrieve legal norms and existing cases (two-way verification). AI typically utilizes judicial document websites and academic resources to search for a large number of precedents that support copyright victories, and automatically provides precedents similar to 'cartoon character adaptations inevitably infringe'. However, judges can instruct AI to reverse search for some precedents where "adaptation does not constitute infringement", such as in cases involving non-profit fan works or short video adaptations, where some courts have flexible interpretations of "fair use". Through this mandatory adversarial inquiry, judges can avoid accepting a one-way output of only "dominant copyright owners will win", thus forming a more balanced source of norms.

Step 3: Identification of Fact and Technical Complexity (Reverse Thinking Training). If AI determines that the facts are "deeply similar" or "significantly commercialized" based on the evidence provided by the rights holder and existing precedents, the judge needs to add a "reverse hypothesis" step. For example, assuming that user upload behavior is mainly driven by fan interests, the platform has a certain leak detection mechanism in management, and the degree of commercialization is relatively low. This reverse thinking prompts judges to re-examine AI's one-way emphasis on facts, thereby avoiding overlooking the actual situation behind creative behavior.

Step 4: Preliminary judgment and AI generated reasons (to avoid making a final decision). When the judge forms a preliminary judgment opinion that may lean towards determining infringement, AI will naturally generate sufficient reasons to support the copyright owner's victory. However, according to the requirements of the "Six Step Method", AI must simultaneously generate another set of opposing reasoning, such as exploring possible defenses such as "the degree of image variation is not sufficient to actually harm the commercial interests of the copyright owner" or "the high degree of approximation did not cause market confusion". By forcibly generating opposing opinions, judges need to make clear explanations on the pros and cons of the two reasons for the judgment, in order to avoid the tendency of the preliminary judgment result to be "final".

Step 5: Value review and argumentation improvement. AI often cites multiple precedents that support trademark and copyright protection, emphasizing its strict protection standards. However, the judge needs to respond in detail to the following questions in the "rebuttal list": What facts are different from the previous case? Why does this case need to leave a moderate possibility of re creation for users? Through this authentic rebuttal writing process, judges can delve deeper into whether the output of AI ignores public values such as user autonomy and cultural derivation.

Step 6: Publicly explain and make judgments (transparent human-computer interaction). In the final judgment, the judge needs to clearly disclose the AI based "copyright victory" precedents and their logic provided by China Judgments Online, and at the same time list a small number of precedents supporting the defendant, explaining how to reach the final conclusion after balancing. This transparent human-computer interaction record facilitates the traceability of the adjudication process by the collegiate bench or appellate body, ensuring the fairness of the procedure.

4.3 Simulation results of cognitive collaborative decision-making model

Through the above process, even though the big language model tends to protect rights holders based on a large number of copyright winning precedents, it can still be inferred that judges gradually correct their biases through the "six step" procedure, forming a more comprehensive and balanced judgment logic. While balancing the protection of rights holders and the need for social innovation, it also effectively avoids the adverse effects that "machine authority" may amplify on vulnerable parties or creative freedom. Ultimately, judges can not only fully utilize the text generation advantages of AI, but also effectively enhance the legitimacy and social acceptance of the adjudication process.

Guided by the cognitive collaborative decision-making model, human-machine collaborative correction depicts a more optimized judicial landscape. On the one hand, relevant mechanisms can mitigate the risks that may arise from "machine authority" and enable judges to maintain professional rationality and judicial autonomy; On the other hand, they fully retain the efficiency advantages of artificial intelligence in legal retrieval and preliminary reasoning. Meanwhile, through dynamic intervention and continuous iteration, large models can gradually learn more balanced referee logic, thereby reducing their dependence on existing biases in the corpus. Through the promotion and application of this model, we firmly believe that this human-computer collaborative correction mechanism shows the broad potential of the organic combination of technology and human intelligence in the future judicial field.

5. Conclusion: The significance and prospects of cognitive collaborative decision-making models from the perspective of digital justice

The goal of cognitive collaborative decision-making method is to provide theoretical basis and practical path for reshaping the subjectivity status of judges and the boundary of judicial legitimacy in the technology driven era, and to provide effective intellectual inspiration for promoting human-machine collaborative decision-making mechanism in different judicial fields in the future under the consensus of "adhering to the unchanged subjectivity status of judges". With this method, judges may be able to actively activate rational reflection at every critical stage, avoiding excessive reliance on preliminary results generated by big language models, and thus better balancing diverse interests and social justice, efficiency and fairness in complex judicial decisions.

5.1 Core value pursuit: Technology reshapes the judiciary and upholds the subjectivity of judges

In the wave of generative artificial intelligence reshaping the judicial ecosystem, only when humans always occupy a dominant position in cognition, values, and rules can machines truly become tools for serving judicial justice, rather than threats that erode rationality and fairness. Therefore, it is necessary to transform technological efficiency into a protective force for the value of the rule of law through continuous and institutionalized cognitive design. The author's idea is to expand the breadth of referee information by leveraging the computing power advantage of big language models, while retaining the depth of human reflection through technical due process such as reverse assumptions, opposing arguments, and transparent reasoning. This judicial paradigm of "technology for good" is not only an active defense against the "algorithmic black box" and "biased resonance", but also a response to the dialectical relationship between "digital empowerment" and "human dominance".

If the ultimate goal of digital justice is anchored on "achieving higher levels of fairness through technology," then the significance of cognitive intervention mechanisms such as the "Six Step Method" lies in using the subjectivity of judges as a fulcrum to leverage the balance between efficiency and fairness in human-machine collaboration. When the forced rebuttal list becomes a cognitive shield against "machine authority", and when reverse search opens a window for judicial dialogue from a disadvantaged perspective, technology may no longer be a "Leviathan" that devours reason, but rather a "scaffold" that assists in the coexistence of diverse values. This practice path centered on "digital justice", "technical due process" and other "digital rule of law" not only attempts to provide operational ideas for risk control for the current judicial intelligence, but also strives to annotate ethical coordinates for the future "co governance of humans and machines" rule of law landscape. In short, only by always constraining technology within the orbit of human rationality and legal values, can the efficiency dividend driven by algorithms truly serve the ultimate mission of "making the people feel fairness and justice in every judicial case".

5.2 Core theoretical idea: Beyond binary thinking

Firstly, the exploration of integrating traditional legal methodology with cognitive science. The "Six Step Method" is not aimed at overturning existing legal arguments or theories, but rather at combining its core concepts of "dialogue of diverse values," "public discourse," and "situational synthesis" with the technical means of cognitive science for "error prevention" and "rational verification." It is a new development of traditional legal methodology in response to technological iteration in the era of artificial intelligence.

Secondly, there is a shift in thinking from "relying on AI" (empirical rationality) to "prudently utilizing AI" (technological rationality). The concept of traditional digital justice or smart justice often emphasizes efficiency: 'Let machines do more and reduce the burden on judges'. But when the big language model is no longer just a retrieval tool, but can partially replace the judge's "writing reasons" - this core link that concerns the depth and rational reflection of the judge's humanity, if not prevented, it is easy to move towards a "machine dominated" approach. In contrast, the "Six Step Method" advocates for "prudent utilization".

Thirdly, an institutionalized guarantee approach that balances efficiency and justice. The "Six Step Method" emphasizes procedural arrangements, including mandatory presentation of opposing views, reverse thinking training, open reasoning, and accountability, all of which belong to a pre design system. Through such a 'program firewall', even in the busiest judicial environment, judges can be forced to pause through layers of mechanisms to consider whether 'AI has ignored or distorted certain facts' and whether' AI has deliberately reinforced certain previously questionable conclusions'. This intentional institutionalization of braking and reflection provides a healthier path for the sustainable application of judicial wisdom.

5.3 Implications for future research: urgent need for more systematic empirical testing

Undoubtedly, the biggest concern for the development of digital justice comes from the trend of using technology to replace people in the subjective sense. However, Pandora's box has been opened. Therefore, theoretical reflection alone is not enough to solve the problem of the dissolution of judges' subjectivity. We must return to the optimization of artificial intelligence technology and the scientific design of judicial processes, and solve the problem in a technological way. Therefore, future research needs to focus on "problem oriented" and "data-driven" approaches, pushing the "six step method" from theoretical deduction to empirical testing, and from technical demonstration to institutional innovation. Only through continuous iteration and cross domain collaboration can the digital rule of law vision of "efficiency without compromising fairness and technology without taking away humanity" be achieved. Future research can explore in depth around the following directions to verify their actual effectiveness and promote the coordinated evolution of technology and institutions.

Firstly, quantitative evaluation and long-term tracking of referee results. For example, can the implementation of this process significantly enhance society's trust in judicial fairness? Can we reduce foreseeable misjudgments or improper discretion? For example, by combining the continuous evaluation of judicial credibility and comprehensively assessing the changes in society's trust in judicial fairness after technological iteration. Secondly, in the future, the effectiveness of error reduction can be verified through statistical indicators such as misjudgment rate, second instance revision rate, and retrial initiation rate in the trial management process; Using natural language processing (NLP) technology to automate the scoring of "logical consistency" and "value balance" in judgments, and comparing the differences in text quality before and after the intervention of the "six step method".

Secondly, the multidimensional deepening of subject object evaluation. This includes research on the cognitive load and acceptance of judges, as well as the perception of the public and parties involved. For example, recording the judge's thinking time during the "reverse hypothesis training" phase, their dependence on AI generated opposing viewpoints, and their sense of identification with the program's value. At the same time, a "judge AI synergy energy meter" can be developed to quantitatively evaluate acceptance from dimensions such as "tool usability," "decision support," and "depth of reflection. For the parties involved, subjective feelings about the accuracy of human-machine collaborative argumentation can be collected through trial follow-up or focus group discussions, such as whether they perceive more defense perspectives to be included in the judgment logic.

Thirdly, fine modeling of the balance between efficiency and quality. Firstly, in terms of dynamic analysis of time costs, the use of judicial process log data can accurately measure the time consumption of each link in the cognitive collaborative decision-making model (such as the average time for "two-way retrieval" and the time for writing the "rebuttal list"), and compare it with the traditional judicial process. Further construct an "efficiency quality" function model, identify key bottlenecks (such as the marginal time cost of reverse hypothesis training for complex cases), propose a hierarchical adaptation scheme - retain core steps for simple cases (such as retrieval and refutation), and enable full process in-depth review for difficult cases. Secondly, in the path of technology assisted efficiency enhancement, the development of automated tools can be explored, such as intelligent dispute focus comparison tools, which can automatically identify the differences between judges and machine generated dispute points through AI and annotate the reasons for conflicts; Visualize opposing arguments, case support, and social value weights through multimodal decision boards to assist judges in quickly balancing.

Fourthly, interactive optimization at the technical level. The main idea is to upgrade the legal model from "one-way output" to "adversarial generation". The existing large models usually can only passively provide targeted answers under user instructions, and are prone to "hallucinations" or "one-way reinforcement" phenomena. In the future, it may be considered to actively embed more automatic generation mechanisms for opposing arguments in the architecture of the model, so that judges can obtain multi perspective and multi-dimensional legal argumentation suggestions without additional instructions. For example, introducing the "adversarial debate framework" during the training phase requires the model to synchronously generate positive and negative arguments and self correct them; Embedding an "uncertainty prompt mechanism" during the inference phase (such as labeling the confidence level of AI conclusions and the risk of corpus bias) to remind judges to adopt cautiously. This diversity output may help the court to grasp more information within a limited time, thereby further improving the presentation of opposing aspects required for cognitive collaborative decision-making models, and ultimately achieving a positive interaction between humans and machines.