Research Article | 
March 2021

The challenges of artificial judicial decision-making for liberal democracy

Christoph Winter

The application of artificial intelligence (AI) to judicial decision-making has already started. Significant progress has been made, not only in the United States, most prominently with regard to bail decisions,1 but also in Russia2 and Mexico.3 China has placed more than 100 robots in courts offering legal advice to the public,4 and Estonia is piloting a program in which small scale civil suits are decided by an algorithm.5 Furthermore, a recent cross-cultural survey suggests that legal scholars believe that, on average, almost 30% of judicial decision-making will be carried out by AI in only 25 years’ time, tripling their estimate that the current role of AI accounts for less than 10% of judicial decision-making.6 Against this background, it seems plausible to assume that future advances in AI will revolutionize the judicial sector. To many, AI not only promises greater fairness, justice, and legal certainty, but it may finally satisfy the legal requirements of the internationally accepted concept of a fair trial, in particular with regards to access to justice, as recognized by, among others, Article 6 of the European Convention on Human Rights (ECHR), Article 10 of the Universal Declaration of Human Rights (UDHR), and Articles 7 and 25 of the African Charter on Human and Peoples’ Rights (ACHPR).7

At the same time, issues of discrimination and transparency have emerged in the context of bail decisions, putting liberal democratic principles under pressure. Although intense debates about the criminal justice system and demands for what has been coined “algorithmic fairness” have been ushered, further challenges that arise from the shift in authority from human intuitions to artificial intelligence within the judicial system remain unclear and neglected in the discourse. As of now, there has been no systematic analysis of the risks to liberal democracy from implementing AI into judicial decision-making. This article therefore sets out to fill this void by identifying and engaging with challenges arising from artificial judicial decision-making.

In this way, this analysis not only offers a first attempt to systematize these challenges, but it also outlines a number of crucial issues which demand further research. Throughout the analysis, I will apply a comparative perspective between human thought processes and artificial intelligence,8 with an emphasis on ethical issues arising from (I.) large-scale discrimination and (II.) a potential lack of transparency within the judiciary. Indeed, I will argue that these issues pose some short- and medium-term threats. However, drawing most notably on recent research by Kleinberg and colleagues concerning bail decisions,9 I am confident that related technical and ethical issues can be solved to yield outcomes better than those from human decision-making, from a wide range of philosophical perspectives on discrimination. While the lack of transparency in AI decision-making, narrowly construed, does pose risks which ought to be addressed, AI can also increase transparency in another—arguably more important—domain, which I will refer to as transparency of options. The far greater long-term threat for liberal democratic values lies in (III.) the overall shift of power from human intuition to advanced AI and, more precisely, in the possibly accompanying threats to judicial independence and the separation of powers. Accordingly, an AI-based judiciary might significantly contribute to the rise of digital authoritarianism.

Before engaging with these questions, it is necessary to clarify and limit the scope of this analysis. A common framework within the discourse is the distinction between whether AI will supplement or replace human beings.10 Will AI help human judges to detect biases or make better decisions, or will it replace human judges altogether? Furthermore, the discussion in law and policy is usually focused on narrow applications of so-called AI that use machine learning or simple decision trees to supplement or partially replace human decision-making. I will concentrate on these partial replacements and corresponding issues of discrimination and transparency in the first two sections of this chapter. However, in the final section I will move on to more advanced forms of AI that could replace the vast majority of human judicial decision-making, thereby going beyond present applications, possibilities, and the main focus of literature. Even though this contemplates advanced forms of AI understood as significantly exceeding current capabilities, one should note that this would not require superintelligence that “greatly exceeds the cognitive performance of humans in virtually all domains of interest.”11 Instead, as I will explain later, it would require the development of an advanced artificial judicial intelligence (AAJI) with more limited domain knowledge and capabilities.

A prevailing argument among researchers and politicians alike seems to indicate that one should focus on present issues rather than trying to solve those which may occur in the future.12 However, research about future developments, such as the implementation of more advanced forms of AI into the judiciary and its entailing risks is crucial because outlining future risks may provide valuable information on what sort of developments ought to be followed with particular care. If, for instance, great risks for liberal democracy are likely to occur in the future as a result of the application of AI, one may try to find ways in the present to mitigate such risks, if one considers liberal democracy to be a desirable political system. To put it simply: it can be beneficial to avoid an encounter with a predator in the first place, rather than trying to run away when it is already in front of you.

I. From Human Bias to Algorithmic Fairness

This analysis will first focus on equal treatment for all citizens which, arguably, is the distinctive part of liberal democracies.13 Much of the academic14 and popular15 debate on AI in the judiciary has been dominated by issues of illegitimate discrimination based on sex or race. This is understandable, given that Google’s face recognition algorithm labeled black people as gorillas,16 algorithms employed in job selection decisions favored white males,17 and white people benefitted more than black people from the COMPAS18 recidivism algorithm,19 among others. It is clear that improvements are needed. Yet, it is a separate question whether these issues justify the abolishment or omission of incorporation of AI into the judicial sector. After all, the much longer history of human intelligence (HI) in the judiciary seems to indicate that human decision-making has not done justice to all groups either. While it may be possible to mitigate discrimination resulting from implicit human biases to some degree,20 we need to identify how feasible it is for AI to overcome such biases compared to HI. If there is a sufficiently good chance that AI is even better positioned to do so in the long term, one should not reject entirely the idea of artificial judicial decision-making due to discrimination, but rather concentrate on how to mitigate discrimination by AI. For AI to play an important role in the judiciary of the future, it does not have to be perfect, but only better than HI. Indeed, we may even be obligated to adopt it in such cases.

Pertaining to discrimination, the law works remarkably similar across jurisdictions in distinguishing between direct and indirect discrimination, also understood as disparate treatment and disparate impact respectively.21 Direct discrimination is typically understood as treating a person less favorably than another person in a comparable situation on the basis of any protected ground, such as sex, racial or ethnic origin, religion, disability, age, or sexual orientation.22 Hence, the disadvantageous treatment is based on the possession of specific characteristics. Here, cases turn on the analysis of the causal link between the protected ground and the less favorable treatment,23 as well as the comparability of persons in similar situations. In some jurisdictions, direct discrimination can be justified,24 in others it cannot.25

Indirect discrimination takes place when an apparently neutral criterion, provision, or practice would put persons protected by some general prohibition of discrimination (e.g., those having a protected characteristic) at a disadvantage compared to others.26 If such discrimination occurs, it will have to be justified, i.e. the provision in question will be upheld only if it has a legitimate aim, and the means of achieving that aim is necessary and appropriate. In contrast to direct discrimination, indirect discrimination is based on apparently neutral criteria which are not formally prohibited. Nevertheless, the consequence of both direct and indirect discrimination is essentially the same: an individual who belongs to a protected group is disadvantaged.

Despite the fact that the application of these definitions can be tricky depending on the exact facts of the case and the evidence available, they work rather well in the context of algorithms, as they explicitly state what criteria are directly taken into account and which ones are not. For example, in the context of bail decisions, algorithms which do not directly take race, sex, religion, or other prohibited factors into account for assessing the flight risk of an individual do not discriminate directly.27 However, as mentioned earlier, the COMPAS recidivism algorithm had worse effects for people of color in comparison to white people. Thus, the algorithm may indirectly discriminate by taking into account apparently neutral criteria such as the zip code, level of education, or a poor credit rating when predicting the flight risk. Since the abovementioned factors have often been influenced by previous discrimination,28 the common worry that AI may perpetuate discrimination, even if the algorithm does not explicitly take race or sex into account, is justified.

However, the COMPAS case, which has strongly shaped the public discourse on AI in the judiciary, has limited value for informing long-term policy, as the exact algorithm is undisclosed and kept as a trade secret.29 While the case shows that it is certainly possible to build algorithms that indirectly discriminate against protected groups (and perhaps discrimination is even likely, absent safeguards in development and application), our primary concern is whether it is possible to build an algorithm or AI that is beneficial from different philosophical points of view in comparison to human decision-making, even with respect to issues of discrimination. In this vein, recent research by Kleinberg and colleagues shows that such improvements are not merely a utopian vision, but feasible to build today.30 Again focusing on bail decisions, the machine learning algorithm developed by Kleinberg and colleagues, which used gradient-boosted decision trees, was trained with a large dataset of 758,027 defendants who were arrested in New York City between 2008 and 2013. The dataset included the defendant’s prior rap sheet, the current offense, and other factors available to judges for making the decision. When tested on over 100,000 different cases, the algorithm proved to be significantly better than human judges at predicting whether defendants would fail to appear or be re-arrested after release, which—depending on one’s policy preferences—can come with diverging yet strong benefits. More precisely, simulations showed failure to appear and re-arrest (“crime”) reductions of up to 24.7% and no less than 14.4% with no change in jailing rates, or jailing rate reductions up to 41.9% and no less than 18.5% without any increase in crime rates.31 Crucially from the perspective of preventing discrimination, the algorithm was able to achieve crime reductions while simultaneously reducing racial disparities in all crime categories.32 The adoption of the algorithm would thus put policymakers in the rather comfortable position of choosing between the options of releasing thousands of people pre-trial without adding to the crime rate or preventing thousands of crimes without jailing even one additional person—whilst reducing racial inequalities. Needless to say, these are not the only options, but they illustrate what kind of trade-offs policymakers will have to make with regards to the balancing of crime and detention rates, as well as the disproportionate imprisonment of minorities, particularly black and Hispanic males in the United States, if AI replaces HI. In fact, Sunstein argues that the algorithm developed by Kleinberg and colleagues “does much better than real-world judges (…) along every dimension that matters”.33

Importantly, these results are not unique to New York City, as Kleinberg and colleagues were also able to obtain qualitatively similar findings in a national dataset.34 However, one might be tempted to argue that it remains uncertain whether these results can be achieved in jurisdictions outside the US. Could it be that there is just something off with judicial decision-making in the US? Might there be a bug in the system which does not exist in other jurisdictions when it comes to bail decisions? After all, common and civil law systems vary,35 and legal education in the United States is fundamentally different from most places, even other common law systems such as Australia, the United Kingdom, or India. Despite this, I fail to see why such differences would lead to significantly different outcomes because, first of all, the law works remarkably similar with regards to bail decisions.36 Second, Sunstein points out that the analysis of the data suggests that cognitive biases can explain why AI so clearly outperforms HI.37 This is a relevant and important insight given that biases occur worldwide irrespective of the jurisdiction.38 To be precise, Sunstein argues that judges make two fundamental mistakes which significantly influence their overall performance.39 They treat high-risk defendants as if they were low-risk when their current charge is relatively minor, and they treat low-risk people as if they were high-risk when their current charge is especially serious.40 Thus, in each of these cases, judges seem to assign too much value to the current offense in comparison to other relevant factors including the defendant’s prior criminal record, age, and employment history.41 Sunstein refers to this phenomenon as the current offense bias.42

He then links the current offense bias to the well-known availability bias, i.e., the tendency to overestimate the likelihood of an event occurring in the future if examples can easily be brought to mind.43 Even though one may be a bit skeptical as to the degree of relatedness between the current offense bias and the availability bias,44 the overall point stands: judges in different jurisdictions are likely to make the same mistakes because there is no reason to assume that Asian, African, or European judges will not suffer from the current offense biases. This is especially so because the related availability bias seems to be a general trait of the mind trying to access the probability of an event occurring based on associative distance.45

We have now seen that crime and detention rates can be significantly decreased without perpetuating or even increasing racial discrimination. This itself could be viewed as a welcome development, as the absolute numbers of minorities in jail would decrease. But does that mean that, if we adopt Kleinberg and colleagues’ algorithm, the percentage of African Americans and other minorities in prison will remain the same? If we assume that human decision-making will eventually decrease discrimination, and thereby reduce not only the absolute but also the relative number of African Americans in prison, the question arises whether algorithms would lag behind HI after some time has passed. It seems plausible to assume that both absolute and relative detention rates for minorities will decrease in the future, if a human-centered judiciary follows humanity’s track record on moral progress.46 However, it will take time to develop the necessary and crucial cognitive features to do so more fully, such as through enhancement of the abilities to override implicit biases.47 At the same time, it is possible to instruct an algorithm in a way that it produces morally or socially desirable outcomes, whatever these may be. For instance, in one scenario, Kleinberg and colleagues instructed the algorithm to maintain the same detention rate while equalizing the release rate for all races. Given this, the algorithm was still able to reduce the failure to appear and re-arrest (“crime”) rate by 23%.48 In another simulation, the algorithm was instructed to produce the same crime rate that judges currently achieve. Under these conditions, a staggering 40.8% fewer African Americans and 44.6% fewer Hispanics would have been jailed.49

We might wonder whether these results really indicate that human decision-making ought to be supplemented or even (partially) replaced by AI from the perspective of discrimination. If it is all about cognitive biases—whether they are related to implicit racial biases or about current offense bias—maybe we can try to improve human decision-making before we get rid of it entirely. Indeed, it is plausible to mitigate some biases to some degree. For instance, the strength of the availability bias can even be reduced by simply being aware of it.50 Additionally, I have previously suggested different institutional and procedural changes as well as mandatory training in behavioral economics and cognitive biases for members of the judiciary in order to mitigate biases.51 However, it does not seem feasible to eliminate them altogether this way.52 To put it simply, in order to completely get rid of cognitive biases at this day and age, one must get rid of human decision-making.53

To summarize, Kleinberg and colleagues’ findings enormously strengthen the case for AI in the judiciary from the perspective of decreasing discrimination. This is especially so because it is possible to reduce crime and detention rates at the same time. Although there is much more to say about human decision-making, cognitive biases, and the role of empathy therein,54 contrary to the present popular belief, AI can be a “force for (…) equity.”55 While it is important to caution that Kleinberg and colleagues’ algorithm was merely concerned with bail decisions, it seems reasonable to assume that it would generally be easier to prove whether direct discrimination occurred under algorithmic than human judicial decision-making. In fact, if the decision procedure is transparent, it would be so obvious that it is unlikely to occur in the first place. The preferred trade-offs among crime, detention, and discrimination would still have to be made, but regardless of how one decides, the result would be an improvement from HI. Having said this, it is important to note—as experienced in the COMPAS case—that the exact trade-offs made may not always be visible. So, should one reject the implementation of AI, even if it can reduce discrimination, for lacking transparency?

II. From Transparency of Procedures to Transparency of Options

The relationship between democracy and transparency is highly contested. Contrary to widespread intuition, it may even be the case that authoritarianism leads to more transparency than democracy. This may be so because greater vulnerability to public disapproval could make democratically elected officials more inclined to promote opacity or withhold information compared to their autocratic counterparts, who have to worry less about public perception.56 The fact that democratic governments have incentives to obfuscate their policies can be illustrated by the recent remarks of Germany’s Minister for the Interior Horst Seehofer, who recently stated that “you have to make laws complicated”.57 Despite the fact that the statement itself may arguably not have been well thought-through from a political standpoint, and Seehofer consequently shortly thereafter tried to argue that he was being “slightly ironic”, the statement does get at the heart of the tension between democracy and transparency. If passing a specific law is better for citizens than any of the alternatives, from the perspective of elected officials, but the law in question is very unpopular, should it still be adopted? And if so, is it permissible to communicate it in a way that makes re-election more probable, despite the fact that citizens disagree with the measure being taken, even if they ultimately benefit from the new law?

The argument that democracy may in fact not always be the ideal engine for transparency does not entail that transparency is not a crucial factor thereof from a normative perspective. Of course, transparency broadly understood helps to fight corruption, promote trust in public institutions, and contribute to the public discourse.58 Yet, in order to answer more specific questions regarding the state of transparency when it comes to artificial judicial decision-making from a democratic perspective, one first needs to consider the function of transparency within a democracy. Interestingly, despite the visibility of transparency as a concept in the public discourse, surprisingly little attention has been paid to its underlying purpose.59 Gupta even calls it an “overused but under-analyzed concept”.60 The basic argument for transparency being an essential part of any democratic system runs along the following lines:61

  1. Democracy requires citizens to take an active part in politics.
  2. For citizens to be able to participate in politics, they need to be able to make political judgments based on relevant information. 
  3. In order to make political judgments based on relevant information, citizens need to have access to that relevant information.
  4. In order for citizens to have access to relevant information, such information must be transparent.

The central question in this framework is what exactly counts as relevant information. Do citizens need to know who made the decision in question and how, or do they also need to be aware of alternative routes which could have been taken? What information should be available to citizens when it comes to decisions by the judiciary, rather than the executive or legislature? 

If, as mentioned above, democracy requires citizens to take an active part in politics, then relevant information could be any information helpful for making political choices such as who or what to vote for, what to protest, and when and how to engage in public discourse more generally. As Bellver and Kaufmann argue, the information provided needs to account for the performance of public institutions because transparency is a tool to facilitate the evaluation of public institutions.62 More importantly, not only the decisions made by directly or indirectly elected individuals, but state actions more generally, including decisions and judgments made by the judiciary,63 need to be transparent in order for citizens to be sufficiently informed.

In order to evaluate public institutions carefully, citizens ideally need to know (or at least be able to know) not only the outcome of the decision itself, the people involved, and their respective roles in the process but—importantly—also alternative options which could have been taken. If, for instance, a decision is not very popular, but the alternatives are significantly worse, then ideally this would be communicated just as it should be transparent if a decision was made which intuitively sounds good, but which arguably had much better alternatives from the perspective of citizens. Accordingly, in order to guarantee that citizens receive relevant information to evaluate public institutions, they should be able to learn what kind of decision was made, the procedures leading to it, and what other options were available.64 To put it simply, transparency about options may be just as important as transparency about procedures and actors from the liberal democratic point of view.

Since it would be an improvement from a liberal democratic point of view if options were more frequently transparent, some AI—as illustrated by the clear trade-offs Kleinberg and colleagues’ recidivism algorithm produces—has strong advantages over HI in this regard. While nowadays citizens and policymakers alike have to trust their (often unreliable) intuitions to assess the situation, the clarity of the trade-offs, for example between public safety, discrimination, and detention rates, as in the case of bail decisions, will become clearer with the implementation of AI. For instance, this means that citizens will not only be able to see the racial composition of the population denied bail, but will also know the consequences of lowering or increasing this rate. In sum, citizens will have better access to highly relevant information in order to participate in the political process.65 Indeed, for better or worse, it will be much more difficult for politicians like Seehofer to use complicated laws to obscure the political decision and evaluation process. This is the case even if programs do not use explicitly programmed decision-making (e.g., a simple decision tree or a more complex algorithm), but instead use machine learning, including deep learning, due to the fact that the relevant trade-offs can still be transparent by evaluating actual and simulated impact, regardless of the complexity of machine learning.

Yet most commentators in the debate on AI and transparency take a different stance and focus. Instead of concentrating on the new levels of transparency regarding the tradeoff being made, they rightly point out that the algorithm itself is often unknown.66 Case in point, the COMPAS risk assessment algorithm, which is now used for bail and sentencing decisions in more than 20 jurisdictions within the United States, is a protected trade secret and thus remains a black box.67 This is already problematic because without knowing the algorithm or machine learning training set, one cannot understand and challenge the decision, in case the design is flawed. From a liberal democratic point of view, this is even more troublesome. If citizens are not able to obtain vital information to evaluate the operations of the judiciary and related procedural norms, they cannot participate in the democratic process in a meaningful way, aside from pointing out that the algorithm should be made public. It may not be unethical per se to outsource typical tasks of the judiciary to private companies, yet it becomes highly questionable if this process leads to the lack of transparency witnessed in the COMPAS case.

Hence, while AI has strong potential benefits with regards to the clarity of tradeoffs, the current lack of transparency (of procedures) in many cases and the potential lack thereof in the future has led scholars to become increasingly skeptical.68 However, in order for it to be sufficient reason to reject the implementation of AI, one would have to show that, at first, HI is in fact better suited to tackle these issues and, secondly, the lack of this kind of transparency is more harmful from the liberal democratic perspective than the lack of transparency regarding the above-mentioned trade-offs. I will consider these in turn.

It may be argued that human judicial decision-making is much more transparent than artificial judicial decision-making because citizens can identify human judges. They can read their arguments and decide for themselves whether the judgment was reasonable or not. In short: Human judges explain why they have come to the decision they came to. Although one can find plenty of ongoing research with regards to the explainability of AI, present applications certainly lag behind human capabilities to explain.69 However, there are notable problems with this line of reasoning.

First, giving human explanations does not equal full transparency of procedures. Explaining judgments is important and can provide a basis for appeal, but an explanation does not necessarily reveal the true underlying cognitive processes. The processes of human (judicial) decision-making are just as unknown as the proceedings of some applications of AI. Following a computational theory of cognition,70 one might even argue that we can at least know the algorithm or data set with regards to artificial decision-making, even if we may not understand its operations, whereas the algorithm running human decision-making remains unknown. It is crucial to note that while a computational perspective of cognition certainly strengthens this argument, it does not rest upon it. Whatever philosophy of mind one might prefer, human decision-making remains a black box as well.

Second, human explanations are not only prone to error and cognitive biases, but also at risk of rationalizing underlying motivations and preferences.71 Intuitively, one might expect that this general tendency might be mitigated by expert knowledge and education, yet as studies have shown over and over again, experts rarely perform better than lay-people.72 Accordingly, the worry is that even though one might know and understand the official reasons given, they might be misleading and the risk remains that the ultimately decisive reasons are still unknown. This is not only important for democratic evaluations but might also decrease the chances of a potential successful appeal.

Third, depending on the complexity of the individual judgment, one might question whether laypeople can fully understand judgments anyway. Surely, one might counter-argue that lawyers will be able to understand the judgment and advise their client accordingly. However, from a liberal democratic perspective, there is not much benefit in lawyers navigating cases if the vast majority of citizens are unable to comprehend and—based on this—evaluate judicial norms and behavior correctly. Arguably, overconfidence may even lead laypeople to think they understand the judgment and give them a false sense of understanding.

These theoretical and practical concerns question the assumption that HI really does provide more liberal-democratically relevant transparency than AI. However, let us assume that this analysis is either mistaken or that there are other overriding arguments which speak in favor of HI-transparency in this regard.73 Consequently, one would still have to investigate whether the lack of human explanations for individual judgments is more harmful than the lack of transparency regarding the above-mentioned trade-offs. Should humans have the possibility to understand judgments in individual cases, or should they rather be able to clearly see the necessary trade-offs behind rules, which not only shape society but also the very individual judgments they care to understand? Should they understand the purpose of the rules, its underlying trade-offs and the explanations given by the ones they voted for (legislature) or the application of them made by often not democratically accountable judges in selected scenarios? Even though the advantages of having access to human judicial explanations ought not to be downplayed, I am afraid that the transparency of the trade-offs in question might simply be even more important.74

III. From Separation of Powers to Judicial Dependence

While improvements with regards to discrimination (I.) and transparency (II.) seem possible, and related issues have been receiving the appropriate academic attention in the past few years, medium- and long-term threats arising from the application of advanced AI within the judiciary for liberal democracy have been almost completely neglected.75 Hence, this part of the analysis aims to raise awareness of these risks. More precisely, I will make the point that advanced artificial judicial intelligence may threaten judicial independence and the separation of powers more generally. Before I proceed with the analysis by outlining why the development of artificial judicial intelligence ought to be followed closely from the point of view of liberal democracy, it is necessary to provide a brief explanation for this seemingly science-fiction scenario.

1. The Possibility of an Advanced Artificial Judicial Intelligence

Advanced artificial judicial intelligence (AAJI) can be defined as an artificially intelligent system that matches or surpasses human decision-making in all domains relevant to judicial decision-making. Crucially, this does not require the development of an artificial general intelligence (AGI) while still avoiding the need for hybrid human-AI judicial systems. Such a system goes significantly beyond the current state of the art; with it, humans would have the ability to outsource decision-making within the judicial sector entirely if they wished to do so. Although such a development is often considered by legal researchers to be extremely unlikely in the medium- or even long-term future,76 machine learning researchers think that even the development of an AGI, i.e., an artificially intelligent system that matches or surpasses human decision-making in all relevant domains,77 is much closer than the common sense among jurists seems to indicate.78 For instance, experts believe there is a 50% chance of AI outperforming humans in all tasks in 45 years.79 With regards to specific activities, they predict that AI will outperform humans fairly soon. This includes translating languages (by 2024), writing high-school essays (by 2026), driving a truck (by 2027), working in retail (by 2031), writing a bestselling book (by 2049), and working as a surgeon (by 2053).80

Surely, one can debate whether it is even possible for an AI to engage in legal reasoning. Crootof, for instance, points out that “the judgment we value in a common law process is a distinctively human skill.”81 In this regard, she further argues that “[g]iven their sensitivity to changing social norms, human judges are uniquely able to oversee legal evolution and ensure that our judicial system ‘keep[s] pace with the times.’” However, if one is primarily concerned about detecting (changing) social norms, AI may arguably be better equipped to do so than HI. While AI can be trained with vast data sets from diverse sources and be more easily and broadly adapted with changing norms, HI relies on personal interpretations of comparatively little and highly selected information from media and interactions with again a very selected group of people. Hence, AI not only has the power to process a greater amount of relevant data, but it is also less likely to fall for the ubiquitous confirmation bias and other misleading selection mechanisms that lead one to paint an inaccurate picture of existing social norms.82

This having been said, following a Dworkinian interpretation of law, one could certainly question whether even an advanced AI would be up for the task of “moral reasoning” as part of the judicial decision-making process.83 At the same time, chances that the judicial system can be outsourced would arguably increase if one would follow Oliver Wendell Holmes’ prediction theory of law.84 Holmes famously stated that “[t]he prophecies of what the courts will do in fact, and nothing more pretentious, are what I mean by the law.”85 The purpose of this section is not to convince the reader of the merits of a specific legal theory, but rather to emphasize the high degree of jurisprudential uncertainty.86 If we can agree that AI would be able to take over the judiciary from some legal theoretical perspectives—even if one personally does not share those views or considers them unlikely to be correct—we have strong reason to start thinking about the accompanying consequences and potentially relevant safeguards. In other words, the mere possibility of an AAJI should lead us to take seriously its implications. 

2. Advanced AI & Judicial Independence

Even though one might argue that a liberal democratic system may in theory be achieved and sustained without the separation of powers, this separation of powers has come to be a cornerstone of any liberal democratic system around the world. The separation of the judicial, legislative, and executive branches of government serves a vital purpose in minimizing the concentration of power and maintaining checks and balances across the government. While checks and balances differ in their configuration across jurisdictions, any significant influence on a given power structure (e.g., branch of government) will likely destabilize the system. Relatedly, this destabilization can take very different forms and outcomes in different systems. For instance, parliamentary democracies in Western Europe may respond very differently to the implementation of an AAJI in comparison to the presidential systems in the Americas. Additionally, those jurisdictions and cultures which consider their judicial branch to have some legislative-like lawmaking powers, such as the European Union87 and the United States,88 may evaluate the threats and opportunities imposed by AAJI very differently than those who consider their judiciary as primarily politically neutral (even if this may in fact not be the case, as legal realism tells us), such as the United Kingdom89 or Germany.90 This fact can be illustrated by the intensity of discussions taking place during the selection procedure of Supreme Court Justices. While appointments in the United States can occupy media attention for weeks,91 there is comparatively little debate in Germany or the UK, which is unsurprising if the judiciary is considered to apply laws neutrally. Finally, the crucial question is not whether but how the interaction and power dynamics between branches of government may be threatened by an AAJI. Will the judiciary become more or less powerful? How would this affect the interaction with and powers of the legislature and executive? Such questions are important from any liberal democratic perspective because, first of all, authoritarianism tends to favor a weak judiciary. Second, an independent judiciary is an important ally in the fight for minority rights and preventing the “tyranny of the majority”.

Michaels argues that the current human-based judiciary is necessary to maintain the separation of powers, for only a human-centered judiciary raises ample attention to the law from the involvement of judges, lawyers, law professors, and so on.92 For Michaels, it seems likely that the legal community would diminish if artificial judicial intelligence were to replace human judges,93 and without a legal community, humans would pay little attention to the law.94 He further argues that such lack of sufficient attention from the legal community might have devastating effects. First, it would be hard to imagine any public response to abuses of authority or concentration of power across the three branches.95 Second, there would generally be “little incentive to construct high-quality legal arguments if there was no possibility that doing so could shape the result.96

However, even assuming that the legal community would in fact diminish or be significantly reduced, it is not clear that this itself would necessarily lead to great power imbalances due to decreased human attention. On the contrary, there may be some significant accompanying advantages resulting from the shift in focus. The attention nowadays spent on specific cases may shift to law-making processes and the analysis of the overall consequences of the laws in question. While society is often occupied with extraordinary individual cases of little systemic value for the long-term, with AAJI, humans would be able to focus on the evaluation of abstract principles and rules rather than on our intuitive judgments about specific, often highly politicized, cases. Given that such perceptual intuitions are especially prone to cognitive biases,97 offer little normative guidance, and have limited value for the whole of society, this may even nudge the public discourse onto a more beneficial path. Instead of focusing on extraordinary individual cases, attention might shift to the much more important trade-offs between detention, crime rates, and discrimination discussed in the first part of this article—potentially leading to lower incarceration, higher security, and less discrimination. Furthermore, the shift in focus from the judiciary onto the legislature (and executive) may carry the advantage of greater public accountability of the legislature, which would again profit from the transparency of the aforementioned trade-offs.

Although one can debate whether the adoption of AAJI would in fact lead to decreased attention on the judiciary, and whether such development would necessarily go hand in hand with a loss of power, the judiciary may nonetheless lose a significant share of its power and, eventually, its independence. This is ultimately so because taking the human out of the equation faces risks of democratic legitimacy. This itself may lead to a reduction of judicial power. Additionally, with reduced democratic legitimacy, it is difficult to imagine that judicial discretion would remain a decisive factor in many cases. On the contrary, I assume that legislatures would use a democratically weak judiciary to justify more and more specific rulings, leaving little discretion for the AAJI.98 With power shifting from the judiciary to the legislature, judicial independence and the separation of powers more generally would be under threat. The graphic below outlines the threat to judicial independence resulting from the implementation of AAJI.

If democratic legitimacy were compromised in this way, the people might even welcome more powerful legislatures (and executives). Such a development could be accelerated by the improved accountability of the legislature resulting from greater transparency of the aforementioned trade-offs. At last, the legislature may increasingly be perceived as responsible for the courts’ rulings. Case in point, if the AAJI does not decide as the public wishes, people may blame the legislature for—in their eyes—a poor design of the laws in question, or poor decision to use a biased AAJI, rather than blaming the AAJI for poor implementation or interpretation of such laws. From this perspective, even governing parties who strongly support judicial independence might give in to public pressure to limit AAJI’s discretion over time.

This is not a definite outcome, and it is highly uncertain where this shift in power would ultimately lead. However, the adoption of an AAJI may pose a great risk for judicial independence which arguably ought not to be taken before necessary safety mechanisms will have been generated. While there are tremendous efforts invested into the development of more advanced AI, including AGI and AAJI, little attention has been paid to creating the necessary political infrastructure that would allow society to harness the potentially enormous benefits of AI without risking the collapse of judicial independence. Although I disagree with Michaels regarding the reason why one might expect a shift in power (attention vs. legitimacy), this analysis strengthens the case for AAJI affecting the separation of powers. More precisely, I argue that it poses a great threat to judicial independence, which ought not be taken lightly.


As part of the legal community, we might be inclined to improve the judicial system as much as possible and defend it from potential threats, even if those threats accompany extraordinary improvements in other domains. As academics, we want to completely understand AI reasoning before we recommend its implementation, even if we do not fully understand human reasoning either. And as humans, we tend to prefer the status quo, even when change would be net-positive.99 All these motivations might explain why so many of us have a strong aversion to the implementation of AI into the judicial sector, but, unfortunately, motivations are not ideal truth-tracking processes.100 At least when it comes to issues of discrimination and transparency, this analysis makes clear that current applications could come with great advantages. Whereas this proposition can be defended from a wide range of normative perspectives on discrimination, the improvements with regards to transparency depend on how or whether one favors transparency of options over transparency of procedures and agents. Having said this, one should bear in mind that the existence of the possibility of such improvements does not mean that there is no risk of large-scale discrimination or a lack of transparency. As seen, the fact that such AI is possible does not mean that AI used in practice will be more transparent or less discriminatory. 

While the analysis hitherto focused on existing technological capabilities, the final chapter introduced the concept of AAJI, defined as an artificially intelligent system that matches or surpasses human decision-making in all domains relevant to judicial decision-making. Concentrating on the potential power shift from the judiciary to the legislature, it has been argued that the adoption of an AAJI threatens judicial independence—and with that one of the central foundations of liberal democracy.

Needless to say, this attempt to outline the challenges of artificial judicial decision-making for liberal democracy is far from complete, and further research is needed in many regards. First, it is necessary to identify what other challenges the adoption of an AAJI may bring about, such as the risk that algorithms may grant laws a new kind of permanency and thereby create an additional barrier to legal evolution.101 Second, it will be crucial to investigate how one can uphold liberal democratic values more generally and, in particular, the separation of powers with an AI Judiciary. The short-term benefits AI could offer over HI with regards to access to justice, transparency, and fairness are so enormous that one may overlook the long-term threats imposed by AAJI. Working out the political and legal infrastructure to reap the fruits of artificial judicial intelligence in a safe and stable manner should become a priority of future research.


Christoph Winter, The Challenges of Artificial Judicial Decision-Making for Liberal Democracy, in Judicial Decision-Making: Integrating Empirical and Theoretical Perspectives, 14 Economic Analysis of Law in European Scholarship 179 (Piotr Bystranowski et al. eds., 2022),

The challenges of artificial judicial decision-making for liberal democracy
Christoph Winter
The challenges of artificial judicial decision-making for liberal democracy
Christoph Winter
URL links