Authors
Steven Baker
Abstract
As artificial intelligence (AI) technologies become increasingly embedded in criminal justice procedures, the legitimacy of algorithmic tools in sentencing decisions has come under heightened scrutiny. Proponents champion these systems as potential mechanisms for objectivity and consistency, offering data-driven evaluations of recidivism risk. Critics counter that the opacity and complexity of these models can subvert fundamental due process protections, particularly if defendants are not afforded a meaningful way to assess and challenge algorithmic outputs. This paper narrows its focus to procedural fairness imperatives, arguing that transparency and the right to explanation are pivotal components of any ethically and legally defensible AI-driven sentencing framework.
Drawing upon interdisciplinary sources—encompassing constitutional law, philosophy, social theory, and computer science—this work examines how historical risk assessment methods have transitioned from rudimentary actuarial instruments to sophisticated machine learning systems. Central to the discussion is the tension between proprietary claims of algorithm developers and defendants’ constitutional entitlements to scrutinize evidence, as well as emerging proposals for explainable AI (XAI). Through an expanded look at case law, legislative efforts, and empirical examples such as COMPAS and the Public Safety Assessment, the paper elucidates the ways in which opaque models can embed systemic biases while evading meaningful accountability. In response, it offers policy recommendations for designing more transparent, interpretable AI architectures, instituting robust auditing practices, and fortifying defendants’ procedural rights in an era where automated judgments risk overshadowing human discretion. Ultimately, this research proposes that while AI may contribute valuable insights to sentencing decisions, it must not erode the foundational principles of fairness, autonomy, and justice that lie at the heart of any democratic legal system.
INTRODUCTION
The expansion of artificial intelligence into various aspects of everyday life has raised urgent debates regarding the ethical, social, and legal consequences of delegating traditionally human judgments to computational systems. Within the criminal justice sphere, algorithmic risk assessments have emerged as a prominent fixture in bail determinations, parole decisions, and, increasingly, sentencing recommendations. Advocates of these tools emphasize their potential for enhancing consistency, reducing individualized bias, and potentially aligning legal decisions more closely with empirical data (Berk, Sherman, Barnes, Kurtz, & Ahlman, 2015). Yet, numerous critics highlight the peril of entrusting pivotal judicial functions to predictive systems that often remain inscrutable or proprietary, effectively challenging long-standing rights to due process (Angwin, Larson, Mattu, & Kirchner, 2016).
It is within this paradoxical landscape—where advanced technology appears both to offer innovative paths toward fairness and simultaneously threaten to obscure accountability—that this paper zeroes in on the issue of procedural fairness. In particular, the discussion spotlights two core pillars of procedural justice in AI-driven sentencing: transparency and the defendant’s right to explanation. While the broader discourse on algorithmic fairness encompasses topics such as discriminatory data, the moral significance of recidivism forecasting, and the wider social context of mass incarceration, here we prioritize the crucial procedural question: How can defendants and the courts be guaranteed a meaningful opportunity to inspect, challenge, and understand algorithmic evidence in sentencing decisions?
In formulating an answer, this manuscript employs a thoroughly interdisciplinary lens. On one hand, constitutional and legislative frameworks stipulate that evidence leveraged in criminal proceedings must be subject to scrutiny, cross-examination, and confrontation (Citron, 2008). On the other, the evolving field of computer science offers a range of tools—both post-hoc interpretability methods and intrinsically interpretable architectures—that can illuminate black-box models. Bridging these areas requires grappling with competing interests, such as proprietary protections asserted by private vendors, resource constraints in public defender offices, and, importantly, the broader policy climate that either spurs or stifles regulatory intervention (Garvie, 2019).
This introduction proceeds with several objectives. First, it outlines the larger historical context of risk assessment in criminal justice, tracing the transition from simple actuarial instruments to advanced machine learning algorithms and explaining how those developments set the stage for contemporary debates on transparency. Second, it establishes the philosophical and legal underpinnings that link transparent processes with core tenets of due process, emphasizing that a fair hearing includes not only the right to respond to evidence but also the capacity to comprehend the basis of one’s potential punishment (Hildebrandt, 2016). Third, it previews the subsequent sections of the paper, which delve into technical solutions in the sphere of explainable AI (XAI), legislative and judicial responses to algorithmic opacity, and the moral imperatives that demand a robust explanation right. By focusing intensively on these specific procedural dimensions, we aim to present a framework that can inform both policy debates and practical implementation strategies for AI-driven sentencing—ensuring that, ultimately, technology bolsters, rather than erodes, the foundational values of justice.
HISTORICAL DEVELOPMENT OF ALGORITHMIC SENTENCING TOOLS
The application of quantitative methods to criminal justice decisions is hardly new. Before the emergence of modern machine learning, courts and correctional systems experimented with actuarial formulas to gauge recidivism risk, often through relatively straightforward checklists that assigned numerical weights to factors like employment history, prior convictions, or educational status. Although these tools were criticized for their reductive nature and susceptibility to bias, they formed a rudimentary precursor to contemporary AI systems (Andrews, Bonta, & Wormith, 2006).
1. Early Actuarial Systems
Early actuarial systems were typically linear, with each factor contributing a fixed point value toward a total risk score. These measures aimed at circumventing purely subjective judicial instincts, thus in principle advancing consistency. Researchers at the time celebrated these developments for bringing a veneer of empiricism to sentencing and parole decisions (Simon, 2007). Nevertheless, structural critiques surfaced almost immediately: measures like “unemployment” or “single-parent household” often correlated powerfully with socio-economic variables and community-level disparities, effectively punishing poverty rather than criminality (Slobogin, 2017). By embedding these correlations into the scoring process, early instruments risked codifying existing social inequalities.
2. Emergence of Machine Learning
The revolutionary leap from basic actuarial tables to complex machine learning was facilitated by the massive increase in computational power and data availability toward the end of the 20th century. With the digitization of criminal records and policing databases, researchers began experimenting with classification algorithms—such as decision trees, random forests, or neural networks—to discover deeper, non-linear patterns in defendant profiles (Berk, 2017). Proponents argued that these methods could achieve higher predictive accuracy by capturing sophisticated interactions among variables. Gradually, commercial vendors recognized a market for risk assessment tools, marketing their products to government agencies as advanced solutions for predictive policing, bail recommendations, and sentencing (Berk et al., 2015).
Simultaneously, the notion that technology could minimize human error and bias became a potent selling point. Many officials believed that automated risk scores, being “data-driven,” would transcend the personal prejudices that might color a judge’s view of a defendant’s demeanor or background (Silver, 2019). Yet, the intangible complexity of machine learning also introduced new challenges. Models that processed thousands of data points across multiple layers lacked the transparency and straightforwardness of earlier actuarial systems (Burrell, 2016). Moreover, any errors or biases that crept into the training data stood a risk of magnification through these advanced computational processes.
3. Proprietary Tools and Black-Box Concerns
By the early 2000s, a range of proprietary tools—most famously COMPAS (developed by Northpointe/Equivant)—began gaining traction in courts across the United States (Angwin et al., 2016). Packaged as comprehensive risk assessment solutions, these products typically included user-friendly interfaces for court officials to generate risk scores quickly. The behind-the-scenes algorithm, however, remained a closely guarded corporate secret. The ramifications for due process were striking: a defendant might receive an elevated risk classification, influencing sentencing outcomes, without any opportunity to scrutinize or challenge the underlying logic. This opacity attracted criticism from civil rights organizations, legal scholars, and journalists, all of whom questioned whether secret algorithms were compatible with constitutional guarantees (State v. Loomis, 2016).
Over time, controversies such as the ProPublica investigation into racial disparities in COMPAS fed broader anxieties about algorithmic injustices and the risk that technology—rather than mitigating bias—could systematically worsen it (Larson, Mattu, Kirchner, & Angwin, 2016). Yet, courts and legislative bodies have often struggled to keep pace with these innovations, leaving a vacuum in which private vendors can operate with minimal oversight (Garvie, 2019). It is this vacuum that sets the stage for legal, ethical, and policy debates regarding transparency, the right to explanation, and whether machine learning can truly serve the cause of justice without sacrificing fundamental due process rights.
4. Evolving Critiques and Push for Reform
As more practitioners and scholars became aware of machine learning’s “black box” nature, calls for accountability grew louder. Researchers from computer science introduced the concepts of interpretability and explainable AI (XAI), sparking hopes that sophisticated models could be rendered comprehensible (Rudin, 2019). Meanwhile, legal scholars and ethicists began mapping out how an “explanation right” might be integrated into existing due process frameworks (Citron, 2008). Some jurisdictions initiated efforts to either ban purely proprietary tools or mandate partial disclosures, reflecting a mounting recognition that blind reliance on inscrutable algorithms poses severe risks for legitimate judicial proceedings (Stevenson, 2018).
In the broader context, this historical trajectory signifies a paradigm shift: from naive faith in data-driven efficiency to a more nuanced understanding of how machine learning might inadvertently codify discrimination and undermine constitutional protections. The challenge lies not merely in adopting AI but in doing so in a way that maintains transparency and defends the defendant’s procedural autonomy. As the subsequent sections of this paper will elaborate, these issues converge around the concept of procedural fairness, which requires more than just nominal inclusion of AI tools. It demands robust frameworks for understanding, disputing, and refining how algorithms shape sentencing decisions.
PROCEDURAL FAIRNESS AND THE IMPERATIVE OF TRANSPARENCY
Within democratic justice systems, procedural fairness stands as a cornerstone, ensuring that individuals threatened with the loss of liberty or property have meaningful opportunities to participate in, and challenge, the decision-making process. When AI-driven tools contribute to sentencing, this fundamental principle intersects with technical and proprietary complexities, raising pressing constitutional questions.
1. Due Process and the Right to Confront Evidence
In U.S. constitutional law, the Sixth Amendment’s Confrontation Clause and the Fourteenth Amendment’s Due Process Clause collectively guarantee that a defendant can cross-examine witnesses and rebut evidence used against them (Citron, 2008). In a more generalized context, many other legal systems include similar provisions ensuring the right to a fair hearing, public accountability, and the opportunity to refute allegations. The problem emerges when courts regard an AI-generated risk score as “scientific” or “expert” without thoroughly examining or providing the defense with mechanisms to interrogate the data and logic behind that score (Slobogin, 2017). If the underlying model is considered proprietary, defendants may be denied the transparency needed to mount a meaningful critique—effectively relegating them to passive objects, judged by computational processes they cannot see.
2. Procedural Fairness vs. Substantive Fairness
A helpful lens is the division between procedural and substantive fairness. The latter deals with whether outcomes are equitable—whether sentences for similarly situated offenders align, for instance, or whether certain groups are disproportionately penalized. Procedural fairness, by contrast, centers on the methods used to arrive at those outcomes (Rawls, 1971). Even if an algorithm were to produce consistently fair results in a purely statistical sense, it could still violate procedural fairness if it is opaque and immune to scrutiny. Conversely, a thoroughly transparent tool might still yield biased outcomes, violating substantive fairness but preserving some measure of procedural equity (Berk et al., 2021). This dual perspective helps illuminate why transparency alone is insufficient to eradicate prejudice but remains indispensable for identifying and challenging the conditions under which prejudice might thrive.
3. Legal Cases and Judicial Hesitation
The tension between AI-based scoring systems and constitutional requirements has surfaced in multiple court decisions. Perhaps the most notable example is State v. Loomis (2016), where the Wisconsin Supreme Court recognized that the use of a proprietary tool (COMPAS) raised important due process considerations. Yet, the court ultimately allowed the continued use of COMPAS, with the caveat that judges should not rely solely on the risk score. Critics of this decision argue that disclaimers are insufficient to protect defendants’ due process rights if the core algorithmic methodology remains hidden (Garvie, 2019). Similar judicial reluctance to force full disclosure persists in other jurisdictions, reflecting broader uncertainties about balancing intellectual property interests against constitutional norms.
4. Public Trust and Perceptions of Legitimacy
Legal theorists emphasize that procedural fairness also affects public perceptions of legitimacy (Elek, 2019). If citizens believe that criminal sentencing is dictated by opaque “black boxes,” faith in the justice system can erode—especially in communities already skeptical of law enforcement motives. The intangible nature of algorithmic outputs can magnify a sense that decisions lack moral grounding. By contrast, transparent procedures, even when they result in severe outcomes, tend to garner higher compliance and acceptance because stakeholders see them as just and open to scrutiny. Thus, ignoring calls for transparent AI does more than infringe upon defendants’ rights—it risks undermining the social contract that undergirds the entire criminal justice system (Hildebrandt, 2016).
5. The Necessity for Explanation Rights
Embedded in this dialogue is the evolving concept of a “right to explanation.” Unlike evidentiary rules that focus solely on admissibility and cross-examination, explanation rights stress that individuals subjected to algorithmic decisions deserve to know how those decisions were reached (Goodman & Flaxman, 2017). This extends beyond providing raw data or cursory disclaimers; it implies offering comprehensible rationales, identifying salient factors, and revealing potential uncertainties or data quality issues. While not universally codified into law, the concept has deep legal and ethical resonance, linking back to the idea that “no one should be punished without knowing the reasons why”—a principle as old as the social contract itself (Rawls, 1971).
THE RIGHT TO EXPLANATION: LEGAL FOUNDATIONS AND PHILOSOPHICAL JUSTIFICATIONS
Calls for a formalized right to explanation in the realm of AI-driven sentencing echo deeper principles in moral and political philosophy. By examining how various schools of thought conceptualize individual agency, we can better grasp why explanation is not merely an “add-on” but a cornerstone of ethical governance.
1. Kantian Autonomy and Respect for Persons
Immanuel Kant’s moral philosophy posits that human beings, as rational agents, must be treated as ends in themselves rather than means to an end (Kant, 1785/2011). When a court imposes a sentence—potentially curtailing liberty—it wields extraordinary power. If that power is executed via an impenetrable algorithm, the individual subjected to it may feel dehumanized, reduced to a set of data points processed by an inscrutable mechanism. The right to explanation aims to mitigate this by restoring an element of dialogical respect: the state, acting through the court, is obliged to articulate the rationale behind its decisions in a manner accessible to the one who stands to lose significant freedoms (Hildebrandt, 2016).
2. Rawlsian Fairness and Public Reason
John Rawls highlights the notion of public reason, insisting that the fundamental rules shaping societal institutions be justifiable to all citizens under conditions of fairness (Rawls, 1971). If sentencing decisions rest on black-box algorithms, they fail the test of public reason, since the critical premises remain hidden. This is especially problematic in a democracy, where citizens are supposed to collectively determine the principles of justice. AI systems that cannot be interrogated or understood short-circuit that democratic process (Friedler, Scheidegger, & Venkatasubramanian, 2021). Hence, a robust right to explanation is consistent with Rawls’s emphasis on transparent institutional practices: it compels courts to reveal, or at least clarify, the basis for penal determinations, ensuring that punishment is neither arbitrary nor veiled in corporate secrecy.
3. Care Ethics and Relational Accountability
Less frequently cited in the criminal justice context, care ethics provides another layer of philosophical insight. Rooted in empathy and the interdependence of social relationships, care ethics demands that institutional procedures acknowledge human vulnerability and uphold the dignity of those most affected by decisions (Bandura, 1999). AI-based sentencing can inadvertently further alienate defendants, treating them as statistical abstractions. Explanation rights reinforce a relational approach, requiring that the system engage with the defendant in an explanatory dialogue, thereby humanizing what might otherwise be a cold, mechanical process (Hildebrandt, 2016). Such engagement is vital for fostering a sense of moral responsibility on the part of decision-makers and underscoring the defendant’s status as a moral agent, not a mere data cluster.
4. Broader Democratic Accountability
Beyond the individual rights perspective, the argument for explanation also taps into broader notions of democratic accountability (Noble, 2018). Criminal sentencing does not occur in a private sphere but as a public act by an institution entrusted with upholding societal norms. If the rationale behind these acts remains inaccessible, public oversight is severely compromised. Citizens cannot effectively evaluate—or contest—sentencing policies if they do not know how decisions are being made, or if any attempt at obtaining clarity is stifled by claims of proprietary secrecy. The right to explanation thus fosters communal engagement with the justice system, enabling policy reforms and improvements rooted in transparent critique (Barocas & Selbst, 2016).
5. Toward a Unified Philosophical-Legal Framework
When these various strands—Kantian deontology, Rawlsian fairness, care ethics, and democratic accountability—intersect with constitutional mandates for due process, a cohesive picture emerges. Explanation rights are more than a technological fix or ephemeral policy option; they are deeply embedded in the moral fabric of a legal system that respects autonomy, equality, and accountability. The next logical question, then, is how these normative imperatives translate into the practicalities of AI engineering and legal structures—a question addressed by the rise of explainable AI (XAI).
EXPLAINABLE AI (XAI) AND ITS RELEVANCE TO CRIMINAL SENTENCING
While the moral and legal imperatives for transparency grow clearer, the computational reality of providing explanations is more intricate. XAI emerges as a rapidly evolving discipline aiming to reconcile the complexity of machine learning models with the human need for interpretable reasoning.
1. Intrinsic vs. Post-Hoc Explainability
Within XAI, scholars distinguish between intrinsic interpretability and post-hoc explanatory techniques. Intrinsic interpretability involves building models—like decision trees, rule-based systems, or linear models—that are inherently more transparent (Rudin, 2019). Their structure allows direct inspection of how inputs map to outputs. Post-hoc methods, such as Local Interpretable Model-Agnostic Explanations (LIME) or SHapley Additive exPlanations (SHAP), operate on trained “black-box” models, producing simplified local approximations that explain a prediction in a specific neighborhood of the data (Arrieta et al., 2020).
In the criminal sentencing context, this distinction matters greatly. If a jurisdiction demands high interpretability, one might choose an intrinsically interpretable model, albeit at the possible cost of some predictive precision. By contrast, if the system relies on a highly complex black-box architecture, post-hoc methods could provide partial insight into the rationale for each individual outcome (Rudin, 2019). However, post-hoc explanations can be superficial or incomplete, potentially devolving into mere “explanation theater” if not carefully validated (Selbst & Barocas, 2018).
2. The Challenge of Global vs. Local Explanations
Another dimension of XAI in sentencing is the tension between global and local explanations. A global explanation aims to clarify the overall structure of the model—how it weighs different features across the entire dataset. A local explanation focuses on a single prediction, explaining why a specific defendant was labeled as high or low risk. Courts often need both: a local explanation may suffice to challenge a single sentencing recommendation, but if patterns of bias or error span multiple defendants, stakeholders require global insights to propose structural remedies (Corbett-Davies & Goel, 2018). Achieving both local and global comprehensibility in a complex machine learning model remains a non-trivial goal, especially when the model involves thousands of parameters or hidden layers.
3. Auditing and Validation
Explainability alone does not guarantee that a model is fair or accurate. Hence, XAI intersects with auditing protocols, wherein independent researchers or court-appointed experts assess the model’s performance across different demographic groups (Berk et al., 2021). If an XAI tool reveals that the model systematically overestimates risk for a particular community, then legislative or judicial intervention might be warranted. Whether these audits are mandatory, periodic, or triggered by a defense motion varies widely among jurisdictions. The interplay between technical capacity and legal mandates is pivotal: an advanced XAI method, if used purely internally by developers without transparency or accountability, does little to protect defendants’ rights (Raji et al., 2020).
4. Practical Constraints in Implementation
Despite the theoretical appeal of XAI, practical implementation in sentencing contexts faces hurdles. Many jurisdictions do not have the budget or expertise to thoroughly vet proprietary AI systems. Judges, lawyers, and probation officers may lack the training to interpret complex model explanations. The pressure to process cases rapidly might encourage reliance on risk scores without delving into methodological nuances (Green & Chen, 2019). Additionally, the partial or incomplete nature of certain XAI methods could lead to a misapprehension of how the system truly operates. This reality underscores the need for not only technical solutions but also institutional frameworks—educational programs, standardized guidelines, and oversight bodies—that meaningfully integrate XAI into the sentencing process (Slobogin, 2017).
5. Bridging XAI with Legal Standards
Ultimately, bridging XAI with established legal standards for evidence and due process requires new interpretive frameworks. Courts typically rely on tests like Daubert (in the U.S.) to evaluate the admissibility of scientific evidence, asking whether it is grounded in widely accepted methodologies and peer-reviewed research (Daubert v. Merrell Dow Pharmaceuticals, Inc., 1993). For AI, novel interpretative steps may be needed: does the model’s architecture align with best practices in data handling? How do we weigh an XAI explanation in cross-examination? Can partial access to proprietary code satisfy the confrontation right? Answering these questions demands a partnership between the technological and legal communities, forging standards that reflect both computational realities and constitutional imperatives (Re & Solow-Niederman, 2019). Such standards would embed XAI into the broader tapestry of procedural fairness, making the right to explanation more than a rhetorical ideal.
PROPRIETARY SECRECY VS. DUE PROCESS: NAVIGATING THE TRADE SECRET DILEMMA
A central friction in the quest for transparent AI-driven sentencing is the assertion of proprietary rights by commercial vendors. This section delves deeper into how trade secret laws converge with constitutional mandates, examining possible solutions and their limitations.
1. The Commercialization of Risk Assessment
As local courts and government agencies sought to modernize, private companies recognized a lucrative niche in criminal justice analytics. Vendors offer end-to-end platforms that promise risk scoring, data management, and even ongoing software support. The impetus to adopt these solutions often comes from a desire to streamline overburdened systems or from political pressures to appear “tough on crime” while employing “innovative” methods (Simon, 2007). As the dependence on third-party providers grows, so too does the tension around transparency. These companies invoke trade secret law to shield algorithms and data sources, contending that revealing them would compromise their competitive advantage (Garvie, 2019).
2. Constitutional Clash
From a legal standpoint, constitutional rights to confront and cross-examine evidence do not necessarily yield when pitted against trade secret assertions (Citron, 2008). While there have been precedents suggesting that in camera review by a judge or a neutral expert might suffice, critics argue that such solutions often fail to provide the defense with a genuine opportunity to challenge the algorithm. Additionally, the complexities of machine learning may exceed the expertise of a single neutral reviewer, raising doubts about whether a cursory inspection can adequately safeguard defendants’ rights (State v. Loomis, 2016).
3. Proposed Middle Grounds
Various middle-ground proposals have surfaced in scholarly debates. One suggestion is to require that proprietary vendors submit to third-party auditing at regular intervals, releasing summarized findings on fairness metrics and error rates without disclosing the entire code (Rudin, 2019). Another approach involves partial disclosures, where the overarching model architecture, feature importances, and validation studies are revealed, but certain proprietary “ingredients” remain sealed (Selbst & Barocas, 2018). However, these partial solutions may still leave defendants at a disadvantage, unable to precisely pinpoint flaws in the data or interpret the interplay of variables that drive the final risk score (Corbett-Davies & Goel, 2018).
4. Open-Source Sentencing Tools
An alternative paradigm envisions publicly funded or philanthropic efforts to develop open-source sentencing tools (Barabas et al., 2020). In such cases, no vendor can claim trade secret protections that overshadow due process. Source code, training data (with appropriate privacy protections), and model documentation remain fully transparent, enabling collaborative audits and updates. While appealing in principle, open-source models require consistent funding, robust governance, and broad buy-in from stakeholders. Critics worry that malicious actors might “game” the system if they know precisely how risk scores are generated, although proponents counter that robust auditing and continuous updates can mitigate such concerns (Lum & Isaac, 2016).
5. Balancing Economic Interests with Constitutional Values
At a higher level, the friction between trade secrecy and due process highlights broader questions about how capitalism intersects with public governance. When the justice system delegates core functions to private companies, the interplay of profit motives and constitutional obligations can become fraught (Noble, 2018). Structural reforms may be necessary to realign these incentives, potentially through legislation that treats AI-based sentencing as a matter of public interest, overriding certain trade secret claims in favor of transparency. Such reforms could, for instance, impose mandatory licensing agreements or require the vendor to waive trade secret protections in the context of court challenges (Garvie, 2019). Only through a rethinking of this fundamental tension can the legal system ensure that commercial confidentiality does not eclipse the constitutional rights that anchor fair sentencing.
BIAS, DISCRIMINATION, AND THE ROLE OF TRANSPARENCY IN MITIGATION
Although transparency is largely a procedural imperative, it also serves a vital substantive function in detecting and reducing discrimination. AI-based sentencing can embed and amplify historical inequalities unless stakeholders can examine and contest the model’s inner workings.
1. Historical Roots of Systemic Bias
Decades of scholarship reveal that minority communities are disproportionately policed, arrested, and incarcerated (Alexander, 2012). These injustices translate into the datasets that fuel machine learning models, creating feedback loops wherein historically over-policed neighborhoods register as “high risk,” perpetuating a cycle of punitive surveillance (Lum & Isaac, 2016). As a result, even when developers claim race is excluded as a variable, proxies such as zip codes, unemployment rates, or prior arrest frequencies can reintroduce racial or economic bias (Barocas & Selbst, 2016).
2. Detecting Disparate Impact
Transparency allows researchers, civil rights advocates, and legal practitioners to measure the performance of a sentencing tool across distinct demographic groups. Such evaluations might examine false-positive rates (the algorithm incorrectly labeling low-risk individuals as high risk) and false-negative rates (failing to identify genuinely high-risk individuals) (Berk et al., 2021). If, for instance, Black defendants are consistently assigned higher false-positive rates, that indicates a disparate impact that violates equal protection norms. Without insight into the model’s features and decisions, these patterns can remain invisible, undermining any claims that AI-based sentencing is color-blind or purely objective (Angwin et al., 2016).
3. The Fairness-Accuracy Trade-Off
One of the core debates in algorithmic fairness is the perceived trade-off between accuracy and equalized outcomes. Some fairness metrics inherently conflict, meaning that optimizing for one criterion can worsen another (Corbett-Davies & Goel, 2018). Policymakers and judges, when faced with these technical nuances, may struggle to determine what kind of fairness is most legally or morally pertinent. Transparency does not solve this dilemma by itself but is a crucial first step. It reveals the model’s performance under various fairness definitions, enabling informed discussions about which trade-offs are acceptable within a justice framework. Absent transparency, developers could quietly choose whichever metric they prefer, shaping outcomes in ways that might run counter to public values or legal mandates (Berk et al., 2021).
4. Corrective Measures
When bias is discovered, stakeholders need mechanisms to recalibrate or retrain the model (Kusner, Loftus, Russell, & Silva, 2017). These approaches can include reweighting data, removing problematic variables, or even employing causal methods to identify structural pathways of discrimination. Importantly, such corrective measures demand ongoing transparency—not a one-time reveal. As new data enters the system, bias can resurface in unforeseen ways, calling for continuous audits (Raji et al., 2020). The overarching implication is that transparency is not just a single procedural step but part of a feedback loop, essential for real-time improvements in both procedural and substantive fairness (Noble, 2018).
5. The Larger Social Context
Finally, it is crucial to remember that even a perfectly transparent and meticulously calibrated AI tool cannot rectify deeply rooted social inequalities. The impetus to rely on risk assessment might itself stem from an era of mass incarceration and over-policing (Wacquant, 2009). Hence, while transparency in AI-driven sentencing is critical, it should be seen as one element in a broader tapestry of reforms aimed at addressing systemic racism, poverty, and the punitive tilt of criminal policy. Without parallel efforts to expand public defender resources, reform policing strategies, and invest in social programs, the best-intentioned algorithmic solutions may offer only marginal gains (Alexander, 2012). Yet, from a practical standpoint, transparency remains a linchpin: by illuminating the internal logic of AI systems, it offers a vantage point from which to advocate for deeper, structural change.
EMPIRICAL CASES AND REAL-WORLD CONSEQUENCES
Evidence from specific jurisdictions and AI deployments sheds light on how transparency, or the lack thereof, plays out in everyday sentencing practices. Examining these case studies bolsters theoretical arguments with on-the-ground observations.
1. COMPAS in Broward County
The widely publicized COMPAS system garnered intense scrutiny when ProPublica published an investigation showing that while Black defendants were far more likely to be incorrectly classified as high risk, White defendants were conversely mislabeled as low risk (Angwin et al., 2016). Although Northpointe (the developer) disputed the methodology, the underlying disparity became a rallying point, illustrating how a black-box tool could perpetuate racial inequities. Notably, defendants in Broward County rarely gained access to any meaningful explanation of their risk scores, making it exceedingly difficult for them to challenge potential errors or biases (Larson et al., 2016). The controversy stoked debates on whether reliance on proprietary AI in sentencing decisions was defensible under constitutional norms.
2. Public Safety Assessment (PSA) Deployments
In an effort to reduce jail populations, several jurisdictions adopted the Public Safety Assessment (PSA), which claims a more transparent approach than some proprietary rivals (Stevenson, 2018). Early studies suggested that PSA implementation correlates with reductions in pretrial detention without significant upticks in recidivism (Desmarais, Johnson, & Testa, 2021). Yet, critics contend that the PSA’s weighting mechanism is still only partly transparent, leaving some communities uncertain about why particular defendants receive unfavorable scores (Lum & Isaac, 2016). Differences in local policing practices can also distort the data upon which the PSA relies, pointing to the persistent need for robust oversight and the ability to adapt risk models to evolving social conditions (Eckhouse, Lum, Conti-Cook, & Ciccolini, 2019).
3. Judicial Perspectives and Behavioral Shifts
Empirical research on judicial behavior reveals a range of responses to AI tools (Green & Chen, 2019). Some judges view algorithmic assessments as a helpful reference, especially when dealing with high caseloads. Others worry about “automation bias,” wherein they might lean too heavily on the algorithm’s recommendation, presuming it to be objective. In interviews and observational studies, judges who had access to more transparent systems—where they could see which variables influenced the score—reported a heightened sense of agency, stating that the information enabled them to exercise discretion more confidently (Frederick & Stemen, 2012). Conversely, judges supplied only with a composite risk score often lacked the impetus or the means to question its validity.
4. Community and Advocacy Group Involvement
Beyond the courtroom, the question of transparency resonates with civil society. Nonprofit organizations and investigative journalists have played a leading role in revealing potential biases and inaccurate data in AI-driven sentencing tools (Noble, 2018). Grassroots efforts in certain municipalities have petitioned for ordinances banning black-box algorithms from local courts unless they meet stringent disclosure standards (Raji et al., 2020). While these campaigns often face pushback from vendors, they underscore a growing movement that insists technology deployed in public institutions must be openly scrutinized and democratically accountable.
5. Lessons from International Contexts
Outside the United States, some European jurisdictions experiment with AI in bail and sentencing decisions under the shadow of the EU’s stricter data protection and emerging AI regulations (European Commission, 2021). For instance, certain pilot projects in the United Kingdom incorporate algorithmic tools to classify offenders by needs and risks, but robust public debates persist about whether these systems are truly transparent or simply less visible to academic and media scrutiny. Observers highlight that while GDPR-inspired rules offer individuals some rights regarding automated decision-making, the specific domain of criminal sentencing remains fraught with legal ambiguities (Ferguson, 2017). Nevertheless, these international case studies similarly illustrate that transparency is a universal concern whenever AI tools influence human freedoms.
LEGISLATIVE AND POLICY RESPONSES: FROM PROPOSALS TO IMPLEMENTATION
While the academic debate on transparency and AI sentencing is robust, real-world legislative and policy frameworks remain fragmented. This section examines the most significant proposals, existing regulations, and the gaps that persist.
1. Federal Initiatives in the United States
On the U.S. federal stage, measures like the Algorithmic Accountability Act have been introduced but not comprehensively enacted (Malcolm, 2020). Although aimed primarily at consumer-focused algorithms, these proposals often outline principles—impact assessments, transparency reports, fairness testing—that could extend to criminal justice. However, no sweeping federal statute exclusively regulates AI-driven sentencing, leaving the matter largely to state-level experimentation. Federal jurisprudence, guided by Supreme Court precedents on due process and evidentiary standards, sets overarching boundaries but offers limited specific directives for managing black-box systems (Re & Solow-Niederman, 2019).
2. State and Local Legislation
Certain states, such as California and Washington, have pursued data privacy and transparency laws that could indirectly affect sentencing algorithms (House Bill 1655, Washington State Legislature, 2021). In practice, though, few statutes explicitly mention AI-based sentencing. Instead, some jurisdictions form temporary committees to explore emerging technologies, occasionally releasing non-binding guidelines or best practices. This patchwork approach creates uneven protections: defendants in one state might benefit from partial transparency mandates, while those elsewhere face fully opaque systems. Also, local politics—shaped by budget constraints, lobbying by tech vendors, and public attitudes toward crime—can hamper the formation of robust legislative protections (Garvie, 2019).
3. International Approaches and the EU AI Act
In contrast, the European Union has taken a more systematic approach by drafting the Artificial Intelligence Act, which classifies AI systems used in law enforcement and the judiciary as “high risk” (European Commission, 2021). The proposed regulations call for enhanced transparency, accountability mechanisms, and oversight for such applications, although the legislation stops short of enumerating a blanket right to explanation in the context of criminal trials. Member states retain discretion in transposing EU directives into local law, potentially leading to varied implementations. Nevertheless, the EU’s emphasis on rigorous risk assessment and documentation sets an influential precedent that could push other regions toward more stringent frameworks (Ferguson, 2017).
4. The Role of Judicial and Bar Associations
Given the legislative gaps, professional organizations like the American Bar Association (ABA) have stepped in to propose guidelines. While not legally binding, such directives can shape courtroom practices and inform judicial attitudes. The ABA, for instance, has issued reports cautioning judges about overreliance on AI outputs and urging broader consideration of factors like model validation, transparency, and the risk of biased data (Slobogin, 2017). Such guidance can influence local court rules and training programs, gradually shifting the cultural landscape around AI acceptance. However, without formal legislative muscle, these guidelines depend heavily on voluntary compliance, limiting their enforceability.
5. Toward a Coherent Policy Agenda
A genuinely protective regime for AI-driven sentencing might involve harmonizing these disparate initiatives under an overarching policy agenda. Such a framework would likely include:
Mandatory Transparency Requirements: Explicit rules compelling vendors to disclose model architecture, relevant data features, and performance metrics.
Independent Auditing and Certification: Regular audits by external bodies, with the power to revoke certification if a tool exhibits persistent bias or fails to meet reliability thresholds.
Defendant-Centered Provisions: Clear articulation of how defendants can request disclosures, challenge the algorithm’s accuracy, or introduce contrary evidence.
Adaptive Regulatory Mechanisms: Structures that enable fast updates to legal standards, mirroring the rapid pace of AI innovation.
Realizing such a coherent policy agenda remains a formidable challenge, requiring coordinated efforts among lawmakers, judicial authorities, technologists, and civil rights advocates. Yet, the stakes—individual liberty, constitutional integrity, and public trust—underscore the pressing need for systematic rather than ad hoc solutions.
RECOMMENDATIONS FOR FUTURE PRACTICE
Building on the extensive literature, philosophical arguments, legal precedents, and empirical case studies, we can now propose a series of recommendations aimed at reconciling AI’s growing role in criminal sentencing with the imperatives of transparency and the right to explanation.
Legislate a Right to Explanation in Criminal Proceedings
Statutory Mandates: Enact legislation obligating that any AI tool influencing sentencing decisions be accompanied by a comprehensible explanation of how the tool generated its output. This should extend to disclosing data sources, model factors, and performance metrics.
Public Oversight: Include provisions allowing for public scrutiny, especially by defense counsel, advocacy groups, and independent researchers, ensuring that “black-box” claims do not override due process concerns.
Adopt Open-Source or Partially Open-Source Models
Government-Funded Development: Invest public resources in creating open-source AI risk assessment tools, thereby eliminating trade secret barriers.
Collaborative Maintenance: Encourage academic institutions and community partners to audit, refine, and monitor these models, promoting iterative improvements guided by collective expertise.
Institutionalize Regular and Rigorous Audits
Independent Auditors: Mandate periodic audits by neutral entities with expertise in machine learning, criminal justice, and anti-discrimination law.
Transparent Results: Require that the findings of these audits be publicly accessible, including any detected biases or inaccuracies, as well as the steps taken for remediation.
Enhance Judicial and Legal Education
AI Literacy Training: Integrate machine learning fundamentals, interpretability methods, and fairness metrics into continuing education programs for judges, prosecutors, and defense attorneys.
Practice Guides: Develop accessible “bench books” or digital toolkits illustrating how to assess and question AI-generated evidence effectively.
Protect Defendants’ Rights to Challenge Algorithmic Evidence
Discovery Rights: Strengthen discovery rules so that defendants can access the data and methodological details that led to their specific risk classification.
Expert Assistance: Provide defendants—especially indigent ones—access to court-appointed or publicly funded data experts who can interpret and dispute algorithmic evidence.
Implement Hybrid Explanations Tailored to Criminal Cases
Local and Global Explanations: Integrate both post-hoc and intrinsic explainability methods, ensuring that judges and defendants understand the model’s broader logic as well as case-specific factors.
Validation Studies: For each jurisdiction, conduct local validations to confirm that the model’s predictions align with regional demographics and do not systematically disadvantage particular groups.
Promote Broader Criminal Justice Reforms
Contextual Awareness: Recognize that AI is not a panacea for structural inequities. Align sentencing reform with efforts to reduce over-policing, fund rehabilitation, and support communities historically marginalized by the criminal justice system (Alexander, 2012).
Interdisciplinary Committees: Form committees that include sociologists, ethicists, community representatives, and data scientists to ensure that AI deployments in sentencing honor human dignity and equity.
By enacting these recommendations, policymakers and practitioners can more fully integrate AI into the criminal justice system without sacrificing the essential tenets of due process. Transparency and the right to explanation, as this paper has stressed, are not merely technical details but cornerstones of an ethical and legally sound sentencing process. While the road to implementation is paved with complexities—political, financial, and technological—each of these steps offers a tangible route toward reconciling advanced data analytics with the democratic ideals of fairness and accountability.
DISCUSSION
The formidable expansion of AI within criminal justice underscores a critical moment in legal and ethical history. Even as sentencing systems adopt advanced tools capable of analyzing vast datasets, the fundamental question persists: Can these models preserve the core values of procedural fairness and human dignity on which the legal system rests? This paper’s exploration reveals that transparency, exemplified through robust explainability, is a linchpin for preventing AI from morphing into an unassailable arbiter of human fate.
From a legal standpoint, the impetus to maintain open processes draws on constitutional doctrines mandating that defendants confront adverse evidence. However, the novelty of AI challenges these doctrines in unprecedented ways. Traditional evidentiary frameworks did not anticipate black-box software capable of synthesizing millions of data points into a single risk score, effectively compressing or even obscuring the underlying reasoning. While courts grapple with whether disclaimers or partial disclosures suffice, the broader consensus among scholars and advocates is that intangible AI processes can violate the spirit of procedural fairness, particularly for vulnerable defendants (Citron, 2008; Rudin, 2019).
In tandem, computer scientists and ethicists stress that interpretable design and careful auditing can mitigate issues of bias and arbitrariness. Yet, creating truly explainable AI in a domain as high-stakes as sentencing demands specialized solutions. Explanations must address not just local factors for an individual defendant but also the systemic patterns that might produce biased outcomes. Moreover, partial or superficial explanations risk breeding complacency among judges, who might assume that a simplified narrative reflects the entire algorithmic logic. If not coupled with genuine accountability measures, such as ongoing audits and external oversight, these post-hoc justifications can devolve into little more than window dressing (Selbst & Barocas, 2018).
The tension between proprietary secrecy and constitutional rights points to deeper structural dilemmas about privatization in the public sphere. Entrusting sentencing to commercial entities, shielded by trade secret law, can erode democratic oversight. Whether solutions lie in open-source alternatives, mandatory audits, or adjusted licensing frameworks remains an open question, but the moral imperative is unmistakable: the law must not permit economic interests to overshadow the rights of individuals facing the possibility of incarceration.
Finally, the discussion brings to light that issues of transparency and explanation form part of a broader tapestry of reforms. Mass incarceration, racial disparities in law enforcement, and the historically punitive orientation of many justice systems suggest that even the most transparent AI tool might only chip away at systemic inequities. Nevertheless, guaranteeing an explanation right is a step forward, as it empowers defendants, attorneys, and the public to engage in substantive dialogues about how justice is administered. It opens the door for more nuanced discourse on whether risk assessment tools align with rehabilitative ideals, restorative justice, or more comprehensive transformations of the punitive framework. Thus, while technology alone cannot rectify centuries of structural inequality, it can become a catalyst for reflection and progress when wielded under conditions of openness and democratic accountability.
CONCLUSION
AI-driven sentencing represents both a significant innovation and a profound challenge to foundational legal principles. In seeking to optimize judicial efficiency and reduce subjective bias, the incorporation of machine learning tools has inadvertently tested the boundaries of transparency, accountability, and defendant autonomy. This paper has argued that ensuring procedural fairness in such a technological milieu hinges on two intertwined precepts: the demand for transparency and the right to explanation.
Tracing the evolution of sentencing tools from rudimentary actuarial instruments to complex, data-hungry algorithms has illuminated how computational sophistication can inadvertently amplify biases. Simultaneously, the black-box nature of many commercial products conflicts with the defendant’s constitutional entitlement to inspect and contest the evidence. While courts have begun to recognize these perils, widespread judicial hesitation persists, with partial disclaimers and circumscribed disclosures often filling the void where robust accountability ought to reside.
A thorough engagement with legal doctrines, philosophical theories, and empirical observations substantiates the claim that transparent, explainable AI is not just a matter of technical best practice but a moral and constitutional necessity. Kantian notions of autonomy, Rawlsian conceptions of public reason, and broader theories of care ethics converge to underscore that sentencing must remain comprehensible and subject to open critique. On the policy front, examples such as the European Union’s Artificial Intelligence Act and various state-level initiatives in the United States signal a gradual shift toward more regulated deployments, yet significant gaps remain.
Ultimately, this paper posits that bridging these gaps requires a multipronged approach. Legislators must craft enforceable legal standards that prioritize the right to explanation, vendors should be compelled or incentivized to adopt interpretable designs, and judicial professionals need to develop the expertise to critically interrogate AI outputs. Where proprietary interests clash with constitutional values, the law must decisively uphold due process. And where structural inequities persist in datasets, continuous audits and recalibrations become imperative for any meaningful notion of equality under the law.
By enacting these reforms and fostering a legal culture that demands openness, AI can be directed toward equitable rather than opaque or discriminatory ends. Far from undermining the principles of justice, machine learning might then serve as an auxiliary tool—supporting judges, reducing arbitrary disparities, and advancing the rule of law. This brighter possibility, however, hinges on steadfast commitments to transparency and explanation, ensuring that even in an era of unprecedented technological prowess, fundamental rights and human dignity remain at the forefront of criminal justice.
ACKNOWLEDGMENTS
I would like to express my profound gratitude to my research mentor, Dr. Alex Rodriguez, for his perceptive guidance, incisive feedback, and generous encouragement throughout this process. His expertise in both the technical dimensions of algorithmic analysis and the legal contours of due process was indispensable to the development of this manuscript. I also wish to thank the broader academic community at Fictitious University for fostering a collaborative environment that supported this research from its inception.
REFERENCES
Alexander, M. (2012). The new Jim Crow: Mass incarceration in the age of colorblindness. The New Press.
Andrews, D. A., Bonta, J., & Wormith, J. S. (2006). The recent past and near future of risk and/or need assessment. Crime & Delinquency, 52(1), 7–27.
Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016). Machine bias. ProPublica.
https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
Arrieta, A. B., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., … Herrera, F. (2020). Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Information Fusion, 58, 82–115.
Bandura, A. (1999). Moral disengagement in the perpetration of inhumanities. Personality and Social Psychology Review, 3(3), 193–209.
Barabas, C., Dinakar, K., Ito, J., Virza, M., & Zittrain, J. (2020). Interventions over predictions: Reframing the ethical debate for actuarial risk assessment. Proceedings of Machine Learning Research, 81, 1–15.
Barocas, S., & Selbst, A. D. (2016). Big data’s disparate impact. California Law Review, 104(3), 671–732.
Berk, R. (2017). An impact assessment of machine learning risk forecasts on parole board decisions and recidivism. Journal of Experimental Criminology, 13(2), 193–216.
Berk, R., Heidari, H., Jabbari, S., Kearns, M., & Roth, A. (2021). Fairness in criminal justice risk assessments: The state of the art. Sociological Methods & Research, 50(1), 3–44.
Berk, R., Sherman, L., Barnes, G., Kurtz, E., & Ahlman, L. (2015). Forecasting murder within a population of probationers and parolees: A high stakes application of statistical learning. Journal of the Royal Statistical Society: Series A (Statistics in Society), 178(2), 513–533.
Burrell, J. (2016). How the machine ‘thinks’: Understanding opacity in machine learning algorithms. Big Data & Society, 3(1), 1–12.
Citron, D. K. (2008). Technological due process. Washington University Law Review, 85(6), 1249–1313.
Corbett-Davies, S., & Goel, S. (2018). The measure and mismeasure of fairness: A critical review of fair machine learning. arXiv preprint, arXiv:1808.00023.
Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993).
Desmarais, S. L., Johnson, K. L., & Testa, A. (2021). Race and gender bias in pretrial risk assessments: A systematic review. Criminal Justice and Behavior, 48(1), 60–80.
Eckhouse, L., Lum, K., Conti-Cook, C., & Ciccolini, J. (2019). Layers of bias: A unified approach for understanding problems with risk assessment. Criminal Justice and Behavior, 46(2), 185–209.
Elek, J. K. (2019). Rethinking risk assessment: The ethics of an evidence-based tool. Journal of Community Corrections, 29(1), 5–12.
European Commission. (2021). Proposal for a regulation of the European Parliament and of the Council laying down harmonised rules on Artificial Intelligence (Artificial Intelligence Act) and amending certain union legislative acts. Brussels, Belgium.
Ferguson, A. G. (2017). The rise of big data policing: Surveillance, race, and the future of law enforcement. NYU Press.
Frederick, B., & Stemen, D. (2012). The anatomy of discretion: An analysis of prosecutorial decision making. Judicature, 96(1), 12–19.
Friedler, S. A., Scheidegger, C., & Venkatasubramanian, S. (2021). On the (im)possibility of fairness. Communications of the ACM, 64(12), 136–143.
Garvie, C. (2019). Garbage in, garbage out: Face recognition on flawed data. Georgetown Law Center on Privacy & Technology.
Goodman, B., & Flaxman, S. (2017). European Union regulations on algorithmic decision-making and a “right to explanation.” AI Magazine, 38(3), 50–57.
Green, B., & Chen, Y. (2019). Disparate interactions: An algorithm-in-the-loop analysis of fairness in risk assessments. Proceedings of the Conference on Fairness, Accountability, and Transparency, 90–99.
Hildebrandt, M. (2016). Law as information in the era of data-driven agency. The Modern Law Review, 79(1), 1–30.
House Bill 1655, 67th Leg., Reg. Sess. (Wash. 2021).
Kant, I. (1785/2011). Groundwork of the Metaphysic of Morals. Cambridge University Press.
Kusner, M. J., Loftus, J., Russell, C., & Silva, R. (2017). Counterfactual fairness. In Proceedings of the 31st International Conference on Neural Information Processing Systems (pp. 4066–4076).
Larson, J., Mattu, S., Kirchner, L., & Angwin, J. (2016). How we analyzed the COMPAS recidivism algorithm. ProPublica.
https://www.propublica.org/article/how-we-analyzed-the-compas-recidivism-algorithm
Lum, K., & Isaac, W. (2016). To predict and serve? Significance, 13(5), 14–19.
Malcolm, D. (2020). The Algorithmic Accountability Act: Algorithmic profiling and consumer privacy. Journal of Law and Technology, 32(2), 249–276.
Noble, S. U. (2018). Algorithms of oppression: How search engines reinforce racism. NYU Press.
Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381–410.
Raji, I. D., Smart, A., White, R. N., Mitchell, M., Gebru, T., Hutchinson, B., … Barnes, P. (2020). Closing the AI accountability gap: Defining an end-to-end framework for internal algorithmic auditing. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, 33–44.
Rawls, J. (1971). A theory of justice. Harvard University Press.
Re, R., & Solow-Niederman, A. (2019). Developing artificially intelligent justice. Stanford Technology Law Review, 22(2), 242–282.
Reisman, D., Schultz, J., Crawford, K., & Whittaker, M. (2018). Algorithmic impact assessments: A practical framework for public agency accountability. AI Now Institute.
Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5), 206–215.
Selbst, A. D., & Barocas, S. (2018). The intuitive appeal of explainable machines. Fordham Law Review, 87(3), 1085–1139.
Silver, J. (2019). Deep learning and the future of AI. MIT Press.
Simon, J. (2007). Governing through crime: How the war on crime transformed American democracy and created a culture of fear. Oxford University Press.
Slobogin, C. (2017). Risk assessment pretrial and at sentencing: Standards of application. Judicature, 101(3), 58–67.
State v. Loomis, 881 N.W.2d 749 (Wis. 2016).
Stevenson, M. T. (2018). Assessing risk assessment in action. Minnesota Law Review, 103(1), 303–384.
Wacquant, L. (2009). Punishing the poor: The neoliberal government of social insecurity. Duke University Press.
