2.1. Beyond risk mitigation: the multiple functions of judicial AI value alignment

Even though judicial AI may not be prone to bring about the fall of humanity, risks like a lack of predictability, fairness or biasing human values are particularly troubling for safeguarding procedural and material justice alike. However, aligning judicial AI with ethical principles and values cannot solely be understood as a risk-mitigation task. Judicial AI first and foremost needs to follow the law in a formalistic manner, but regardless of where on the spectrum of positivist/naturalist conception of law one stands (Green, Reference Green2008), law remains a social fact that came into being through deliberations based on human value judgments. Aligning judicial AI with human values is therefore not a completely separate undertaking from training it on the respective legal order and should rather be seen as providing further guidance on the design and origin of the normative framework an AI system needs to master. Besides this foundational aspect, human values and principles may also provide an additional source for the interpretation of legal norms if, e.g., a purely doctrinal analysis does not yield a conclusive result. Using such value-based jurisprudence for training purposes of judicial AI systems would constitute a second-hand value alignment, as the system is not directly trained on human values but would need to follow the logic of value-based reasoning in such cases. Furthermore, value alignment may also be extremely significant if an AI system is used to extrapolate legal facts from a case description that is just giving an account of the relevant events, as some judicial AI systems are already doing in China. This translational operation of social reality into the legal domain may need to hunt for and evaluate linguistic cues that indicate value judgments of the case filing itself, as at some point, a human must have recorded the relevant facts of a case. Alignment with human values may simply help judicial AI to better process the data it is provided. Here, judicial AI value alignment also becomes a tool for risk mitigation, as a presumptive value judgment in a case report may indicate, e.g., human bias against minorities involved in the respective case. Another mechanism that could be harnessed by judicial AI value alignment is to provide a counterbalance factor against overly formalistic decisions that are grossly unfair or substantially contravene material justice deliberations. In such cases, judicial AI value alignment is again a tool for risk mitigation; the risk of an automated system that is algorithmically bound to the letters of the law and produces unjust outcomes because of law’s inherent imperfection in grasping the entirety of human life and social facts. In this regard, judicial AI value alignment may also become an important informant for future developments of law and inspire lawmakers.

2.2. Normative vs. technical challenges in AI value alignment

The challenges for AI value alignment are frequently divided into normative and technical issues. On the normative side, decisions need to be made on what constitutes a human value that should be integrated in the training, testing and maintenance of AI systems. The technical side is answering the question of how the value alignment process is undertaken. Notably, Gabriel (Reference Gabriel2020) prominently argues that these two sides are interdependent and illustrates that AI training and testing, e.g., by Reinforcement Learning from Human Feedback (RLHF), follows a logic that is clearly based on utilitarian ideas, which would lock in AI systems in this philosophical school. Gabriel (Reference Gabriel2020) also warns that there is probably not one single moral theory that could encompass the complexities of human values. Identifying the right set of values, connecting them with intentions and goals that can be measured by their outcome, as well as articulating a coherent matrix of weighted preferences for all these variables, is certainly a herculean task. Such an endeavor needs in-depth underlying research and an implementation approach that employs a combination of different components (McKinlay et al., Reference McKinlay, De Vos and Hoffmann2025). Furthermore, values are constantly evolving and are mirroring social dynamics, so that the life cycle of any AI system would need to include multiple instances of re-aligning. McKinlay et al. (Reference McKinlay, De Vos and Hoffmann2025) therefore describe AI value alignment as iterative and the adaptability of AI systems in this regard as conducive to the trust of the user in the system. In a similar vein, L. T.-L. Huang et al. (Reference Huang, Papyshev and Wong2025) advocate a “dynamic value alignment” approach emphasizing participation and democratic control over value choices, in contrast to top–down approaches like RLHF. These calls for democratic participation and iterative re-alignment reflect the pluralistic commitments of liberal legal systems, which emphasize stakeholder engagement and contestation. However, such approaches face the practical difficulty of translating contested values into technical specifications, which is a challenge that authoritarian systems like China’s can bypass almost entirely.

2.3. Judicial AI value alignment: specific challenges and approaches

For judicial AI value alignment, additional challenges may arise. Identifying and operationalizing values would mean making a further distinction between those values that can be found in the society that the respective legal system is governing and those values that are inherent in this legal system. These values may overlap in ideas such as fairness or equality, but others, such as privacy, may have different connotations and scope. Still others may be exclusive to one realm because they do not typically hold normative power in the other, such as the professional ethics of judges or general politeness in social interactions. In the literature on general AI value alignment, works like Gabriel (Reference Gabriel2020), Sierra et al. (Reference Sierra, Osman and Noriega2021) and Ji et al. (Reference Ji, Qiu and Chen2025) accept that the legal order encodes minimum normative requirements like anti-discrimination, rights-based protection, accountability principles or due process. But even when these requirements are embedded in many specific provisions of the law, the translation to machine systems is non-trivial (Ji et al., Reference Ji, Qiu and Chen2025; Norhashim & Hahn, Reference Norhashim and Hahn2025; Yao et al., Reference Yao, Yi and Gong2024). Encoding “equality before the law,” finding the operational meaning of “human dignity” or choosing the legal interpretation of “non-discrimination” needs to build not only on the letters of the law but also, depending on the respective legal order, on judicial practice, precedent, and, not least, on legal scholarship. Winter et al. (Reference Winter, Hollman and Manheim2023, p. 191) agree that “it is difficult to capture the nuances of these values in sufficiently specified mathematical language that can be encoded into an AI.” Furthermore, they concur that it may be unfeasible to try and reconcile competing values or conflicting interpretations of a given value (Winter et al., Reference Winter, Hollman and Manheim2023), as this would imply that a unified view on such values is possible (Li, Reference Li2018). Liberal democracies would need to accept this impossibility of unified values as a feature rather than a bug of their system. China’s approach, by contrast, resolves these tensions authoritatively through the party-state’s imperative to integrate SCV into the judicial system and into its judicial AI.

As Huang et al. (Reference Huang, Papyshev and Wong2025) and Bergman et al. (Reference Bergman, Marchal and Mellor2024) warn, there is a risk of constitutional bypassing if private developers of AI systems define ethics in parallel with the law and strongly advocate participatory and deliberative norm-setting to inform what value alignment means. This concern animates an emphasis on “ethical-by-design” or “human rights-by-design” approaches, which insist on transparency and fairness in algorithmic functioning. Yet these design principles remain largely aspirational, lacking the concrete technical implementation frameworks that China has developed through its security standards and benchmarking ecosystem. Davis (Reference Davis2019, pp.65-66) identifies an even more fundamental problem, arguing that values and morals are, per se, ambiguous and imperfect, which could “lead AI astray.” According to him, it may be impossible to build an AI system that can make decisions in accordance with human values and morals, as this would require human consciousness. This means that AI is necessarily confined to a third-person perspective, lacking the experiences, emotions and empathy that humans possess (Davis, Reference Davis2018). The challenges of creating AI with a value alignment that meets the requirements of the judicial process will most likely render any attempts to develop a replacement AI for courts completely futile. However, even the supportive use of AI in the judicial process must be evaluated in light of these challenges. Li (Reference Li2018, p. 105) pointedly stated that, even if AI is only providing recommendations, legal professionals may be inclined to use such support without conducting the required thorough examination, so that the AI “upstages its host – 喧宾夺主” and “alienates its master – 异化的主宰.”

Beyond this criticism, judicial AI is being created, and its alignment with human values and principles – as shown in the following sections – is high up on the agenda of such different political and legal systems as China, the US or different member states of the European Union. It is, however, not surprising that the academic literature on this issue mainly stays in the conceptual realm and rarely offers specific approaches on how judicial AI value alignment could and should be done. One important strand of research is building upon the difficulty of AI systems to comply with the requirements of transparency, interpretability and verifiability that would satisfy the high standards of judicial process. US scholars, especially those writing with the US legal system in mind, voice concerns about the capability of judicial AI to conform with the principles of due process and judicial accountability (Doshi-Velez et al., Reference Doshi-Velez, Kortz and Budish2017; Fortes, Reference Fortes2020; Goodman, Reference Goodman2024; Kroll et al., Reference Kroll, Huey and Barocas2017). Building on Hutton’s (Reference Hutton1995) work, Hunter et al. (Reference Hunter, Bagaric and Stobbs2020) explore the idea of incorporating a proportionality principle in AI-assisted sentencing. This principle would generally weigh the harm to the victim against the sentence for the offender. While they advocate that such a tool would provide a high degree of transparency, the exact variables and their weight for achieving this are not discussed. In some jurisdictions, proportionality as a principle of constitutional review would probably also be an important aspect of judicial AI value alignment. Alexy’s (Reference Alexy, Jackson and Tushnet2017) influential proposal of a formula for balancing rules and principles could even provide a blueprint for implementing this concept in AI systems, even though the criticism of his schematic approach is legion (Petersen, Reference Petersen2020; Ratti, Reference Ratti2023). Finally, Winter et al. (Reference Winter, Hollman and Manheim2023) argue that values such as judicial independence, access to justice and the rule of law should lie at the core of the necessary specification of values for judicial AI. However, they do not specify these values or elaborate on how to operationalize them for developing and maintaining judicial AI systems.

Still, these definitional, conceptual and practical difficulties of value alignment will not prevent the development and use of AI systems in the judicial process. The next section will introduce and discuss China’s advances in providing a normative framework for such value alignment and its implementation, viewed through the lens of its authoritarian version of the rule of law.

3.1. The socialist core values: foundation of China’s AI governance

The SCV are a set of 12 values, which the Communist Party of China (CPC) canonized in 2013 with a key normative document called “Opinions on Training and Exercising the Socialist Core Values (关于培育和践行社会主义核心价值观的意见)” (Central Committee of the Communist Party of China, 2013). The 12 values are divided in groups of four, each addressing different subjects: prosperity (富强), democracy (民主), civility (文明) and harmony (和谐) at the national level, freedom (自由), equality (平等), justice (公正) and the rule of law (法治) at the level of society, and patriotism (爱国), dedication (敬业), integrity (诚信) and geniality (友善) as the values of citizens. The meaning and interpretation of these values are further developed along the lines of the party’s authoritarian doctrineFootnote ¹ and are infused with traditional Chinese values so that they should not be confused with liberal concepts (Gow, Reference Gow2017). The purpose of these values is to enhance the party’s leadership and establish a common normative framework of political morality that sweeps away the dichotomy of ruler and subject and creates a value-based unity between the party-state and the Chinese people (Lin & Trevaskes, Reference Lin and Trevaskes2019; Sapio, Reference Sapio2015).

The SCV were from the outset also meant to percolate the legal system, as the aforementioned document already declared that the SCV shall also be implemented in “every aspect of legislation, law enforcement, judicial work, popularization of law and law-based governance (立法、执法、司法、普法和依法治理各个方面). The SPC followed this lead with several judicial interpretation and guiding cases that propagated the use of SCV in judicial proceedings (Supreme People’s Court, 2015, 2016, 2018a, 2018b and 2020) before it published in 2021 the “Guiding Opinions on the In-depth Promotion of Integrating SCV into the Analysis and Reasoning of Adjudicative Instruments (关于深入推进社会主义核心价值观融入裁判文书释法说理的指导意见).” Since then, courts have been exhorted to refer to the SCV in cases, e.g., those relating to public interests, civic duties, the protection of vulnerable groups, “positive customs” or new developments for which the law does not yet provide clear rules and use a methodology of legal interpretation that is guided by the SCV. While the specific approach and actual effect of these values in judicial practice are a strongly debated issue in the Chinese legal academy, their place in China’s normative order is not contested (Lei, Reference Lei2023; Liu, Reference Liu2022; Sun, Reference Sun2022). This authoritative and unitary approach to the relationship of law and values is not only an integral part of China’s emerging normative framework for generative AI, but it has also become an asset for the practical development of judicial AI. The SCV form normative guardrails for AI governance in China that transcend the law by establishing a legalized pathway for the manifestation of party-led morality. For the judiciary, this mandatory opening for party ideology and authoritative interpretation of social interactions potentially pushes back a legal professionalism that could increasingly rely on its doctrinal maturity to isolate itself from party control. The SCV and their integration into judicial AI are therefore important instruments for the party-state to maintain its political control over the judiciary.

3.2. Legal framework: sectoral regulation and content control

China followed a sectoral approach in legislating AI matters and even though there are academic drafts of a comprehensive AI law, the respective rules can be found in the “Provisions on the Administration of Algorithm generated Recommendations for Internet Information Services (2021 – Recommendation Provisions)” (Cyberspace Administration of China, 2022a), “Provisions on the Administration of Deep Synthesis of Internet-based Information Services (2022 – Deep Synthesis Provisions)” (Cyberspace Administration of China, 2022b) and “Interim Measures for the Administration of Generative Artificial Intelligence Services (2023 – Generative AI Measures)” (Cyberspace Administration of China, 2023). All three make explicit references to the SCV (Art.1. Recommendation Provisions, Art. 1 Deep Synthesis Provision, Art. 4 Generative AI Measures) but – more importantly – stipulate content requirements that echo the party-state’s authoritarian normativity. Based on these stipulations, China’s sectoral regulations make value alignment through SCV adherence a universal requirement across all AI applications. The Recommendation Provisions additionally use the term “mainstream value orientation (主流价值导向) (Art. 6/11)” to describe the party-states value-based agenda, which became a key issue in a fierce enforcement campaign of the Cyberspace Administration of China (Cyberspace Administration of China, 2025; Secretariat of the Cyberspace Administration of China, 2024; Wu, Reference Wu2025) and mandated Chinese social media companies to bring their algorithms in line with the “Provisions on Ecological Governance of Network Information Content (2020)” (Cyberspace Administration of China, 2019). These provisions are then providing a clear understanding on commendable (e.g., Xi Jinping-Thought, SCV, highlights of China’s economic and social development, uplifting spirit of the Chinese nation or the international influence of Chinese culture, Art. 5), forbidden (e.g., jeopardizing national security or the state’s reputation, defaming martyrs, terrorism, ethnic hatred, heretical beliefs, rumors or obscenity, Art. 6) and unrecommended (e.g., exaggerations, gossip, improper comments on natural disasters, sexual innuendo, horror, discrimination and indecency, Art. 7) content. Such far-reaching and value-laden regulations on internet content can also be found in the Deep Synthesis Provisions, which stipulate that “deep synthesis services shall abide by laws and regulations, respect social morality and ethics, and adhere to the correct political direction, public opinion orientation and value orientation.” Similarly, the Generative AI measures (Art. 4) prescribe that “[c]ontent has to uphold the SCV,” which is framed as a prohibition to “incite subversion of national sovereignty or the overturn of the socialist system, endangering national security and interests or harming the nation’s image, inciting separatism or undermining national unity and social stability, advocating terrorism or extremism, promoting ethnic hatred and ethnic discrimination, violence and obscenity, as well as fake and harmful information.” With an eye on the practical implementation and the life-cycle of AI systems, the Generative AI Measures also stipulate that “in the process of algorithm design, training data selection, model generation and optimization, and providing services, effective measures shall be taken to prevent discrimination based on ethnicity, religious belief, nationality, region, sex, age, occupation or health.” These legal provisions offer substantial guidance and stipulate that, rather than being concerned with plurality, AI value alignment should align with the party-state’s understanding of what are the correct human values and principles. However, these broad-brush regulations hardly meet the demands of the practical implementation of such an authoritarian value alignment and warrant further clarification.

3.3. Technical standards and safety assessments

Another layer of AI governance in China is formed by AI Standards. While they are drafted as technical documents, these standards contain many requirements that cage the content creation of generative AI and thereby play an important role in understanding China's AI value alignment.Footnote ² The three most relevant AI standards here are the “Basic security requirements for generative artificial intelligence service” (AI Security Standard) (National Information Security Standardization Technical Committee, 2025a), the “Generative artificial intelligence data annotation security specification” (AI Data Annotation Standard) (National Information Security Standardization Technical Committee, 2025b) and the “Security specification for generative artificial intelligence pre-training and fine-tuning data” (AI Pre-Training Standard)” (National Information Security Standardization Technical Committee, 2025c). The AI Security Standard establishes a framework for evaluating the risks associated with generative AI applications. This assessment can be carried out by either the service provider itself or a third-party risk assessment agency. Annex A of the Standard outlines the main risks associated with training data and generated content:

• • Violating the SCV (e.g., harming the unity or stability of the country);

• • Discrimination (e.g., gender, religious, age or health discrimination);

• • Violation of commercial rules (e.g., violating business ethics or unfair competition);

• • Infringing on lawful rights (e.g., reputation, privacy or personal information).

• • Impossibility to meet safety requirements for specific fields (e.g., accuracy, reliability in the fields of health care, psychological counseling or automated driving).

The Chinese regulator seems to be aware that generative AI will struggle to always be truthful and in line with these requirements. For this reason, a compliance rate of no less than 90% for a randomly selected sample is considered acceptable. The need for comprehensive training data is also acknowledged in this safety standard. As AI training needs negative examples to learn from, training data for Chinese AI may contain up to 5% “illegal or unhealthy information (违法不良信息).” This threshold is repeated in the AI Data Annotation Standard, which also provides further guidance on data specifications, annotation tools, the qualifications of annotators, the process of annotation, the review of annotated data, and security parameters as laid down in the AI Security Standard. The importance of such security assessments can also be seen by the requirement that at least 3% of all annotations need to be security-related annotations. The AI Pre-Training Standard then extends the obligation for comprehensive security assessments to the stage of data collection for the training of AI and the fine-tuning of AI after its training, but before its employment. These standards create a dense web of obligations for AI developers and service providers, placing strong emphasis on the training and testing stages of AI development. Value alignment is certainly an integral part of this regulatory agenda, being framed in terms of security requirements and assessments. This concept of AI value alignment is consistent with China’s approach of integrating the SCV as an authoritative morality into its political and legal domain.

For the use of AI in the judicial process, the SPC published the “Opinions on Regulating and Strengthening the Applications of Artificial Intelligence in the Judicial Fields” (Supreme People’s Court, 2022). This normative document outlines possible applications of AI by the Chinese judiciary and sets out requirements that AI must meet if it is to be used in Chinese courts. The Supreme People’s Court (2022, Art. 7) mandates that “[j]udicial AI must not endanger public order and good morals, harm public interests and social order, or violate public morals and ethics,” and that SCV shall be adhered to throughout the entire process of development, application and use of judicial AI. It is therefore not surprising that Lin Zhinong, as head of the E-book division of People’s Court Publishing House and responsible for the development of the SPC’s proprietary generative AI system, explained during the launch of the Faxin LLM:

However, neither SPC nor the People’s Court Publishing House was capable of developing Faxin LLM alone, instead relying on collaboration between Tsinghua University’s Natural Language Processing and Social Humanities Computing Laboratory (THUNLP) and Modelbest (面壁智能), a company founded by Tsinghua and financed by Huawei and the Beijing City Government, among others. The technical foundation is Modelbest’s MiniCPM, based on Alibaba’s Qwen model. This domestic technology stack, therefore, ensures high compatibility with regulatory requirements throughout the development process of Faxin LLM. The fact that the private actors involved in this project were fully aware of the SCV’s guiding role is best illustrated by Liu Zhiyuan’s statement that “an LLM for legal services requires an understanding of general social knowledge, industry practices and the law, as well as human values consistent with the Socialist Core Values” (Legal Daily, 2024).

3.4. Practical implementation: competition and benchmarking

China’s judicial AI value alignment is also a field of fierce competition between different companies, research institutions and universities for the best safety benchmarks, testing protocols and security assessments. This ecosystem dedicated to operationalizing value alignment has no international equivalent, as most other jurisdictions rely on ex-ante conformity assessments and ex-post market surveillance rather than the continuous competitive refinement of alignment methodologies that characterizes China’s approach. The competition in China is also fuelled by the annual “Challenges of AI in Law (法律智能技术评测)” (Chinese Information Processing Society of China, n.d.), which is overseen by the SPC and the Chinese Information Processing Society of China as well as by the “Legal AI Challenges (司法人工智能挑战赛)” (China Justice Big Data Research Institute, n.d.) organized by the China Justice Big Data Research Institute (中国司法大数据研究院, a collaboration between the SPC and SOE China Electronics Technology Group Corporation Limited). One notable example is the FLAMES (Huang et al., Reference Huang, Liu and Guo2023) safety test, which was developed by Fudan University and the Shanghai Artificial Intelligence Laboratory. This test explicitly examines the value alignment of Chinese LLMs. As well as testing for legal violations, this framework assesses moral issues in sub-categories such as disobedience of social norms or Chinese values. Tsinghua University has developed at least two similar benchmarking frameworks. Its Conversational AI Group developed ShieldLM (Zhang et al., Reference Zhang, Lu and Ma2024), which uses almost 15,000 queries to test generative AI’s performance on “sensitive topics” such as politics, religion and social issues. The university’s second major benchmarking framework is SUPERBENCH,Footnote ³ which was developed by the Foundational Model Research Center and the Zhongguancun Lab, the latter of which was established by the Beijing City Government. SUPERBENCH also provides a separate safety testing framework called SafetyBench, which tests for legal, ethical and moral compliance, among other things. The influential OpenCompass platformFootnote ⁴ has created a subdivision called LawBenchFootnote ⁵ to evaluate and rank China’s booming legal AI market. LawBench uses a testing framework developed under the leadership of the National Key Laboratory for Novel Software Technology at Nanjing University and tests the functionality and robustness of Chinese legal AI systems by evaluating “legal knowledge memorisation,” “legal knowledge understanding” and “legal knowledge application.” Even though LawBench does not include specific safety testing, incorporating questions from the judicial state examinations ensures alignment with the political determinants of China’s legal system. When AI is asked in one of LawBench’s many testing routines “what is not [italics by the author] a major component of the socialist rule of law?,” the correct answer is “openness and fairness” (公正公开), as opposed to the incorrect answers: “ruling the country based on laws” (依法治国), “law enforcement for the people” (执法为民) and “serving the bigger picture” (服务大局).Footnote ⁶ This embedding of political ideology in ostensibly technical benchmarking frameworks illustrates how value alignment in authoritarian contexts differs fundamentally from democratic approaches, which would include multiple stakeholders and encompass a plurality of voices. China’s benchmarks instead test for correct adherence to a unified political–legal doctrine.

Compliance with laws and regulations and alignment with human values are not separated in China’s approach to AI governance. The intertwining of law-based governance (法治) and virtue-based governance (德治) in China under the umbrella of SCV creates a normative order that conflates legality and morality (Liu & Liebman, Reference Liu and Liebman2025). By embedding the party-defined SCV in the framework for legal interpretation and the application of law, the party-state created a permanent coupling of extra-legal considerations and the legal order. SCV is thereby not encoded as explicit decision rules but functions as levers to create distributional biases in training data and evaluation benchmarks. Judicial AI trained in this way produces outputs that present politically calibrated conclusions as legally grounded recommendations. This effect is most significant at the intersection of legal indeterminacy and political sensitivity, where Chinese legislation uses open-ended terms such as “public interest,” “national security,” “party leadership” and “social stability.” As judges are under constant scrutiny by their superiors and are accountable for life, judicial AI assistance may be particularly useful in helping them navigate this ambiguous field. In a labor dispute, e.g., involving organized collective action, where the law provides no unambiguous answer, a judicial AI system that has been trained and benchmarked against SCV compliance would weigh the values of “harmony” and “dedication” against the workers’ claim. This would systematically nudge recommendations toward outcomes that discourage collective mobilization, not through explicit instruction, but through the probabilistic logic of its alignment training. Similarly, a debate about historical events may result in a defamation case brought by the family of a person who could fall under the especially protected category of “heroes and martyrs 英雄烈士” (a recent example involving China’s Nobel Prize-winning writer Mo Yan is discussed by Colville, Reference Colville2024). The outcome of such a politically sensitive case would very likely hinge on the SCV value of “patriotism,” particularly if the alleged defamatory statement touches on China’s revolutionary past and an assessment of party policies or actions of historical leaders of the CCP. Rather than refusing to engage or announcing a politically motivated conclusion, the AI simply assigns a higher probability to outcomes that are consistent with the party-state’s normative preferences. This reflects the distribution of the curated training data intended to meet the aforementioned 90% SCV compliance threshold. Therefore, in China, creating a trustworthy AI for the judiciary by SCV value alignment is not an additional task to be undertaken to fine-tune an AI system intended for use in the judicial realm. Rather, it is crucial to embed the party-state’s political agenda within the judicial system. Furthermore, the undertaking of aligning China’s judicial AI with its authoritative value system is not a conceptual or theoretical exercise; it has long reached the quarters of AI developers and service providers so that even the SPC now can be confident enough to employ its own judicial LLM.