2.1 Principal services of courts

Why should individuals and organisations defer to the judgments of a court, including its general interpretation of legal rules? Raz’s account of legitimate authority addresses this as a special case of the more general puzzle of authority: how can we justify that subjects should defer to commands, somewhat independent of the particular content of those commands, so that the commands preempt or exclude some of the subjects’ other reasons for action (Raz, Reference Raz1986, pp. 369–429)? Raz proposes that a body enjoys legitimate authority over a subject insofar as it enables the subject to better act as it has appropriate reasons to do. These reasons may be to promote its own short- or long-term interests and values, as well as to respect and promote the interests of others. The service an authority provides to a subject may, for instance, be due to it being swifter or wiser than the subject, less subject to bias or particular weaknesses, or better able to help subjects avoid coordination problems or include more relevant information about the interests of others. Thus, an authority may be legitimate for some subjects but not for others, and it may be a more legitimate authority than alternative arrangements. An authority may even issue some bad decisions while still giving subjects reason to act – as long as the subject is overall able to act somewhat better by deferring to the authority (Raz, Reference Raz1986, p. 75, Reference Raz and Raz1994, p. 214, Reference Raz2006, p. 1014).

So, what service do courts provide their subjects? We propose that the core services or tasks of courts are, first, to provide impartial adjudication based on judicial method, including the use of legal reasoning and method, meaningful participation of parties and provision of appropriate remediation (Fuller & Winston, Reference Fuller and Winston1978; Sourdin & Zariski, Reference Sourdin, Zariski, Sourdin and Zariski2018; Sturm, Reference Sturm1990). Indeed, many AI-inflected legitimacy criticisms against the judiciary focus on the right to a legal hearing “before a competent, independent and impartial tribunal” (Satterthwaite, Reference Satterthwaite2025, p. 10). Second, beyond the parties to the hearing, courts’ interpretation of legal rules often creates a jurisprudence – precedents which help stabilise expectations about the future behaviour of others (Sabel, Reference Sabel, Hollingsworth and Boyer1997). Notably, arbitrators in private adjudication are not understood to provide this broader service, which is one reason their decisions are not made public (Sweet & Grisel, Reference Sweet and Grisel2017).

Courts may, thus, be legitimate authorities in this sense insofar as they serve as agents of legislators or trustees of their subjects, in the broad sense, by providing these services the subjects have interests in. Courts can help individuals and organisations with dispute resolution. The parties avoid violent determination of conflicts by having the court’s decision preempt or replace the losing parties’ preferences for another outcome. The courts also help secure subjects’ interest in predictability. Courts assist each subject to better figure out what the law will be in somewhat similar and yet different situations – and hence what other law-abiding actors will do. Again, the court’s decision replaces some of the actors’ other reasons for acting otherwise. Courts may also contribute to provide assurance that public authorities and private actors are subject to the rule of law. Together, these services give reasons for subjects to defer both to the general authority of courts and to the authority of specific judgments.

2.2 Legitimacy issues in delegation to AI

Legitimacy concerns with courts can arise with both the assumption and delegation of judicial power, but their nature is different. What is at stake for the current topic is a question of delegation. Does the use of AI take authority hitherto allocated to and exercised by the judge alone, such that the judges modify or relinquish their legitimate role and authority?

In some respects, this concern is akin to the debate on “hidden actors” in courts, e.g., secretariats and clerks. Hidden human actors often provide important judicial services, such as improved efficiency, institutional memory and domain expertise (Baetens, Reference Baetens2019). The roles of these hitherto understudied actors have motivated a stream of scholarship: Computational and qualitative methods have helped identify who has written judgments in different courts (see review in Lie & Langford, Reference Lie and Langford2024). This research is often driven by a concern that such de facto delegation or shared authority with these hidden actors raises legitimacy issues, such as whether these hidden services are sufficiently transparent to affected parties, how they influence case selection and the quality of reasoning and whether they are sufficiently grounded in the informed free consent of parties and other branches of government (Pauwelyn & Pelc, Reference Pauwelyn and Pelc2022; Rosenthal & Yoon, Reference Rosenthal and Yoon2010).

In the case of AI, we can think of this technology as yet another hidden actor. Judges draw on the services of AI, ad even delegate to AI different judicial tasks. How this delegation occurs in practice may vary. Before considering the legitimacy concerns in such an act let us clarify the nature of various forms of services AI may provide, including delegation in various senses.

AI is often discussed in the singular, but it is, in practice, a cornucopia of different technologies with different uses – from highly automated and autonomous goal-driven systems to those which are more informational, dialogical or specific in nature (Bengio et al., Reference Bengio, Cohen, Fornasiere, Ghosn, Greiner, MacDermott and Richardson2025; Palmiotto, Reference Palmiotto2024). A useful way to categorise these uses is through delegation theory. We can discern at least six possible delegated relationships between courts and AI, where AI offers various opportunities and poses different challenges, some more complex than others. These six should only be regarded as a heuristic typology which is neither mutually exclusive nor exhaustive. To the contrary, we note below added challenges when the relationships are unclear, multiple or contested. Moreover, we note that there is an anthropomorphic quality to some of these labels, raising the common risk that they ascribe to AI “moral agency” or blur “ontological boundaries” (Salles et al., Reference Salles, Evers and Farisco2020, p. 93). However, several terms are common in delegation theory, and our intention is to communicate metaphorically the nature of the relationship, rather than any intrinsic human-like qualities – indeed, our final critique rests on fundamental epistemic and ontological differences between AI and human actors. We illustrate these six forms of delegation in Table 1, which run along a spectrum of increasing delegated discretion.

Table 1.

Forms of delegation to AI

Table 1 Long description

Six ideal types of AI service are aligned with delegation theories, stating each role’s core principle, delegation function, and main risks. Servant corresponds to a servant–master dyad focused on execution and performing tasks, with inaccuracy as the key risk. Socratic aligns with facilitator–interlocutor dialogue for informed conversation, risking weak dialogical competence, undesirable preference formation, and manipulation or advice giving. Advice matches advisor–advisee recommendation to provide suggestions or nudges, with risks including an inaccurate information base, lack of creativity, and anchoring or manipulation. Partnership uses a peer-to-peer equality model to reach an independent decision counted among others, with risks of inaccuracy, lack of explainability, and manipulation. Agency reflects a principal–agent autonomy model to reach a goal, highlighting moral hazard, information asymmetry, and inaccuracy; Trusteeship reflects beneficiary–trustee interests to avoid bad decisions, with paternalism or manipulation, information asymmetry, and inaccuracy. Across roles, inaccuracy appears repeatedly, while manipulation-related risks become more prominent as the AI shifts from executing tasks to shaping decisions or acting autonomously.

First, AI may be deployed as a servant. It carries out tasks which the judicial master could perform, to save the principal’s time, mental acuity, etc. Thus, it is simply automation of existing court processes regarding such things as filing of cases, checking deadlines and assigning cases. In classical AI technologies, such automation of rule-driven processes in courts and legal practice was often a central focus, although machine learning-based AI has been increasingly mobilised for such purposes (Ashley, Reference Ashley2017; Fonseca, Reference Fonseca2025). In these AI-as-servant cases, AI has no authority over the judge, its master: the reasons AI provides do not displace the reasons the judge already has. Nonetheless, there may be the standard risks that arise from any form of AI, especially inaccuracy but also a routinisation of tasks that require greater human, legal or domain-specific competence (Pasquale, Reference Pasquale2019).

Second, AI may be an interlocutor in the sense found in Socratic dialogues, serving either as the questioner or the responder. A judge may question AI in the Socratic process of “eliciting the truth by means of questions aimed at opening out what is already implicitly known, or at exposing the contradictions and muddles of an opponent’s position” (Blackburn, Reference Blackburn2006). This dialogical approach leans on the knowledge capacities of large language models (LLMs) and their apparent goal of helping find solutions (Chalmers, Reference Chalmers2025) and captures part of the idea of “Scientist AI,” advocated by some of the pioneers of deep learning. “Inspired by a platonic and idealized version of a scientist,” Scientist AIs are “AI systems designed for understanding rather than pursuing goals” (Bengio et al., Reference Bengio, Cohen, Fornasiere, Ghosn, Greiner, MacDermott and Richardson2025, p. 4). A Scientist AI is thus “trained to provide explanations for events along with their estimated probability” as part of the scientific process, e.g., of hypothetico-deductive reasoning, including “first understanding or modeling the world and then making probabilistic inferences based on that knowledge” (Bengio et al., Reference Bengio, Cohen, Fornasiere, Ghosn, Greiner, MacDermott and Richardson2025, p. 5). Of course, there can be dangers with such Socratic dialogue, which have been discussed since the time of Plato. As a questioner, a judge may provide, intentionally or not, overly leading or manipulative questions that provide a false sense of confidence in the answers provided by AI. Alternatively, the dialogue may be captured over time by the AI interlocutor who asks leading questions of the judge.

Third, AI may be used as an advisor. It assists judges, who are AI’s advisees, by helping the judges reflect on their overarching interests/values, and their more specific preferences, especially in applying judicial methods and choosing amongst interpretive strategies to pursue. In this respect, this advice service involves no formal delegation of decision-making authority and may be closer to the types of conversations judges may have with clerks, secretariats and other judges when they move beyond a Socratic dialogue and request direct advice. There is danger that AI can informally influence the adjudication process to a great extent and skew the deliberation by mechanisms such as framing and nudging (Chong & Druckman, Reference Chong and Druckman2007; Thaler & Sunstein, Reference Thaler and Sunstein2008). Thus, Schmauder et al. (Reference Schmauder, Karpus, Moll, Bahrami and Deroy2023, p. 799) argue that “blindly outsourcing the development and implementation of nudges to ‘black box’ AI systems means that the ultimate reasons for why such nudges work, including the underlying human cognitive processes that they harness, will often be unknown.” So there is a gray area where AI may serve both as advisor and as agent or trustee (see below), and where it is difficult to determine which roles it serves – and how.

Fourth, AI may be a partner in a judicial process. Here, AI reaches a decision on some matter, and this counts formally in the decision-making. However, its decision is not decisive on its own – the decisions of others also count under some procedure of triangulation or deliberation. The ultimate example of such partnership in courts might be AI as a member of a panel, e.g., as one of three judges in a case. More prosaic examples might be in matters of litigation procedure. For example, machine learning models have been shown to perform well in estimating the days needed for a court hearing in the overwhelming majority of cases but struggle with outliers (Bentsen et al., Reference Bentsen, Skiple, Langford, Birkeland Deichman-sørensen and Grendstad2026). A partial automated system to determine the number of days might, therefore, take the average of the human judge and AI’s estimation, weighting them equally or even differently for different types of cases. Such partnership approaches will often rest on conscious institutional design and may, in principle, optimise the respective advantages of human and machine decision-making. However, there remain considerable risks, including inaccuracy and lack of explainability, as well as a possibility for manipulation over time. Further important questions concern whether and how AI should acquire such joint authority to make the speech act of legally binding judgments (Austin, Reference Austin1975).

Fifth, AI may be an agent. The delegation of authority by the judicial principal furthers the judge’s express and fixed preferences in the face of alternative strategies, some of which may be best uncovered by an agent in a manner that is efficient and enhances quality (Gailmard, Reference Gailmard, Bovens, Schillemans and Goodin2013). If the principal is sufficiently dissatisfied with the agent, the principal may correct it in the particular case or replace it more permanently. Simple examples of dissatisfaction might be when an AI service hallucinates judicial precedents or translates relevant foreign sources incorrectly.

While AI as the agent for the judge is formally constrained (Stone Sweet & Brunell, Reference A. and Brunell2013, p. 65), such delegation raises a familiar suite of principal-agent challenges. Does the agent have the methods or competence for the particular task (competent agent problem)? For instance, Latif argues that “Persistent reliance on AI outputs invites a growing dependence on algorithms that lack genuine comprehension of legal precedents, statutes, and contextual nuances” (Latif, Reference Latif2025, p. 10). Moreover, will the agent act to further the expressed interests of the principal (strong agent problem)? Among several concerns, this raises the challenge of moral hazards. These often arise from a hidden misalignment of the interests of the principal and agent (Gailmard, Reference Gailmard, Bovens, Schillemans and Goodin2013) and/or asymmetry of information between them, making it difficult for the principal to verify and monitor the actions of the agent. Thus, AI may provide a formal presentation in support of particular outcomes yet hide the particulars of the case that explain the choice of one particular output or decision – which might be appropriate but could also be due to a misalignment of values, principles and rules (Bai et al., Reference Bai, Kadavath, Kundu, Askell, Kernion, Jones and McKinnon2022). Other strong agent challenges are information problems arising from the AI “black box” and lack of explanation (Miller, Reference Miller2019) and lack of transparency in the underlying collection and treatment of data (Yu & Barter, Reference Yu and Barter2024).

These challenges may be even more salient in AI models and systems with high levels of agency, such as LLMs and “agentic AI” (Bandi et al., Reference Bandi, Kongari, Naguru, Pasnoor and Vilipala2025). The latter form of AI is “capable of interacting, reasoning, and adapting in real-time to achieve specific goals autonomously” (Hughes et al., Reference Hughes, Dwivedi, Malik, Shawosh, Albashrawi, Jeon and De’2025), and is now widely available (e.g., Claude, Codex) (Caballero, Reference Caballero2026). When AI agents also coordinate other AI agents for a common goal the agent-principal challenge is further compounded. Some responses are alignment attempts such as “Constitutional AI,” which seek to deeply embed rules and principles in AI models (Bai et al., Reference Bai, Kadavath, Kundu, Askell, Kernion, Jones and McKinnon2022).

Finally, AI may assume the role of trustee for the judge. In such cases, AI acts to further the judge as beneficiary, promoting the judge’s overarching interests or values, such as proper judicial method, even if they are contrary to the judge’s express current preferences. Here, the beneficiary cannot easily revoke the delegated powers of the Trustee, especially not in the particular case (Alter, Reference Alter2008; Hale, Reference Hale2024). For such authority to be legitimate, as we shall argue, a necessary condition is that the trustee must somehow be better placed to regularly identify and act on the true interests and needs of the beneficiary – and be seen to reliably do so. There may certainly be scenarios for the justifiable use of AI as Trustee, for instance, to remove persistent bias or to reduce delays (Bentsen et al., Reference Bentsen, Skiple, Langford, Birkeland Deichman-sørensen and Grendstad2026; Peer & Gamliel, Reference Peer and Gamliel2013). Nonetheless, there is a risk of unjustified and unconstrained paternalism, common with all trustee models (Bellamy, Reference Bellamy2008), which can amplify the problems of agency models identified above. Indeed, there may be an ultimate danger of “epistemic dominance” by AI in an extensive trusteeship model, which can undermine the “trust in one’s own epistemic capacities and hinder the exercise of those capacities” (Coeckelbergh, Reference Coeckelbergh2023, p. 1341).

On its face, AI seems more desirable and legitimacy-enhancing when AI contributes as a servant, interlocutor and advisor, and, to a certain extent, partner and agent, rather than as a trustee. The first forms of use are typically judicial decision-support systems that support judges in their decision-making while formally leaving the final decisions in the hands of the judge (Hendrickx, Reference Hendrickx2025). The contribution of AI becomes often more controversial and problematic when there is more de facto delegated discretion and authority to AI.

Three points merit mention. First, delegation of discretion and authority to an AI system may in practice occur across the spectrum from servant to trustee. Second, there may be a distinction between theory and practice for each of these six relationships, such that their actual deployment blurs the boundaries. This occurs in cases involving “hidden actors” in general and with regard to AI in particular. For example, consider when the secretariat, a human clerk or an LLM suggests a particular legal categorisation or analysis, and the judge agrees and adopts it without much attention to alternatives or details – any public statements and formal responsibility notwithstanding. The role of advisor then blurs toward being an agent or even a trustee. What should matter for legitimacy analyses is neither the formal allocation of authority nor what the parties claim, but the de facto relationships in the ongoing judicial practices. Thus, more research is needed to address further the existence – and contestation – of such blurring of roles. Moreover, it can be important to determine which forms of a particular AI delegation are more likely to lead to such slippage: e.g., is a single nudge as an advisor more likely to lead to anchoring effects compared to a suggestion of different scenarios? Various nudging and framing arrangements create legitimacy challenges also outside AI, in ways that may illuminate the new issues wrought by AI.

Third, and importantly for our purposes, if AI-assisted human judges, or AI acting alone, demonstrably come to outperform human judges acting on their own on some such tasks, such a service may come to have a stronger claim to authority than human judges acting without AI. Such a scenario may seem unthinkable in many areas of judging; the rapid advances in AI counsel us to consider multiple scenarios. We regard it as strengths of this Razian account that, given that these scenarios are possible, the account indicates a fruitful way to address the issues and some of the conditions for such conclusions. It also allows for dynamic changes in the legitimate authority of such roles of AI over time, as AI and its tasks and services vis-à-vis human judges change.

Conclusions regarding each type of delegated role, and their various combinations, must take account of the particular types of legitimacy factors that emerge with its specific use and the phase of the adjudication cycle, to which we now turn in Sections 3 and 4. These factors may legitimate trustee approaches in certain contexts and delegitimate servant, interlocutor, advisor and agent approaches in others.

3.1 Legitimacy clusters

Criticisms of institutions couched in terms of illegitimacy often indicate the underlying reason for the critique. In the case of theories about legitimacy, the reasons are often clustered around the input, process or output aspects of exercises of authority (Follesdal, Reference Follesdal2020, pp. 481–482; Langvatn & Squatrito, Reference Langvatn and Squatrito2017; Peter, Reference Peter and Zalta2010; Scharpf, Reference Scharpf1999). What unifies this tripartite list is that these elements are the main ways that institutions can fail in their tasks and hence fail to be legitimate authorities.

To put it more formally, those subject to institutional power may challenge:

1. (1) Input legitimacy – the framework which establishes the relevant rule-making power: for instance, many criticisms of claims to authority are of the form that the authority is, in essence, undemocratic – those subject to the authority had little or no influence on the establishment, shape or selection of the relevant institutions or of the particular office holders.

2. (2) Process legitimacy – how the outcomes were arrived at: for example, that agreed procedures, hearings, etc. were not followed, that the existing practices were biased or that judges used their powers inappropriately, even if originally grounded in consent.

3. (3) Output legitimacy – the nature of the benefits from those authorities, be it that the results are meagre or unfairly distributed.

While these three aspects have a chronological character, tracking the origins and exercise of power, they should be understood as interacting yet distinct. Regardless of the field of application, these three types of legitimacy concerns may arise independently of each other, and – as we shall see – their different nature often results in questions over whether they are sufficiently commensurable for purposes of comparison and balancing.

Finally, normative legitimacy analysis must often include a fourth element: (4) sociological assurance. When an institution increases its power (e.g., judicial review to judges) or delegates it to another (e.g., a judge to a jury or AI), one additional issue regarding legitimacy is whether there is public knowledge that others are likely to comply with that institution’s decisions – e.g., for our topic, whether most subjects respect the rule of law and will obey the orders of a court. Widespread noncompliance with an institution’s rulings may challenge the normative legitimacy of a body (Follesdal, Reference Follesdal2020). This may be because some desired outcomes require general, if not universal, compliance, or because the unchecked presence of free riders lessens the incentive of others to comply or accept an institution’s power, making the collaboration highly unstable. Assurance about others’ compliance partly depends on sociological legitimacy (Gibson, Reference Gibson2007). The latter concerns whether “the governed” believe and accept that an institution has, or maintains, legitimate authority over them. In this Weberian sense, we may ask whether individuals affirm “a system of authority” and lend it “prestige,” such that obedience may follow (Weber, Reference Weber1964, p. 382).

This sociological assurance elucidates and justifies one sense of trustworthiness for the normative legitimacy of the judiciary. The social functions of peaceful adjudication and development of law require that the judiciary both (a) actually secures these core tasks regularly and most of the time – that it is worth the subjects’ trust; and (b) that the general public can determine whether this is indeed the case (Laux et al., Reference Laux, Wachter and Mittelstadt2024). This combination is often crucial to induce general compliance among “contingent compliers” who are prepared to do their share in social settings as long as they believe that the objectives and rules are fair and they have reason to believe that others think likewise and comply (Follesdal, Reference Follesdal2006; Goodin, Reference Goodin1992). Such complex public trust in the judiciary – including assurance that public authorities comply with the rule of law – is important to secure the collective good of nonviolent dispute settlement by legal methods.

We submit that these four elements often form the basis for legitimacy assessments, which in our case concerns the delegation of judicial power to AI models and systems. It is useful to parse each of these elements to identify the types of legitimacy issues that AI tends to expose.

Regarding input legitimacy, many criticisms of institutions question, for example, whether the institution and its mandate are set up in the proper way. In the case of courts, it might be whether adjudicators are appointed according to appropriate procedures and are sufficiently impartial among the parties to a dispute, or whether the laws they interpret and apply are sufficiently fair – or rather perpetuate domination of particular genders, social classes, ideologies or states. Such requirements are often phrased as resting on the free and informed consent of subjects (Buchanan & Keohane, Reference Buchanan, Keohane, Wolfrum and Röben2008). Such consent might be given directly in a contractual sense – e.g., consent to arbitration in a private law agreement. But often any such consent is at best hypothetical or implicit, e.g., contained in essence in constitutions or democratic based-law making. Input legitimacy can also be interpreted more broadly and integrate broader rule of law principles (Fuller, Reference Fuller1964). For international institutions, the central question is often whether states rather than non-state actors have consented to the delegation of power (Buchanan & Keohane, Reference Buchanan, Keohane, Wolfrum and Röben2008).

In the case of AI, there may be concerns that the use of AI models and systems – including data for them – contravenes existing laws that, for the sake of the argument, express and constitutive, from a normative perspective, the best interests of the subjects. Indeed, administrative law in many countries requires some form of human control in decision-making, while instruments like the EU’s GDPR, and to a large extent the AI Act, heavily constrain the use of automated decision-making in courts as well as reuse of data for new purposes. Use of AI by institutions surrounding the court can also raise broader input legitimacy issues, such as internal authorisation procedures. In the same way that there is an expectation that appointments of court secretariats and judicial clerks must meet some basic procedural and competence requirements (e.g., interviews for administrative officials: law degrees for clerks), there can be an expectation that delegation of judicial power to AI meets some basic administrative requirements.

A further set of legitimacy challenges for an institutional court relate to the process behind an act of authority. For instance, a central concern in the debate over the use of powers of judicial review to invalidate or direct legislation is whether courts lack the appropriate form of democratic legitimacy. They may be engaging in questions that should be determined by elected representatives (Waldron, Reference Waldron2006), broader participatory processes (Bellamy, Reference Bellamy2008) or structured deliberation (Gargarella, Reference Gargarella2022). In the case of AI, including in its use in courts and decision-making in public administration, we tend to see these concerns molded together in broader concepts like “democratic governance” of the use of AI to ensure that it is trustworthy (EU, 2019). In practice, there has been a strong focus on participatory democracy, in which critics lament the lack of citizens’ and users’ (judges/caseworkers) ability to influence when and how AI is mobilised (Broomfield & Reutter, Reference Broomfield and Reutter2022; Holten Møller et al., Reference Holten Møller, Shklovski and Hildebrandt2020). This is now complemented by a growing emphasis on representative and deliberative democracy, given the potential far-reaching consequences of AI and the complexity of decisions concerning its deployment. Buhmann and Fieseler (Reference Buhmann and Fieseler2023: 146) argue that responsible AI innovation “calls for public deliberation: well-informed ‘deep democratic’ debate that involves actors from the public, private, and civil society sectors in joint efforts to critically address the goals and means of AI”.

Besides the decisions to use (and continue to use) AI, there are a range of other process concerns. Typical aspects of adjudication that trigger normative inquiry often include issues such as transparency, impartiality and independence – which are necessary for the social functions of courts – but also broader AI issues such as cybersecurity and sustainability of energy use (Jegham et al., Reference Jegham, Abdelatti, Koh, Elmoubarki and Hendawi2025; Kayode et al., Reference Kayode, Adebola, Akerele, Fagbohun, Agbo, Bantale and Nwokocha2025). For instance, transparency demands openness in methods – e.g., court hearings should be public, and arrangements should ensure that other actors than judges do not make the final decision (Baste et al., Reference Baste, Cyndecka, Esayas, Langford, Lison and Weitzenboeck2025).Footnote ¹ On our view, this is one aspect of legitimacy which is at stake in the “black box” problem in AI-supported decision-making, with inductive machine learning methods which have occupied much attention. Likewise, potential bias in datasets and misalignment in AI algorithms raise concern about the impartiality of judges, which could be seen as a process legitimacy flaw.

In relation to the giving of reasons, we do not see it as a central part of the process which leads to the delivering of judgments. We rather see this as an essential output of the decision itself – perhaps unlike some other authors (Tasioulas, Reference Tasioulas2022, p. 13). Reason-giving by courts both enables better predictions about the legal rules in future situations and provides the requisite assurance that justice is not only done but seen to be done. The use of an AI model as part of judicial reflection and consideration – e.g., AI in advisor modus – that does not provide sufficient explainability could be seen as raising a process concern. Such a situation can be analogised to a judge accepting information provided by a clerk who is unable to explain its source. Whether the information is reflected in the final judgment as reasons is irrelevant – the concern here is processes leading to a judgment that might influence a judge in a problematic manner.

A third cluster of legitimacy objections concerns the output of the institution: that the judges fail to deliver reasoned judgments, for instance, due to a failure to identify all relevant legal norms or precedents, because their interpretations are incorrect or inconsistent, or that the mill of justice grinds too slow. The justification for AI models is often such output legitimacy, with arguments that it can improve effectiveness as well as fairness, reducing bias and arbitrariness (Tasioulas, Reference Tasioulas2023). However, three concerns tend to arise.

The first is accuracy proper. While predictive and generative models are improving considerably in many fields (Deuschel et al., Reference Deuschel, Foltyn, Roscher and Scheele2024), their performance in the field of law remains jagged and somewhat uneven – especially when there is insufficient data or doctrinal uncertainty and complexity (Siino et al., Reference Siino, Falco, Croce and Rosso2025). Moreover, it is challenging to quantify responsibly the resulting level of uncertainty and then qualify it in such a way that is actionable for use in legal decision-making (Marusich et al., Reference Marusich, Bakdash, Zhou and Kantarcioglu2023). The second is fairness. Underlying bias in the learning data for models (historical but also through reinforcement learning) may be reflected in the resulting algorithms. The empirical debate over whether COMPAS software for criminal sentencing in US courts incorporates biases against African American defendants is a case in point (Donohue, Reference Donohue2018; Kehl & Kessler, Reference Kehl and Kessler2017). The third is explainability. While there is now an array of “explainable AI” (XAI) approaches, for more advanced AI models it remains difficult to know in what sense, if any, the reasons given were in fact those “used” in a “decision.” This is a longstanding challenge in understanding human judging (as the legal realism debate attests), where AI-supported judging based on machine learning rather than pre-programmed expert rules creates new challenges, for instance, regarding what the AI input is “about” (Cappelen & Dever, Reference Cappelen and Dever2021). Moreover, there is uncertainty over whether AI can take into account both past and future contexts, which are relevant for such reason-giving, and potential and second-order effects – a concern we address below.

Finally, the need for sociological assurance underscores the importance of communicating signals to citizens that the institution is legitimate and worthy of compliance – and is regarded sufficiently as such by many others. For courts, this can include forms of transparency about procedures, etc., to show that the judiciary is indeed a “competent, independent and impartial tribunal.”Footnote ² It can also require that the judiciary is accountable in certain ways even though it is independent. It cannot be directly subject to control by the executive or the legislature – that would prevent its ability to provide the services. But it must be possible for outsiders to determine, for instance, whether the court actually applies judicial method to the appropriate legal sources in the particular case, or whether existing forms of automation exclude meaningful participation. This arguably contributes to foster general compliance with the judgments and consistent interpretations by the public and public knowledge about both. Without stable expectations about such future compliance, the judiciary cannot provide the services of dispute resolution and authoritative interpretations.

Empirical research on citizen attitudes toward AI-based adjudication confirms the need for a sufficient focus on assurance. A consistent finding is that citizens’ trust in judicial decision-making is lowered if there is no “human in the loop” (Laux et al., Reference Laux, Wachter and Mittelstadt2024). Moreover, we are prepared to accept some “human failures” more than “AI failures” (Longoni et al., Reference Longoni, Cian and Kyung2022). However, there are several important nuances to this growing body of research. It appears that for public decision-making, we trust AI less than a human, but we can often trust a human who uses AI more than a human who doesn’t use AI at all (Laux et al., Reference Laux, Wachter and Mittelstadt2024). Note that there are important variations. Sometimes, minority groups are more likely to trust machine-based decisions given previous patterns of discrimination, while citizens in countries with weaker rule of law are more likely to trust AI compared to human decision-making. Recent research in Europe also indicates that citizens are less concerned about AI if it speeds up decision-making and more concerned about the transparency of algorithms rather than whether it is publicly or privately owned (Bentsen & Johannesson, Reference Bentsen and Johannesson2026). Moreover, the importance of explanation diminishes somewhat if there is an opportunity to appeal. This all suggests that the precise forms of delegation and the jurisdictional context will matter a lot, and that the perceived importance of a “human in the loop” may well change over time.

3.2 Cluster conflicts

Even though many agree on these four clusters or elements, several questions are unanswered: is this just an eclectic collection of considerations or is there some unifying perspective that might also help in cases where the various elements need to be “balanced” or traded off against each other (Buchanan & Keohane, Reference Buchanan, Keohane, Wolfrum and Röben2008)? For instance, can “input” legitimacy be sacrificed completely if “output” is excellent? For example, can more reliable and swifter AI-assisted decisions be more legitimate than AI-free judicial processes – even though they are less transparent? Or must each element meet a threshold for the institution to be legitimate?

Recall that we regard these clusters as the principal ways by which institutions can fail in their tasks and hence fail to be legitimate authorities (Follesdal, Reference Follesdal2020). At least four issues deserve attention. The first is to be aware of the fallacy of accumulation. While these elements are interconnected, they are not simply cumulative: A complete lack of input legitimacy cannot always be compensated by better output legitimacy (Follesdal, Reference Follesdal, Grossman, Cohen and Follesdal2017). In the case of AI, Li et al. (Reference Li, Qi, Liu, Di, Liu, Pei and Zhou2023) make the same point: “AI trustworthiness cannot be achieved through hillclimbing on a set of disjoint criteria.” To avoid this fallacy of accumulation or “hillclimbing,” many argue that at least some threshold of each element must be met, and we would agree with this (Goodin, Reference Goodin2013; Follesdal, Reference Follesdal2006; Taylor, Reference Taylor1987).

There is likewise disagreement over the “weighting” of each element. For some, the minimum or “gold” standard is process legitimacy, and in particular democracy (Peter, Reference Peter and Zalta2010, pp. 15–18) – meaning that, at the very least, the threshold for this aspect may be higher than for the others. In our context, this seems to be the assumption for many normative scholarly analyses of the use of AI in courts which emphasise respect for democratic principles and process, including inclusivity in design (Laux et al., Reference Laux, Wachter and Mittelstadt2024); and has been affirmed in some recent empirical studies (e.g., Magalhães et al., Reference Magalhães, Arnesen, Kern, Koenig, Schiff and Tyler2026). In contrast, the EU’s (2019) definition of trustworthy AI is silent on this issue, although it does pay more attention to the rule of law. To be sure, it may be argued that a rigid process threshold, requiring fully fledged legitimate processes, is not needed in all circumstances. When normatively urgent substantive interests are at stake, we may also have reason to protect them against certain procedural concerns. Nonetheless, such sacrifices can have problematic long-term consequences, and, under these conditions, other standards of process might be imposed: e.g., transparency, due process, reason giving or consultation, or outputs that help enforce democracy (Bodansky, Reference Bodansky1999).

The second is the challenge of ontological commensurability. The intrinsic qualities of the different aspects of legitimacy make it difficult to undertake proportional “balancing” for trade-offs. Even when we can agree on the relative importance of each legitimacy cluster, it is challenging to create a single metric for a balancing exercise, a common challenge in the mixing of any method (Beach & Kaas, Reference Beach and Kaas2020). To illustrate: it is often not possible – or indeed meaningful – to say that an institutional design with 10 percent more output legitimacy is as legitimate as an institution with 10 percent more process legitimacy. In the case of AI, there have been attempts to develop metrics that permit these calculations for selected elements of trustworthiness. For example, Giudici et al. (Reference Giudici, Centurelli and Turchetta2024) provide commensurable ranking metrics for robustness, accuracy, fairness and explainability in AI models which permit analysis of trade-off. However, the model excludes other important aspects of trustworthiness, such as democratic governance, privacy, safety and environmental sustainability. And issues of how to “weigh” each of these factors remain to be addressed. In practice, commensurability may be achieved through rough triangulation (e.g., through comparing simple ordinal-like units like “none,” “none+,” “small-,” “small,” “small+,” “some-,” “some,” “some+” and “large-” (Kern, Reference Kern2018)), or some other incomplete quasi orderings (Sen, Reference Sen2017); pragmatism (each cluster is considered on its “own terms” and put subsequently into a conversation with the other (Tait & Taylor, Reference Tait and Taylor2022)); or some form of democratic process based on representation, participation and/or deliberation. We here note that our legitimacy framework leaves these issues as open questions for further attention, including for which purposes any such single metric would be helpful.

The third is tensions between the clusters. Even if we could measure in a commensurable way each cluster, a feature of legitimacy analysis has been the inescapable dilemmas that arise in institutional design. There may be internal conflicts between and within a cluster that are difficult to navigate and require inescapable trade-offs. To illustrate, Dunoff and Pollack (Reference Dunoff and Pollack2017) argue that there are trilemmas of institutional design for international courts which prevent them from being fully independent, transparent and accountable to states. In the case of AI, such trustworthy dilemmas have garnered increasing attention in the past 5–10 years (Li et al., Reference Li, Qi, Liu, Di, Liu, Pei and Zhou2023). For instance, improvements in explainability often come at the expense of accuracy, as deep learning models are controlled to ensure greater transparency and interpretability. Or improvements in privacy can come at the cost of nondiscrimination and fairness, as the masking of personal data removes the possibility to correct for bias. Thus, not all elements or clusters of legitimacy seem to be required in all cases. Rather, close scrutiny of the development of AI combined with systematic normative arguments is required to understand how well the various clusters of rules ensure justifiable institutions.

This makes it necessary to consider whether we can and should speak of more or less legitimacy in such contexts. The aim may be to improve on, satisfice or optimise legitimacy in a certain context without attempting to maximise legitimacy or meet a high threshold for each element. In our view, it is possible to speak of and discuss whether delegation of judicial power to AI enhances or diminishes legitimacy of adjudication, especially when a legitimacy assessment is carried out in a particular context. Given that a singular policy, rule or action can enhance one aspect of legitimacy yet disturb others, determining whether legitimacy is increased or decreased requires often a full assessment of the various aspects and how they fit together in the particular case. AI may thus often improve efficiency, reduce human bias/fallibility and improve access to public goods, but it encounters questions over a potential weakening of input, process and output legitimacy together with increased sociological uncertainty due to the potential for free riders. We should especially determine how much such conflicts matter from a normative legitimacy perspective. For instance, it may be less problematic if the various standards for input and process legitimacy must be satisfied to some threshold level rather than maximised. Thus, more attention may be appropriately directed to these inescapable trade-offs which require optimisation solutions (subject to thresholds) or hard consequence-informed choices.

The final challenge concerns lack of output legitimacy. For instance, a classical question is whether an institution can be legitimate if it does not generate sufficient benefits. A state may have impeccable democratic credentials but is it legitimate if its citizens are significantly and avoidably deprived? In the case of AI, one is often faced with the question as to whether AI must be deployed if it provides clear benefits, including for adjudication. It is reasonable to presume that some output standard of justice and/or the common good must be met. One important question is what standard. In the context of the debate on judicial review, the debate on this question is complex. Some argue that courts must at least ensure avoidance of “extreme injustice” (e.g., a small core of civil rights and minimum subsistence) (Buchanan & Keohane, Reference Buchanan, Keohane, Wolfrum and Röben2008, p. 44), others claim that court must ensure more adequate outcomes (good governance, social trust, significant protection of rights and sufficient education) (Kane & Patapan, Reference Kane and Patapan2010, p. 605), yet others make thick demands for optimal outcomes or full rights realisation (Peter, Reference Peter and Zalta2010, pp. 13–14). In the case of AI, the technology is still too nascent to gain a sense of its full potential, but it is likely that the duty to use AI in courts could be advocated as it becomes ever more powerful and effective.

Two preliminary conclusions from this discussion are that the legitimacy of institutions depends on these four factors (input, process, output, assurance), but the assessment is not simply monotonic: neither more democratic control, more transparency and due process, nor more beneficial outputs will always increase the legitimacy of the institution. And legitimacy assessments often require a detailed understanding of the intricate details of rules and actual practices within the institution and its complex interconnection within a web of other institutions.

4.1 Legal process

Legal process can include all aspects of a case that have a strong administrative dimension, whereby the steps taken have an indirect or direct bearing on the content of the substantive decision. This can include case filing, allocation of judges, holding of interim hearings, ordering of injunctions based on clear predetermined criteria,Footnote ³ transcribing proceedings, interpretation and translation of foreign languages, translating evidence and other sources and summarising and publishing judgments. In the legal process, AI can be deployed, for example, to extract important case details at various stages (Kolkman et al., Reference Kolkman, Bex, Narayan and van der Put2024, p. 7) and cluster and/or analyse cases when they first enter the court system, suggesting an order of importance or urgency or recommending an assignment to the appropriate court, or assist with similar and repetitive cases that often overwhelm courts (Fikfak & Helfer, Reference Fikfak and Helfer2025, pp. 85, 99; Fonseca, Reference Fonseca2025).

In several of these legal process areas, the use of AI tends to be regarded as largely “uncontroversial” for AI (Fikfak & Helfer, Reference Fikfak and Helfer2025, p. 93). During the last decade, AI has been adopted for such functions and, with the arrival of generative AI, its usage has accelerated (Langford, Reference Langford, Zalnieriute and Limante2026a). In some instances, it can be argued that deployment of AI in legal processes contributes to enhancing significantly process and output legitimacy. For example, in Norway, if AI were used to transcribe proceedings in courts (which currently does not already occur), it could improve the quality of judgment writing and public accountability. Likewise, the use of AI to de-identify personal data more quickly could allow the publication of most, rather than a small proportion of, judgments (Baste et al., Reference Baste, Esayas and Langford2024).

However, there is also some dissent. Satterthwaite (Reference Satterthwaite2025) notes that LLMs are often poorly trained on small – so-called “low resource” – languages, which raise risks when AI is used even for seemingly simple tasks, from fact summaries to translation and transcription. Given that most litigation is fact-intensive, in that facts are often more determinative of the outcome than a legal interpretation, these risks should not be underemphasised. Small errors can be decisive, suggesting that highly automated trustee models of delegation without any human control – whether in innovation processes or in actual practice – can be problematic. Further challenges arise in highly automated systems that are not adaptive to changes in procedural law and the broader ecosystem of litigation. Procedural uses of AI tied to legal rather than administrative aspects of process require greater vigilance; and abrupt changes in nature and forms of claims also require reflexivity. Indeed, such sudden changes can or will be wrought by AI itself. For instance, several public administration agencies report that in certain areas of law there has been an exponential increase in the volume and length of complaints since generative AI was made publicly available, while LLM-generated deepfakes for court evidence are a growing problem (McConvey et al., Reference McConvey, Burkell, Deng, Ebrahimi, Grossman, Mahdavimoghaddam and Bagheri2026).

4.2 Inventorising

Ahsmann (Reference Ahsmann2011) describes an intermediate process of decision-making as “inventorizing.” In this phase, the judge “reads the case file to determine what it is about, what has been claimed, and what legal questions are involved” and “determine the applicable rules and case law, list the relevant facts, and consider the propositions put forward by the claimant and the defence” (as summarised by Kolkman et al., Reference Kolkman, Bex, Narayan and van der Put2024, p. 3). The extent to which this phase of issue and rule identification and factual presentation is driven respectively by parties/lawyers and judges varies between jurisdictions and different areas of law. Ultimately though, judges will need to decide on the inventory that they take as relevant in any oral hearing and eventual decision/s.

Early AI models purported to assist legal research in specific domains (Hill, Reference Hill2025), while domain-specific predictive models, such as Legal-BERT and RoBERT, performed well in identifying arguments within the European Court of Human Rights corpus (Al Zubaer et al., Reference Al Zubaer, Granitzer and Mitrović2023). However, the “democratisation” of generative AI with LLMs and their steadily increased power has seen their greater use by parties and judges. This has fostered rapid inventorising of relevant legal rules, which can help judges identify and cluster similar cases/applicable laws, and potentially enable a judge to see new patterns and subsume the case within new legal categories. However, the performance of such models “in the wild” can vary significantly given their dependence on underlying judicial data. The challenge of hallucinations is well known, with a new global database documenting the incidence of fictional citations in judgments (Curran et al., Reference Curran, Lansley and Bethell2023; Charlotin, Reference Charlotin2026). Even when models are trained on domain-relevant data with routine augmentation, Magesh et al. (Reference Magesh, Surani, Dahl, Suzgun, Manning and Ho2025, p. 216) find that AI research tools made by LexisNexis and Thomson Reuters each hallucinate between 17 percent and 33 percent of the time. This suggests that there should be significant caution in using current AI models to legally inventorise in an agent or trustee role. AI models are arguably most useful for identifying potential legal rules and arguments in a servant, interlocutor or – with some caveats – an advisor manner. We might assume that “false positives” will diminish over time – that phantom rules and precedents will not be offered – and that courts and judges will build up peer-reviewed and experiential knowledge on when AI can operate as an agent in such contexts.

The same conclusion on the dangers of trustee/agent uses of AI arguably applies to the identification and collation of the relevant issues and facts. Issue identification concerns the relevant facts in a case – i.e., what questions are legally relevant in the case? It partly precedes the identification of legal rules but is also an iterative process that occurs in dialogue with legal rules. Solid legal training to “see like a lawyer” is crucial. Issue identification is thus a skill that is often tested later rather than earlier in legal education, especially in more complex fact and legal scenarios. Nonetheless, as an advisor, AI may assist in identifying unseen but applicable patterns in the case law due to human blind spots. However, human control is often needed given AI’s blind spots, especially its reliance on past cases – the problem of status quo bias (Fikfak & Helfer, Reference Fikfak and Helfer2025, p. 102).

In relation to the listing of relevant facts, current AI technologies can certainly help in organising and summarising facts. Beyond such time-saving uses, its most powerful role might be in helping judges understand facts, especially given that legal training involves little focus on evidence and that courts often must cover every societal domain. Nonetheless, the familiar challenges of accuracy (hallucination) and selection bias (absence of training data) should be taken into account. The greatest danger may be cognitive offloading or automation bias (unreflexive deference to AI-generated responses) in cases with voluminous facts or many cases with similar facts (Gerlich, Reference Gerlich2025). High levels of delegation can diminish judicial understanding in a particular case or cases over time as judicial learning and development is restricted.

4.3 Problem-solving

The next phase of adjudication concerns what we may think of as problem-solving. It involves attention to the facts – both their “selection” (establishing them and listing the points of dispute) and their “assessment” (determining whether relevant facts have been stated and if these facts are sufficiently substantiated) – as well as the decision, including remedies and sentencing (Ahsmann, Reference Ahsmann2011; Kolkman et al., Reference Kolkman, Bex, Narayan and van der Put2024). Problem-solving can also include determinations of the law when its content is not clear. Such cases are often “hard cases,” arising “when no settled rule dictates a decision either way” (Dworkin, Reference Dworkin1974).

In our view, problem-solving also includes the necessity to harmonise second-order considerations. Consideration of the systemic effects of a general legal interpretation, subsumption (application of law to facts) or a particular remedy is often a key part of judicial reasoning (Fuller & Winston, Reference Fuller and Winston1978; Ruhl et al., Reference Ruhl, Katz and Bommarito2017). Legal solutions can affect future cases or other areas of law, while remedies and punishment can have significant consequences for public policy, individuals and corporate entities (Horowitz, Reference Horowitz1977). To be sure, some admonish courts to largely abstain from these sorts of considerations (Vermeule, Reference Vermeule2006), especially the “creative” part, as attested by descriptions such as “legislation by judges,” “judicial activism,” etc.

However, harmonisation is often unavoidable for courts, even if it is not always acknowledged explicitly. Such instances particularly arise when new challenges arise that disturb the current legal equilibrium, be it due to social change, new technologies or reforms in other parts of the law. Most thus agree that some harmonisation by a court is sometimes necessary if it is going to be able to deliver legally sound judgments. The important disagreements concerning the harmonisation by courts thus should instead be about how often and how much it occurs.

This process of harmonisation – analogous to “reflective equilibrium” (Rawls, Reference Rawls1971) – often involves coming to see new patterns in perceptions, which seem better justified than those of the past. As MacCormick (Reference MacCormick1990, pp. 557–558) states:

The key challenge is then how to govern or guide this judicial role to reduce risks of domination by unaccountable adjudicators. Who should have the authority to decide when adjudicators should not stretch interpretations, or how to “rebalance” adjudicators’ priorities among legal norms? From the perspective of legitimacy, three points appear central to such discussions: who performs these processes of discovery and creation; how their quality can be monitored; and how the processes or their outcomes can be changed, and by whom.

Using AI in problem-solving can bring many advantages to judicial problem-solving, whether as advisor or agent. AI can be used to identify previous bias (Piccolo et al., Reference Piccolo, Katsikouli, Gammeltoft-hansen and Slaats2023), correct for human heuristics in the assessment of law and facts (Piperides et al., Reference Piperides, Allen, Dhami, Flessner, Hastie, Koehler, Wagner, Gigerenzer and Engel2006) and provide better access to legal information (Satterthwaite, Reference Satterthwaite2025; Sunde, Reference Sunde2023). Indeed, some studies of “legal reasoning” in LLMs show a tendency in the models toward a legal formalism that is less affected by sympathy for one of the parties (Posner & Saran, Reference Posner and Saran2025), and AI may thus reduce mistrust that adjudicators are biased. AI systems may also assist in enhancing judicial quantitative skills, such as calculation of damages and penalties (Ruohonen & Hjerppe, Reference Ruohonen and Hjerppe2020). Moreover, they may help navigate complex cases, in which levels of epistemic uncertainty are high due to the complexity of the problem caused by multiple, interconnected and dynamic elements (Funke, Reference Funke2010). AI together with data science methods like network analysis and agent-based models can enable “systems thinking” by enabling formal or informal modeling of consequences of different legal approaches (Ruhl et al., Reference Ruhl, Katz and Bommarito2017; Schwartz, Reference Schwartz2020).

However, the limits of AI in judicial problem-solving tend to be visible as the levels of complexity increase in any central aspect of a legal problem. This is generally a challenge for LLM models addressing any complex problem-solving (Shojaee et al., Reference Shojaee, Mirzadeh, Alizadeh, Horton, Bengio and Farajtabar2025). In the case of judging, it is doubtful whether AI models alone can navigate multiple legal sources, address multiple and incommensurable factors in subsumption (e.g., proportionality analysis) or consider the distribution or magnitude of second-order effects. Several issues must be addressed by further research. First, it is unclear whether AI can or could perform the creative work – the sorts of conceptual/perceptual shifts and paradigm changes – involved in the reflective equilibrium or harmonisation processes. For example, in the UK, judges suddenly overturned established law that a husband cannot be guilty of raping his wife (UK, Reference UK1991). Second, in some complex cases, it is arguable – contrary to Dworkin (Reference Dworkin1986) – that no unique right answer can be found by proper judicial method, despite best efforts, be it with human or artificial Herculean cognition. For example, should the right to strike or the right to move goods across state borders in the EU have priority (De Vries, Reference De Vries2013)? In such cases, it is arguably human presence that is currently important in arriving at a legitimate decision. AI models – even as they improve – should be most likely used in interlocutor or advisor modus in addressing complex cases. This is because it is adjudicators who are currently authorised and made publicly accountable. To be sure, other branches of government may adjust such legal norms over time, but it is currently unclear what would be justifiable allocation of such authority with AI.

Moreover, when it comes to subsumption, even if LLMs are potentially less biased, they are limited in their sensory inputs, and it may be that the more visceral judicial experience of the courtroom provides additional information that may guide a fairer outcome (Posner & Saran, Reference Posner and Saran2025). Bias may arise in other aspects of problem-solving, such as the determination of sentencing as the Compas experience shows (Neil & Zanger-Tishler, Reference Neil and Zanger-Tishler2025). This suggests that AI models that seek to highly automate just one aspect of problem-solving or many aspects in very simple cases – from factual and legal assessment to subsumption and conclusions – should be subject to significant testing and strong and affordable rights of review. There is a risk that important information, including outliers and exceptions, is lost in the process.

4.4 Reasoned judgment

Much critical attention on the judicial use of AI has focused on explanation (Deeks, Reference Deeks2019), which is necessary for authoritative interpretation and justifying findings of fact, subsumption and remedies/sentencing. Richmond et al. (Reference Richmond, Muddamsetty, Gammeltoft-Hansen, Olsen and Moeslund2024, p. 6) argue that AI must meet at least three specific explanatory functions to count as legal reasoning: it must be comprehensive (“supply evidence, support or reasoning for each output”), comprehensible (“provide explanations that its users can understand”) and correct (an explanation should accurately reflect the process the AI system used to arrive at the output). However, the particular legal demands for explanation can vary significantly across different jurisdictions and fields of law.

Earlier classical rule-based AI models fulfilled these criteria to a certain or large extent, while struggling to ensure accuracy (Ashley, Reference Ashley2017). The same could be said for simpler quantitative models such as linear and logistic regression, general additive models, Bayesian models and decision trees, in which the relevant factors are specified in advance (Burrell, Reference Burrell2016). The “black box” problem has, however, bedeviled more recent and powerful AI models – whether predictive or generative AI. Reason-giving is not an intrinsic part of their design, as they extrapolate the most likely text based on past data.

This challenge is not unique to AI. There has been a similar and long-standing critique of juries as they rarely give reasons (Stern, Reference Stern2025). As a result, many government and national court systems have been under pressure – including from international human rights courts – to justify this delegation of judicial power for determining factual questions (Roberts, Reference Roberts2011). This has led some countries to “circumscribe or abolish their systems of lay participation” (Hans et al., Reference Hans, Diamond, Kutnjak Ivković and Marder2024, p. 141).

The broader field of “XAI” has sought to enhance reason-giving through different “black box” methods (Hassija et al., Reference Hassija, Chamola, Mahapatra, Singal, Goel, Huang, Hussain, Spinelli, Mahmud and Hussain2024; Linardatos et al., Reference Linardatos, Papastefanopoulos and Kotsiantis2020; Roscher et al., Reference Roscher, Bohn, Duarte and Garcke2020). These may be post hoc – e.g., through Shapley values and LIME (Local Interpretable Model-agnostic Explanations) approximations – or counterfactual examples that interpret through selecting instances of the dataset, although challenges remain in identifying causal effects in complex AI systems (e.g., deep learning), accounting for inferences across domains and sub-domains, and communicating meaningful explainability (Jullum et al., Reference Jullum, Redelmeier and Aas2021; Miller, Reference Miller2019; Richmond et al., Reference Richmond, Muddamsetty, Gammeltoft-Hansen, Olsen and Moeslund2024). Addressing these challenges has been enhanced by the recent wave of “AI reasoning models” that are designed to “think” through problems using step-by-step logic, using methods such as chain of thought pretraining, backtracking and “self-reflection,” and greater use of reinforcement learning (Chen et al., Reference Chen, Qin, Liu, Peng, Guan, Wang and Che2025; Wang et al., Reference Wang, Liu, Wang, Li, Wang, Yan and Xu2026). However, questions still remain over the extent to which these methods sufficiently meet high legal explainability standards in decision-making, raising questions over the degree of output legitimacy and sociological assurance they create (Barry, Reference Barry2024; Richmond et al., Reference Richmond, Muddamsetty, Gammeltoft-Hansen, Olsen and Moeslund2024).

From a systemic perspective, explanation also plays an additional role. If the concern is to avoid favouring one party to the dispute, AI may be a helpful mechanism, though other concerns must also be considered, such as attention to the societal impact of the interpretation, whether such choices can be monitored, so that legislatures may correct what they regard as misinterpretations by judges. In particular, as a judge has normative power to determine what is valid interpretation, they will sometimes set a new precedent. As Tasioulas (Reference Tasioulas2023, p. 34) states, “the explanation of its decision-making obviously does not consist in the application of existing law to the case at hand.” Thus, explanation is not only important for a particular case (Richmond et al., Reference Richmond, Muddamsetty, Gammeltoft-Hansen, Olsen and Moeslund2024) or ensuring judicial accountability and assurance (see below), but it also provides the service of prediction, giving reasons for a departure from existing precedent so that it can be extended to other similar cases. Are there reasons why subjects should grant AI similar authority?

4.5 Legitimation

The fifth aspect of judging concerns assurance that the task is well done. Obviously, transparency about premises and explicit legal reasoning is an important aspect of providing assurance that the various and necessary elements of a judgment are sufficiently in place, such that it deserves compliance and that the court deserves general trust (Voeten, Reference Voeten2012). Thus, courts have increasingly devoted greater attention to public communication (Spencer, Reference Spencer2017). In addition, scholars observe that most courts are concerned with their sociological legitimacy – if not, survival – meaning that, in making judgments or imposing remedies, they tend to lag changes in public opinion or seek to triangulate between opposing societal interests to maintain their judicial capital (Roux, Reference Roux2009; Tushnet, Reference Tushnet2008).

At the same time, there is a strong public preference for effective decision-making in courts and public administration, such that weaknesses in process legitimacy may be partly traded off against gains in output legitimacy. This may be the case with AI. Empirical evidence shows that speedy adjudication is valued highly by citizens (Laux et al., Reference Laux, Wachter and Mittelstadt2024), with effectiveness sometimes even preferred over fairness (Bansak & Paulson, Reference Bansak and Paulson2024). Thus, if AI can improve the efficiency and effectiveness of courts, then this might lower the demands on process legitimacy. However, some process legitimacy thresholds must be arguably met. Even in this respect, AI may help by providing more transparency about whether a process and decision are within the “normal” range, whether sources and precedents are correctly applied, etc. Thus, it can provide more “second opinions” and enhance the task of peer control by other jurists. Moreover, AI may make it easier for affected parties to decide whether to appeal. Over time, AI systems could develop in a multi-agent direction, so that different AI systems monitor each other and could signal to observers – whether there is consensus about the outputs or not – as certain observatory organisations do today.

Even assuming that phantoms and hallucinations are somehow eliminated in later versions, it is less clear which contributions AI can provide where transparency is required concerning AI models’ training and testing data, and their underlying algorithms and evidence on reliability, validity and statistical regularities. Managing legitimacy in complex cases with high levels of epistemic uncertainty, or in controversial cases with high levels of legal discretion and moral uncertainty, may remain challenges for AI in the modes of delegation. These cases require high levels of reflexivity and discretion in negotiating the legal, political and communicative space that judges operate within (Langford, Reference Langford2026b). Attaining this type of competence is already a challenge for embodied human cognition (Dixon, Reference Dixon2023), and it is less clear how artificial cognition can master the nuances of such contextual multifaceted thinking and decision-making. Thus, it is notable that most empirical surveys tend to find little support for purely AI-based decision-making in courts (Laux et al., Reference Laux, Wachter and Mittelstadt2024), showing that the trustee models currently lack sociological assurance or legitimacy.

4.6 The legitimacy challenge of discretion

One aspect of the judicial function which poses particular challenges for legitimacy assessments of AI is the normative power of discretion. Judges exercise discretion in several of the stages of cases discussed above. This includes inventorising (how to (re)classify or subsume a case under some category), problem-solving (how to “balance” and “harmonise” legal norms within a range consistent with proper judicial method and rule, specifying a remedy or punishment) and giving reasoned judgment (based on some decision about which explanations are the most relevant to articulate).

From a legitimacy perspective, it is crucial to ask questions concerning input (are judges now authorised to share or even delegate such discretion to AI?) and process (should they do so?). Notably, until a few years ago, output legitimacy would have been an equally central question in most cases of the use of AI, as there was considerable doubt (including with first-generation LLMs) over whether AI could engage in even standard legal interpretation and application which involved aspects of discretion. The improved predictive and reasoning abilities of AI, especially where there is sufficient training data and less need for creative problem-solving, mean that the discussion has shifted to focus more on the input and process dimensions – although certain output aspects remain central, such as explanation.

It may be useful to analyse the question of discretion by beginning with the initial delegation of adjudicative power to courts. Courts typically act as trustees for the parties to a dispute. As subjects of the courts, the parties have a shared overarching and second-order preference for an (sufficiently) impartial arbiter to settle their first-order dispute by particular means – legal sources – according to processes of judicial methods. An important part of these methods is that judges are authorised to exercise discretion in interpreting and applying the law and imposing remedies. To be sure, there are not only methodological limits to their discretion: subjects may also revoke the authority enjoyed by the domestic or international court through legislative or constitutional changes or treaty amendments, respectively. Consider what happens when judges delegate parts of these trustee tasks to AI. Two challenges arise, even in the best of cases in which AI, over time, has actually promoted appropriate judicial values. First, how can the judges using AI be sure that AI aligns dynamically with the underlying judicial values and preferences over time for this purpose (an issue of output legitimacy) (Abiri, Reference Abiri2024)? Second, it is not at all clear whether judges can or should be able to delegate their discretion regarding interpretations and remedies without loss to the authority of resulting decisions (an issue of input/process legitimacy). What reasons are there for subjects of the court to replace their other reasons for action and defer to the discretionary decisions not of a judge but of AI? That is, do subjects have reason to accept such shifts in who exercises discretion and how, so that these modified authorities remain legitimate – or gain or lose legitimacy?

We underscore that whether the judges’ normative power requires humans in the loop is not simply a conceptual issue. Instead, it seems to be a received, often shared assumption – though with exceptions. Our point here is limited to suggesting that the (tacit) agreement to defer to judges’ discretion must be rethought to determine whether those or other arguments also apply to nonhuman agents, against a backdrop of which are the best alternatives. For instance, one reason to rely on the discretion of impartial judges in interpretation is that the parties may prefer that sort of bounded arbitrariness to being subject to decisions ultimately controlled by one of the parties to the conflict. This may, for instance, be the reasoning behind states’ decision to agree to “incomplete contracts,” which sometimes explicitly and implicitly authorise arbitrators to exercise discretion (Pauwelyn & Pelc, Reference Pauwelyn and Pelc2022). Note that this concern may be equally well secured by the mechanical solution of flipping a coin in cases where legal considerations “run out.” However, perceptions also matter (Keren & Teigen, Reference Keren and Teigen2010). Similar avoidance of bias by the bounded arbitrary exercise of discretion may be secured by relying on AI – if AI can be trusted to be impartial. Indeed, AI may well come to be seen as more reliably impartial, and that may in turn support attitudinal changes. But again, AI might not gain such a reputation.

A second reason to accept judges’ discretion may be the recognition that categorisation, interpretation, etc. must heed several somewhat conflicting considerations – simplicity, consistency, predictability, legislators’ intent, etc. – in some sort of (reflective) equilibrium. And the vetting process of judges combined with deliberative processes will arrive at some non-objectionable judgment which incorporates all such considerations. As Fuller (Reference Fuller1964, pp. 41, 45) argued, the simultaneous pursuit of all aspects of the rule of law is unlikely to be feasible, making it “necessary to pursue a middle course which involves some impairment” of the relevant values or “desiderata.” Similar signaling, proxies and mechanisms of legal reasoning are hitherto absent from AI based on text prediction, as discussed above, even if AI is able to propose optimisation scenarios and – perhaps – increasingly explain them. Again, such features and requisite transparency may change in the near future. What may be lacking is a sufficiently overt authorisation by the subjects of the law to AI to exercise discretion alone or in conjunction with human judges, in ways that carry normative power.