RESEARCH DOSSIER
Stereotypes, Disparities, and Legal Decision-Making
Abstract. Legal decision-making allocates suspicion, liberty, credibility, culpability, punishment, and access to remedies through institutions that combine human judgment with formal rules. This monograph examines how stereotypes, social categorization, evidentiary uncertainty, discretionary practices, and institutional arrangements produce or sustain disparities. It synthesizes experimental social psychology, mock-jury research, observational and quasi-experimental studies of policing, bail, charging, sentencing, and research on predictive algorithms. It distinguishes descriptive disparities from causal discrimination and develops a decision-stage framework for transparent measurement, institutional reform, and independent fairness evaluation.
Keywords: forensic and legal psychology; legal decision-making; stereotypes; racial disparities; judicial sentencing; bail decisions; implicit cognition; jury bias; disparate impact; procedural justice; algorithmic fairness; causal inference; Artificial Era
Scientific Overview: The Object of Inquiry
The distinctive object of this Research Dossier is the production of unequal legal decisions under conditions of social categorization, uncertainty, institutional discretion, and differing access to legal resources. A legal decision is an action with institutional consequences: investigating or stopping someone, authorizing a search, deciding a charge, setting release conditions, assessing testimony, reaching a verdict, determining punishment, or deciding eligibility for relief. Each action has a separate decision rule and a separate evidentiary record. The psychology of disparity must identify the particular action before proposing its causes.
A stereotype is a learned association or generalized belief connecting a social category to attributes, dispositions, or behavior. Stereotypic information may be explicit, rapidly activated, normatively endorsed, contested, or deliberately inhibited. Prejudice refers to evaluative orientations toward a group, whereas discrimination concerns differences in treatment or opportunity linked to group membership. A disparity is a difference in measured distributions of an outcome between populations. Discrimination, disparate impact, and unequal baseline exposure are related explanatory possibilities, not synonyms. Kang et al. (2012) Kang (2012) discuss these forms of bias in legal procedure; Arnold, Dobbie, and Hull (2022) Arnold (2022) show how an observable release-rate gap can be decomposed using quasi-random assignment rather than equated automatically with individualized animus.
The research questions are consequently layered. Which social representations alter perception, memory, threat evaluation, and credibility? Which legal stages amplify or constrain those mechanisms? How do formal rules, available alternatives, and asymmetric access to representation affect case outcomes? Under what designs is the effect of perceived race or another protected characteristic identifiable? Which improvements can change decisions without worsening other dimensions of accuracy, legitimacy, or equality?
This monograph is a critical integrative synthesis of original experiments, administrative research, quasi-experimental identification, methodological studies, and peer-reviewed research reviews. It is neither a preregistered systematic review nor a new experiment or quantitative meta-analysis. A recent systematic review of 347 methodologically rigorous studies on racial and ethnic differences in noncapital sentencing is treated as a secondary synthesis rather than reproduced as an independently audited census of the literature Spohn et al. (2026) Spohn (2026). The geographical concentration of much of the causal evidence in the United States is identified explicitly; its institutional assumptions must be reconsidered before transfer to other legal systems.
Historical Formation of the Research Problem
Investigations of legal inequality predate contemporary laboratory studies of implicit association. Social-psychological research contributed constructs for categorization and stereotyping; legal scholarship developed analyses of discretion, unequal procedural resources, race, gender, and institutional legitimacy. The later empirical turn brought more systematic comparisons of defendants, adjudicators, case assignments, prosecutorial decisions, police stops, and algorithmic prediction. The resulting field is plural: experimental psychology interrogates mechanisms, empirical legal studies model decision distributions, and causal inference examines the identifying assumptions under which those distributions support explanatory claims.
The Implicit Association Test described by Greenwald, McGhee, and Schwartz (1998) Greenwald (1998) provided a reproducible task for evaluating relative associations between categories and attributes. The IAT altered the vocabulary of implicit cognition without replacing direct evidence of conduct. Studies of professional judges later questioned whether knowledge of law alone eliminated automatic group associations. In a multi-part judicial study, Rachlinski et al. (2009) Rachlinski (2009) reported implicit racial associations among judges and findings compatible with both bias effects and motivated control. The evidence supported studying judicial judgment as psychologically regulated decision-making rather than presuming perfect impartiality from professional status.
Forensic research also increasingly studied individual decisions in relation to institutional pipelines. Rehavi and Starr (2014) Rehavi (2014) linked federal cases from arrest through sentencing and emphasized charging choices as an important pathway. Arnold, Dobbie, and Yang (2018) Arnold (2018) used quasi-random bail-judge assignment to study differential treatment at the margin. Pierson et al. (2020) Pierson (2020) combined large-scale stop data with temporal and threshold comparisons. These advances moved the question from whether two observed groups have unequal average outcomes to how differences arise at identifiable decision points.
The history matters because a single contemporary vocabulary can obscure older distinctions. Discretion may describe lawful latitude, under-specification of evidence, or opportunity for discriminatory action. Judicial independence may protect against political coercion yet leave unexplained between-judge variation. A numerical risk score may standardize parts of judgment while inheriting the measurement assumptions of past decisions. Each historical development changed the locus at which inequality could be measured and contested.
Construct Definitions and Levels of Explanation
A coherent analysis distinguishes the individual, interpersonal, organizational, and system levels. At the individual level, perception, attention, associative accessibility, and thresholds for classifying ambiguous evidence matter. At the interpersonal level, interview styles, perceived status, language, demeanor, and representation shape an encounter. At the organizational level, guidelines, workloads, disciplinary incentives, supervision, courtroom practices, plea norms, and recordkeeping influence repeated decisions. At the system level, legal categories, criminalization, neighborhood exposure, unequal resources, data-generating mechanisms, and cumulative histories determine who appears at each subsequent stage.
A conditional disparity* compares outcomes after accounting for specified measured variables. Its interpretation depends on whether controls represent legally relevant case differences, downstream consequences of previous unequal treatment, colliders created by selection, or variables measured with differential error. In criminal justice, a prior record might be legally relevant to an immediate sentencing rule while also reflecting earlier patterns of enforcement. Conditioning on it can answer a narrow legal question while concealing a broader cumulative pathway. A defensible report states which question is being estimated.
An individualized counterfactual* asks what would happen to this decision under a specified change in information or perceived group assignment. A policy counterfactual* asks what happens to a population under a change in thresholds, admissible inputs, access to counsel, or decision procedure. Individual and policy counterfactuals have different treatment definitions. A standardized hypothetical case can manipulate apparent race experimentally; an administrative database cannot generally reproduce that intervention without assumptions about omitted facts and selection.
Legal outcomes also have multiple units. A police-stop disparity is commonly measured per driver, resident, journey, traffic exposure, or observed stop, none of which has the same denominator. A bail decision can be recorded as release, monetary conditions, detention duration, or later alleged misconduct. Sentence severity may concern the incarceration decision, sentence length conditional on incarceration, or the combined expected custodial burden. Researchers must preserve those denominators and time horizons before interpreting a gap.
Stereotype Activation, Perception, and Threat Appraisal
Theories of associative activation predict that prior exposure and context affect how rapidly particular interpretations become available. In five studies involving police officers and undergraduate participants, Eberhardt et al. (2004) Eberhardt (2004) found bidirectional links between racialized cues and crime-related attention in laboratory tasks. The work is significant for identifying perceptual mechanisms: social meaning may shape what information becomes salient before a formal evidentiary judgment is articulated. Its paradigms do not directly measure wrongful arrests or the average effect of an actual policing policy.
The shoot/don't-shoot task developed by Correll et al. (2002) Correll (2002) tested rapid discrimination between armed and unarmed targets, examining how racial cues interact with response deadlines and error rates. The crucial dependent variables include latency, false alarms, and missed signals. The task illustrates how a decision threshold and information sensitivity can be separated conceptually; performance in a simulated task is one component of research on threat stereotypes rather than a direct estimate of officer-involved shooting rates.
Research by Wilson, Hugenberg, and Rule (2017) Wilson (2017) also investigated racialized estimates of physical size, formidability, and threat. Such findings connect seemingly descriptive perception to evaluative consequences. Their legal relevance lies in hypotheses about judgments of danger, reasonable force, and culpability. Those hypotheses require tests of the whole decision situation, including environmental uncertainty, departmental instructions, and accountability.
Visual stereotypicality is distinct from a categorical group label. Eberhardt et al. (2006) Eberhardt (2006) reported that perceived stereotypical appearance among Black defendants in cases with White victims predicted capital-sentencing outcomes after controls. Because the study was observational and case-specific, the finding cannot identify a unique mental mechanism in each juror or establish identical associations in other countries. It nevertheless shows why equal treatment analysis sometimes needs within-category measures, not only broad nominal race variables.
Implicit Cognition and the Problem of Behavioral Prediction
Implicit measures index task performance under structured association conditions. They are informative about associative regularities and may predict behavioral criteria in aggregate. They should not be treated as direct readings of a person's motive or as calibrated tests of whether that person will discriminate in a particular courtroom tomorrow. This distinction is critical for legal fairness, where unsupported individualized diagnostic inferences would themselves threaten procedural integrity.
The meta-analysis by Kurdi et al. (2019) Kurdi (2019), based on 217 research reports, found significant but heterogeneous relations between implicit measures and intergroup behavior, with separate contributions from explicit measures. Heterogeneity means that task, criterion, setting, base rate, and sample composition matter. Payne, Vuletich, and Lundberg (2017) Payne (2017) argued for attending to the aggregate environmental and cultural meaning of implicit associations. These approaches reorient assessment from blaming a single hidden mental trait toward examining how shared cues interact with institutional settings.
It follows that education alone cannot be assumed to produce durable changes in courtroom behavior. A network meta-analysis of procedures designed to change implicit measures by Forscher et al. (2019) Forscher (2019) found that shifts in task measures were not accompanied by reliable evidence of corresponding changes in explicit behavior across the evaluated interventions. This result is an important constraint on claims that brief bias awareness sessions, by themselves, repair structural disparities. Lasting effects require direct measurement of actual decisions, sufficient follow-up, and explicit implementation fidelity.
From Individual Categorization to Institutional Disparity
Individual errors and group disparities can become coupled through repeated legal stages. A person may face a different likelihood of police contact, evidence search, charge severity, pretrial detention, plea bargaining leverage, conviction, and sentence even when no single stage explains the entire final difference. Early decisions change the composition and observed characteristics of cases reaching later institutions. The cumulative record becomes both a result of previous decisions and an input to subsequent ones.
This mechanism creates an identification problem. A model estimated only on sentenced defendants excludes those never stopped, prosecuted, or convicted. A model of convictions among charged individuals conditions on a group selected by charging discretion. An observed conditional gap therefore may understate cumulative effects even when a given prosecutor or judge follows identical rules within a selected case set. Conversely, some raw disparities reflect offense composition, evidence, or legally relevant distinctions, and these differences must be measured rather than assumed away.
Cumulative disadvantage should be investigated as a sequence of transitions, preferably with linked individual-level administrative histories and explicit selection models. Rehavi and Starr (2014) Rehavi (2014) demonstrated the interpretive value of connecting early-stage charges and final sentences. Pager (2003) Pager (2003) demonstrated, using a field audit of job applicants, that a criminal record can subsequently alter access to employment. This downstream finding concerns labor-market decisions, not judicial motives, but it explains why an institutional penalty can propagate into later conditions relevant to well-being, economic status, and reentry.
What Counts as Evidence of Discrimination?
A group outcome difference is a descriptive finding. An experimental difference produced by randomized presentation of defendant attributes is evidence about the manipulated signal within the experimental task. A regression-adjusted disparity estimates an association conditional on model specification. A quasi-experiment using judge assignment provides a stronger identification strategy under its allocation, exclusion, and monotonicity assumptions. A legal conclusion of discrimination depends additionally on governing law and its applicable evidentiary standards. These categories should be reported without collapsing one into another.
Some decisions are evaluated by outcome disparity; others by threshold disparity. For searches, comparing hit rates alone can be misleading because the risk distribution among those searched is itself selected by enforcement. Threshold tests use additional assumptions to estimate the level of suspicion at which action occurs. For bail, the relevant misconduct outcome is observed only when defendants are released. Selective observation obstructs direct comparisons of risk between detained and released groups. Arnold, Dobbie, and Hull (2022) Arnold (2022) address this challenge using quasi-randomly assigned judges.
This scientific separation allows more precise institutional improvement. A fairness problem may arise through unequal exposure, measurement error, differential thresholds, inaccurate prediction, unequal legal resources, procedural disrespect, or an interaction among them. Each candidate mechanism calls for a different empirical test and a different intervention.
Police Contact, Search Thresholds, and Unequal Exposure
Police discretion begins well before an arrest. Deployment decisions determine where officers are present; enforcement priorities determine which violations receive attention; suspicion judgments determine which people are approached; search decisions determine which encounters escalate. Measuring a stop rate against a residential population can confound differences in travel, area policing, and traffic exposure. Measuring search success only among searched people can introduce selection bias. A scientifically usable account therefore defines the unit of encounter, the comparison population, the legally relevant decision rule, and the temporal ordering of its inputs.
Pierson et al. (2020) Pierson (2020) analyzed a large administrative dataset of nearly 100 million traffic stops across the United States. Their study used variation in visibility around sunset, alongside search and contraband-detection analyses, to examine racial differences in stop and search decisions. The veil-of-darkness design compares stops at times when race is differentially observable, subject to assumptions about traffic composition and other time-varying factors. The search-threshold work uses information beyond raw hit rates to infer differences in applied standards of suspicion. These methods support evidence of racial disparities in decision thresholds in studied settings while leaving important questions about unobserved exposure and local implementation.
Laboratory work on rapid threat judgments proposes psychologically plausible processes underlying some encounter decisions. Correll et al. (2002) Correll (2002) and Eberhardt et al. (2004) Eberhardt (2004) illuminate response competition and category-linked attention. Translating these effects into actual police outcomes requires accounting for selection, procedures, tools, supervisory practices, officer experience, and whether the relevant behavioral criterion is an erroneous stop, unnecessary search, use of force, or unjustified arrest. The existence of an associative response bias does not identify any specific officer's culpability.
An important reform implication follows from this multi-stage structure: better records can be as consequential for evaluation as better individual training. Encounter-level data with consistent reason codes, legal grounds, search outcomes, location, encounter time, and missingness indicators make specific questions testable. Auditors should identify changes in traffic patterns and enforcement priorities before attributing differences to a change in attitudes.
Pretrial Liberty and the Psychology of Risk
The bail decision combines perceived public safety, anticipated court appearance, statutory constraints, evidentiary uncertainty, and the cost of detention. Individuals may experience pretrial detention as punishment in practice even though the proceeding determines release conditions before adjudication. In psychological terms, a judge or magistrate must estimate a future event from incomplete evidence while applying rules about acceptable error. Such judgments are sensitive to framing, representative examples, workload, time pressure, and perceived threat, yet observed differences require methods that can distinguish these mechanisms from legal-case composition.
Arnold, Dobbie, and Yang (2018) Arnold (2018) introduced a marginal-outcome framework using quasi-random assignment of defendants to bail judges in Miami and Philadelphia. Their comparison asks whether released defendants at the decision margin have equal rates of pretrial misconduct by race under conditions in which judges are otherwise unbiased according to the specified model. The evidence indicated racial bias against Black defendants and suggested inaccurate danger stereotypes as a possible source. The framework identifies differential decision behavior more strongly than it identifies an individual judge's conscious rationale.
In a later analysis, Arnold, Dobbie, and Hull (2022) Arnold (2022) estimated disparate impact in New York City using quasi-random judicial assignment and a method designed to address missing risk outcomes for detained defendants. The study reported that approximately two-thirds of the observed Black–White release-rate disparity was attributable to disparate impact under its identification framework. This estimate is specific to the studied population, outcome definitions, periods, and identifying assumptions; it is not an invariant coefficient of prejudice across legal systems.
The deeper psychological issue is that a risk forecast is neither the same as a legal recommendation nor a warrant for a particular liberty restriction. A judge could accurately assess a chance of nonappearance but still choose a policy that weighs the human and procedural cost of detention differently. Research must distinguish accuracy of forecasts, appropriateness of thresholds, equality of treatment conditional on relevant evidence, and distributive effects of the chosen policy. These are distinct evaluative variables, not competing labels for one quantity.
Prosecutorial Charging and Plea Dynamics
Prosecutors possess institutional powers that determine the menu of legally available outcomes. Charge selection, mandatory minimum exposure, offers, evidentiary disclosure, and willingness to dismiss allegations affect bargaining leverage before any judicial sentence is announced. Defendant risk preferences and uncertainty interact with detention conditions, access to counsel, expected trial outcomes, and time costs. These pressures make an observed plea difficult to interpret as a simple expression of freely revealed preference.
In linked federal administrative records, Rehavi and Starr (2014) Rehavi (2014) examined how observed racial sentence differences arise across stages. Their analysis found substantial sentencing differences conditional on recorded arrest offenses and other characteristics and identified initial charging decisions, especially charges carrying mandatory minimum sentences, as major contributors. The results challenge any study that treats the sentence imposed by a judge as if it were determined independently of prosecutorial choices. The reported numerical findings are tied to the federal data and the authors' modeling strategy, with continuing scope for omitted characteristics and cross-jurisdictional variation.
Psychological research into plea decisions must also attend to defendants' comprehension of probabilistic outcomes. Perceived alternatives are shaped by legal advice, language accessibility, detention, fatigue, trauma histories, and trust in institutions. An apparently uniform plea offer may be experienced differently by defendants with unequal access to resources. Experimental work can isolate specific informational frames, whereas linked administrative data can trace actual choices and later consequences. Both designs are necessary because the empirical decision is constructed by the institution and interpreted by the individual.
The related canonical dossier on Plea Bargaining and Defendant Decision-Making is a distinct scientific object in the forensic map; this monograph addresses plea processes only as a stage in the production of disparity. Comprehensive analyses of plea validity, comprehension, legal rights, and defendant cognition belong in that separate study.
Jurors, Group Deliberation, and Evidentiary Interpretation
Juries transform multiple individual readings of evidence into a collective legal determination. The process includes initial impressions, evaluation of source credibility, memory of testimony, group discussion, normative pressure, and negotiated interpretations of legal instructions. Social categories can influence early inferences, but deliberation can also expose inconsistent assumptions and widen the set of considered facts. Understanding this system requires examining both individual verdict preferences and the deliberative process that produces a collective decision.
The meta-analytic review by Mitchell et al. (2005) Mitchell (2005) examined racial bias in mock-juror treatment of defendants. Mock-jury studies provide experimental leverage over defendant cues and evidentiary scenarios; their outcomes are sensitive to scenario content, participant recruitment, instructions, and the absence of real-world legal consequences. Thus, effect patterns should not be assumed to apply uniformly to actual criminal juries or all legal jurisdictions.
Sommers (2006) Sommers (2006) studied racial composition and deliberation in mock juries considering the case of a Black defendant. Racially heterogeneous groups exchanged a wider range of information than all-White groups. Importantly, changes among White participants in the diverse groups included more discussion of case facts and greater willingness to address race, indicating that group composition can influence the deliberative environment rather than merely the arithmetic of member votes. This experiment highlights an institutional design variable, although controlled mock-jury conditions do not establish the size of any intervention effect in every courtroom.
The group context also complicates simple claims about awareness. Bias can be reduced when a concern becomes salient and a decision maker adopts a correction strategy; the same salience can change legal narratives, introduce overcorrection, or trigger competing norms. The appropriate outcome is accurate application of the legal standard to valid evidence with accessible reasons, rather than the achievement of a predetermined verdict distribution.
Judges, Discretion, and Sentencing
Judicial sentencing decisions combine legal categories, guideline ranges, aggravating and mitigating evidence, institutional practices, and interpretations of culpability and rehabilitation. Professional experience may improve recognition of legally pertinent evidence, but individual and contextual variation remains an empirical question. Studies of judicial implicit associations and those of judicial real-case sentencing address different inferential targets.
Rachlinski et al. (2009) Rachlinski (2009) offered experimental evidence that professional judges are capable of implicit association effects and motivated correction in particular laboratory tasks. Observational research instead examines whether real sentence distributions differ after controlling for available case characteristics. These approaches should complement one another without converting a laboratory score into an explanation for every real sentence.
Cohen and Yang (2019) Cohen (2019) examined more than 500,000 federal defendants linked to randomly assigned sentencing judges. They reported politically patterned differences in the treatment of Black and female defendants, with larger differences when sentencing discretion was greater. Their empirical strategy makes judge assignment a lever for studying decision heterogeneity, but interpretation still depends on institutional assignment procedures and the definition of comparable cases. Their findings do not license claims that all judges of a given political background will act identically.
The study by Crow and Goulette (2022) Crow (2022) analyzed district composition and sentencing outcomes for federal drug and economic offenses. Its offense-specific and mixed results underscore the ecological problem: proportions of judges with specified identities within a district are not randomized assignments of individual defendants to a judge. District associations can be useful descriptive evidence while remaining less direct indicators of an individual decision mechanism.
A 2026 systematic review by Spohn et al. (2026) Spohn (2026) synthesized 347 methodologically rigorous studies of race and ethnicity in noncapital sentencing. Its reported pattern of harsher treatment of Black and Hispanic defendants provides important cumulative evidence, while variation across federal and state contexts, decisions to incarcerate, and sentence length calls for differentiated interpretation. The review's own study-selection protocol and included methods determine the strength and limits of its aggregate conclusions.
Capital Sentencing and the Interpretation of Stereotypicality
Capital sentencing is a particularly consequential test of how social information can interact with discretion because the outcome is irreversible and evidence is filtered through highly specific legal standards. The empirical literature includes studies of defendant race, victim race, regional practice, and physical stereotypicality. Each variable may have a different role, and the observed association between perceived appearance and outcome may depend on the subgroup of cases examined.
Eberhardt et al. (2006) Eberhardt (2006) reported that, among Black defendants charged in cases involving White victims, greater perceived stereotypicality was associated with a higher likelihood of capital punishment after statistical controls. This provides a basis to investigate nonverbal social-category cues within a nominal group. It does not establish a universal causal effect of appearance, nor can it replace analysis of defense resources, aggravating evidence, plea or trial selection, and jurisdictional differences.
The conceptual lesson extends beyond capital punishment. Formal legal equivalence at one stage may coexist with unequal interpretation of ambiguous information. A rigorous design should record the case-selection process, the information actually available to deciders, measurement reliability for stereotypicality ratings, and any variables used as covariates. It should also distinguish predictive association from legally adjudicated proof of discrimination in a particular case.
Gender, Intersectionality, and Other Dimensions of Legal Inequality
Legal outcomes are differentiated by gender, race, ethnicity, disability, age, class, language, and their intersections. A stereotype may be relevant because it links a category with perceived dangerousness, credibility, caregiving, deception, or competence. The meanings of such categories depend on the legal decision at issue; gendered expectations in custody adjudication need not operate like gendered expectations in criminal sentencing or credibility evaluation.
Cohen and Yang (2019) Cohen (2019) offer an example in which differential treatment patterns by defendant gender and race were analyzed jointly as characteristics of judicial decisions. Intersectional questions additionally require modeling how combinations of identities are treated, rather than assuming the sum of independent average group effects describes every subgroup. Data sparsity and measurement ambiguity can make fine-grained interactions unstable; uncertainty intervals and theoretically specified hypotheses are essential.
A disparity should never be represented as an immutable psychological property of a population. Differences in measured outcomes frequently arise through social exposure, institutional treatment, legal access, discriminatory history, and measurement practices. Strong research specifies these mechanisms and identifies the conditions under which an intervention could change them.
Credibility, Authority, and the Social Meaning of Testimony
Perceived credibility has central legal significance because witnesses, victims, defendants, experts, and institutions provide information that decision makers must evaluate under uncertainty. Group categories and status cues can shape initial plausibility and the standards of corroboration applied to statements. Yet testimony reliability is also affected by interview procedures, memory reconstruction, suggestive questioning, confidence presentation, and evidentiary provenance. These interacting causes should be separated rather than folded into a single general notion of credibility bias.
The forensic map assigns eyewitness confidence, testimonial contamination, investigative interviewing, and deception detection to separate canonical dossiers. FAL-13 instead studies differential credibility attribution across social positions and its aggregate implications. A practical design would use paired case materials in which legally irrelevant category cues vary while factual evidence, statement sequence, and instructions remain stable. The analysis must predefine whether the outcome is rated credibility, willingness to request corroboration, evidentiary weight, or ultimate disposition.
Expert status introduces a further issue: a technically confident narrative can be persuasive even when its error rates or uncertainty are inadequately disclosed. This problem concerns evidence quality as well as stereotyping, and the effects are best studied by jointly examining information provenance, expert disclosures, and decision-maker interpretation.
Disparities Beyond Criminal Adjudication
Legal decision-making includes civil litigation, administrative adjudication, eligibility determinations, family law, employment disputes, housing disputes, and access to protective remedies. These settings have different evidentiary burdens and institutional procedures. Extrapolating numerical findings from federal criminal sentencing to disability accommodations or asylum proceedings would be a serious methodological error, though social cognition and institutional discretion provide shared conceptual questions.
The courtroom analysis by Kang et al. (2012) Kang (2012) explicitly included both criminal and civil employment paths, illustrating how different stages can expose different actors to bias-related decision opportunities. The audit experiment by Pager (2003) Pager (2003), while outside courtroom adjudication, demonstrates downstream institutional consequences of a criminal record for employment opportunity. Together these lines of work support tracing consequences across the boundary of the legal institution, while keeping each outcome and sampling frame distinct.
In comparative research, rules for exclusion of evidence, legal representation, protections for affected persons, and the extent of written reason-giving vary considerably. The mechanisms of a stereotype can be studied across societies, but the estimand of a legal disparity must remain situated in its jurisdictional and historical structure.
Risk Instruments, Human Judgment, and Algorithmic Decision Support
Risk scores are increasingly studied as inputs to decisions about release, supervision, and recidivism. Their proper object is a defined outcome within a population and period: rearrest, reconviction, alleged nonappearance, technical violation, or a form of violence. These labels measure different phenomena and are affected by policing intensity, reporting, surveillance, and follow-up time. A model built to predict rearrest may reproduce variation in enforcement as well as individual behavior. A model that estimates an event probability does not itself specify what legal action should follow.
In Kleinberg et al. (2018) Kleinberg (2018), the bail setting was used to compare machine predictions with human judgments and explore the tradeoffs among pretrial crime, detention, and fairness. Their approach makes the selective-label problem central: when detention prevents observation of an individual's hypothetical behavior if released, straightforward supervised learning cannot know what the detained person would have done. The authors' results depend on substantive assumptions about judicial decisions and the observed case distribution. A useful system evaluation must study actual deployment, not merely compare retrospective prediction accuracy in a database.
Dressel and Farid (2018) Dressel (2018) evaluated the accuracy and fairness of a widely discussed recidivism prediction instrument in comparison with human judgments and simple models. Their results caution against treating proprietary complexity as evidence of superior validity or equity. Performance should be compared with transparent baselines on relevant populations using the same outcome definition. Aggregate accuracy, especially in imbalanced datasets, can obscure consequential group differences in false-positive rates, false-negative rates, and calibration.
Chouldechova (2017) Chouldechova (2017) demonstrated that several familiar mathematical fairness criteria cannot generally be satisfied simultaneously when outcome prevalence differs by group. This is an incompatibility result about defined metrics under specified conditions, not an instruction to disregard fairness. A legal institution must articulate which errors matter, who bears their costs, what evidence is legitimately available, and how accountability will be assigned when different criteria conflict.
The 2025 analysis by Arnold, Dobbie, and Hull (2025) Arnold (2025) extends the methodological discussion to building nondiscriminatory algorithms with selectively observed labels. Using quasi-random bail-judge assignment, the authors developed procedures to address conditional input disparities and selection. The significance of the paper is methodological: an algorithm's input features can preserve group-conditional distortions even when its stated optimization target appears neutral. Identifying those distortions demands credible assumptions about latent outcomes and selection rather than an automatic transformation of the observed data.
The systematic review by Lawson et al. (2024) Lawson (2024) examined research on disparate effects of risk assessment instruments. Its findings were mixed across settings and instruments, with numerous studies reporting no measured exacerbation or some reduction in disparities and others finding adverse effects. The review's assessment of study quality also limits strong generalizations. The evidence challenges the assumption that an algorithm invariably increases unequal outcomes, just as it challenges the assumption that numerical standardization is a sufficient fairness remedy. Instrument validity and the policies governing its use require separate evaluation.
Disparate Impact and the Fairness Metric Problem
Fairness debates become operational only after the relevant equality principle is specified. Demographic parity concerns aggregate selection or action rates. Equalized error rates concern the balance of mistakes conditional on an outcome. Calibration concerns whether a particular predicted score corresponds to comparable observed event rates across groups. Equal opportunity may prioritize a chosen class of errors or benefits. Legal fairness may also require individualized reason-giving, contestability, proportionality, and protection from prohibited factors. These criteria address different objects and can diverge in the same dataset.
The incompatibility explored by Chouldechova (2017) Chouldechova (2017) shows why a system cannot be certified as globally fair simply because one favored number is equalized. Nor is a statistical disparity in a predictive score by itself an adequate causal diagnosis. One must distinguish the model, the policy threshold, the human decision using the score, and the resulting downstream outcome. A well-calibrated score can coexist with unequal detention policy, and a score with identical average outputs can hide group-specific misclassification costs.
Fairness auditing should therefore report both predictive and allocative outcomes. Predictor validation addresses temporal stability, transportability, sampling, missing data, and measurement error. Policy evaluation addresses thresholds, alternatives to detention, sensitivity to legal constraints, override practices, appeal procedures, and distribution of burdens. Human overrides can mitigate or amplify model-related disparities; the direction is empirical rather than guaranteed by human oversight alone.
Data-Generating Processes and the Selection Problem
Administrative data encode prior decisions. An arrest label depends on behavior, exposure to police, enforcement strategy, report practices, and recording rules. A sentencing dataset includes only persons who passed through earlier procedural stages. A risk-model training set may be conditioned on earlier human judgments. Consequently, a predictive model can be accurate for its selected sample and still fail when deployed to a different population or used to change who becomes observable.
Selection can also operate through missing outcomes. For pretrial studies, misconduct if released is a potential outcome for persons who were detained, but it is not directly observed. A comparison of released people by race may be biased if judges release defendants at different thresholds. The marginal decision designs of Arnold, Dobbie, and Yang (2018) Arnold (2018) and Arnold, Dobbie, and Hull (2022) Arnold (2022) use variation across assigned judges to construct more informative tests under explicit assumptions.
A related challenge is the distinction between prediction and policy feedback. If a risk score causes more supervision in one group, it can increase detection of technical violations in that group and then make later models appear to confirm initial group differences. Effective longitudinal studies should record intervention exposure, monitoring intensity, data provenance, and the precise event definition; otherwise a feedback loop may be mistaken for persistent underlying risk.
Human–Artificial Systems as Joint Legal Decision Makers
Algorithmic decision support sits within a socio-technical arrangement: a person or institution selects the purpose of the model, defines acceptable error, chooses training outcomes, presents predictions, sets policy thresholds, and interprets exceptions. From the standpoint of Artificial Era research, the scientific object is a coupling between human and artificial decision processes with observable allocations of informational and institutional authority. This research perspective does not imply that a model possesses legal authority, consciousness, human psychological motives, or responsibility under any jurisdiction's law.
An artificial system can exert influence through ranking, default options, warning labels, confidence presentation, interface friction, and the availability of explanations. These features may change how judges and other professionals allocate attention or trust. Yet the behavioral effect must be measured by randomized interface comparisons, shadow deployments, prospective audits, or other designs that identify actual human reliance. Historical algorithm scores alone cannot show automation bias, appropriate deference, or deliberate refusal to use a recommendation.
The principal empirical question is whether the complete human–artificial workflow improves legally relevant outcomes with justified burdens. A system may increase average forecast accuracy while shifting error costs toward particular populations. It may reduce between-judge variation while standardizing flawed historical labels. It may improve documentation while weakening the ability of affected persons to understand or challenge the basis of a decision. Research therefore evaluates the joint policy rather than treating the artificial component as an independent moral explanation.
These issues overlap with the separately mapped dossier on Algorithmic Risk Assessment and Human–Artificial Decision Support in Justice. FAL-13 evaluates algorithmic fairness as one important pathway of legal disparities; the adjacent dossier is responsible for the full technical and behavioral science of risk scoring and human–artificial interaction in justice.
Structural Disparities and Institutional Design
A legal institution does not begin with a uniform distribution of cases. Social inequality influences patterns of exposure, legal resources, residential context, evidence availability, and prior records. Historical decisions can shape current administrative variables, and those variables may then be treated as neutral summaries of an individual's situation. Structural explanations address these linked distributions; psychological explanations address how people process and act upon the information and environments they encounter. The scientific challenge is to estimate how the levels interact.
A single outcome equation can conceal these interactions. For example, suppose observed sentence length depends on the initial charge, detention, plea decision, and judicial discretion. If protected characteristics affect early decisions, controlling statistically for charge and plea may estimate a relatively narrow direct judicial component while obscuring total institutional effects. If charge also reflects genuine case evidence not captured in the dataset, failing to control for it can exaggerate an apparent direct judicial disparity. The correct specification follows the question, a plausible causal diagram, and the data-generating history.
The distinction between direct and total pathways is illustrated by Rehavi and Starr (2014) Rehavi (2014), whose linkage of criminal cases across stages identifies how prosecutorial charging contributes to later observed sentences. Likewise, the police-stop analysis by Pierson et al. (2020) Pierson (2020) uses a decision-specific comparison rather than claiming that a general population disparity automatically reveals the exact source of a stop. Process-level accounts make intervention targets more explicit.
Fairness, Legitimacy, and Procedural Experience
Justice involves more than the final distribution of dispositions. People evaluate whether they were heard, treated respectfully, given an intelligible explanation, and judged under predictable and consistently applied procedures. Tyler (2006) Tyler (2006) developed an influential framework for procedural justice and legal legitimacy. Its significance for disparity research is that an institution's credibility can be affected by how decisions are reached and communicated as well as by the eventual outcome.
Procedural legitimacy is not a substitute for examining discriminatory outcomes. A respectful interaction can accompany an unjustified search; an explanation can reproduce an inaccurate stereotype in formal language; nominally identical procedures can burden people differently when interpretation services or legal assistance are unequal. The analytic task is to measure perceived fairness, procedural safeguards, accuracy, and distributive consequences separately while studying their interaction.
Experimental evidence on jury group composition Sommers (2006) Sommers (2006) suggests that decision environments can alter information exchange. Such findings complement broader theories of procedural legitimacy because both concern institutional structures affecting participation and reason-giving. A mature evaluation would test whether specific changes improve actual decision accuracy, participation, and equity over time and across affected groups.
Psychological and Institutional Consequences of Disparate Decisions
Legal decisions produce downstream exposure to stress, economic loss, restricted mobility, stigmatization, and uncertainty. A pretrial detention decision can affect employment or access to counsel; a conviction can alter later eligibility for housing or professional opportunities; a record can influence subsequent expectations of danger or trustworthiness. These processes provide plausible feedback mechanisms between legal treatment and later social outcomes. The strength and nature of each pathway require direct longitudinal evidence.
The field audit by Pager (2003) Pager (2003) tested how a criminal record affected employer responses and revealed a consequential postjustice mechanism of stratification. This is not a study of court sentencing but demonstrates why evaluating institutional decisions solely at the moment of judgment understates their possible social effects. A later outcome may reflect both the direct penalty and the way other institutions interpret that penalty as a social signal.
Research on perceived legitimacy Tyler (2006) Tyler (2006) further suggests that experience with legal authority may affect willingness to cooperate or comply. Causal pathways should be investigated without interpreting a group's reduced trust as a fixed disposition. Trust is a response to expectations and experiences that can vary as institutions change.
Comparative and Cross-Cultural Research
The evidence reviewed here draws heavily on U.S. policing, bail, federal sentencing, and English-language social psychology. The empirical phenomena are not limited to the United States, but legal definitions, prosecutorial systems, professional education, available demographic indicators, data protections, and historical relationships among groups vary. A procedural rule can have different consequences in adversarial, inquisitorial, administrative, and customary environments. No single U.S. estimate should be transferred as the expected effect size elsewhere.
Comparative work needs construct equivalence before numerical comparisons. A category called race, ethnicity, caste, citizenship, disability, or minority status may be measured differently across systems and can have distinct social meanings. A country's reliance on administrative records may prevent researchers from observing some protected attributes while enabling better measurement of others. Researchers should specify how categories were constructed, whose perspective shaped them, and whether reporting of group membership changes behavior.
A promising international program would harmonize estimands rather than force identical administrative variables. Studies could compare reasons for stops, standardized grounds for pretrial detention, evidentiary disclosure, judicial discretion ranges, and appeal outcomes using jurisdiction-specific lawful data practices. Theory would operate at the level of social categorization, uncertainty, institutional incentives, and procedural safeguards, while empirical estimates would remain local until transportability is demonstrated.
Research Designs and the Hierarchy of Inference
Studying legal disparity requires matching the design to the causal or descriptive question. Laboratory manipulations of race cues, witness language, appearance, or decision instructions isolate controlled stimulus effects and test psychological mechanisms. Their internal validity may be high while their ecological validity varies with realism, perceived stakes, and sample composition. Field audits extend randomization into consequential environments but are limited by accessible outcomes and ethical and legal constraints. Administrative records capture actual institutional decisions but ordinarily include unmeasured differences in evidence, discretion, and selection.
Quasi-experimental designs use sources of treatment variation that approximate random assignment under explicit conditions. Judge leniency as an instrument can identify effects near a decision margin when assignment is sufficiently as-if random, the instrument strongly influences release, and assumptions such as exclusion and appropriate monotonicity hold. The bail studies by Arnold, Dobbie, and Yang (2018) Arnold (2018) and Arnold, Dobbie, and Hull (2022) Arnold (2022) exemplify why a clearly defined margin can offer more causal leverage than a broad raw comparison.
Regression designs can be valuable when the model is explicit and covariates are reliably measured. Their estimates should be accompanied by sensitivity analyses for omitted confounders, alternate case classifications, changes in sample restrictions, and different definitions of severity. Researchers should avoid controlling automatically for variables that are consequences of the discrimination being investigated. A prior charge, arrest, plea, or previous detention may simultaneously act as a legally relevant predictor and a downstream mediator of earlier institutional treatment. A causal diagram identifying the temporal order helps make the estimand transparent.
The proper evidence hierarchy is question-dependent. An experimental mock trial may be best for testing whether a name alters perceived credibility; a linked administrative panel may be better for evaluating cumulative charge pathways; a quasi-random judge allocation may best identify marginal release effects. Triangulation is strongest when methods have different failure modes and converge on a specified conclusion. Agreement among several observational studies sharing the same measurement bias does not by itself establish causation.
Measurement, Reliability, and Missing Data
Operational definitions determine what a study can claim. The same term, bias, may denote an association in an IAT, a speeded-response difference, an unequal threshold, a residual conditional coefficient, a population outcome gap, or a legal violation. These indicators should have separate names in data dictionaries, reports, and synthesis tables. Without this discipline, an empirical claim can silently shift from one construct to another.
Measurement reliability matters when coding race or appearance, case seriousness, offense type, prior history, witness reliability, judicial discretion, and sentencing outcomes. Interrater agreement is particularly important for subjective measures such as perceived stereotypicality or narrative assessment of danger. Coders should be protected from knowledge of the outcome when feasible. The unit of analysis must reflect nesting of people within officers, judges, districts, and time; standard errors and validation splits should respect such clustering.
Missing data can be informative rather than random. Ethnicity may be absent from records differentially by agency; detained people lack observed release outcomes; some police encounters are recorded only if they result in searches or citations; appeals and postconviction corrections may be incompletely linked. A complete-case analysis can create a misleading comparison. High-quality studies state the missingness process, report its magnitude, conduct plausible sensitivity analyses, and differentiate empirically identified effects from bounds or assumptions.
A research synthesis should also distinguish absence of demonstrated bias* from demonstrated absence of bias*. Null effects can reflect a truly small effect, inadequate precision, weak measurement, restricted variation, or a procedural control that successfully mitigates the mechanism. Confidence intervals, sample design, and prespecified minimum meaningful differences give the null result interpretive value.
Replication, Generalizability, and Publication Bias
Replicability has more than one meaning in legal psychology. A laboratory effect should reproduce under equivalent manipulations and measurement conditions; a judicial disparity estimate should withstand alternate specifications and independent data; a policy intervention should preserve its effect after implementation at scale. A replicated association in student mock juries remains evidence about those conditions until external validity is tested. Similarly, a robust U.S. federal estimate is not a universal property of all courts.
Systematic reviews can map patterns and expose heterogeneity more effectively than selective narrative examples. The review by Spohn et al. (2026) Spohn (2026) gives broad coverage of the race–sentencing literature, but its reported numerical inventory should not be mistaken for a new census performed in this monograph. Researchers should inspect inclusion criteria, overlapping samples, differing legal controls, and the possibility that more publishable findings have been reported more often.
For implicit cognition, Kurdi et al. (2019) Kurdi (2019) documented heterogeneous relationships with behavioral criteria. For intervention research, Forscher et al. (2019) Forscher (2019) separated changes in implicit task measures from changes in behavior. These findings illustrate why a psychologically persuasive mechanism and a validated institutional remedy are different objects of evidence.
Intervention Research: Changing Procedures and Outcomes
Research-informed interventions can operate through improved evidence quality, less discretionary variation, clearer decision standards, disclosure of uncertainty, structured review, blinded evaluation where lawful and feasible, independent auditing, and reliable appeal processes. A favorable intervention claim requires evidence about the outcome it is intended to improve. For example, a debiasing program that reduces an IAT score should not be reported as reducing unjust detention without data on detention decisions.
The analysis by Kang et al. (2012) Kang (2012) articulated courtroom intervention options in light of behavioral science. Subsequent evidence constrains broad inferences from brief psychological exercises; Forscher et al. (2019) Forscher (2019) showed why direct behavioral follow-up is essential. Training can be one component of a package, but procedural control and outcome audit need independent evaluation rather than relying on good intentions or favorable participant surveys.
Interventions should be designed at the stage where the hypothesized harm occurs. Search-threshold changes should be evaluated against search rates, thresholds, evidence yields, and any displacement to other forms of enforcement. Bail reforms should examine detention, appearance, community outcomes, legal assistance, and uneven burdens. Sentencing reforms should examine both decisions to incarcerate and sentence duration, while monitoring changes in charging and plea strategy that can offset an apparent judicial improvement.
Randomized policy evaluations are particularly valuable where ethically and legally permissible. When randomization is unavailable, staggered implementations, threshold discontinuities, interrupted time series, matched comparisons, and natural variation can be informative if assumptions are tested. Programs should preregister key outcomes, publish error definitions, and plan independent replication. Unintended effects, including selective enforcement shifts and altered recording practices, belong in the primary evaluation.
Scientific Discussion: Competing Explanations
Several explanations can produce similar observable patterns. An unequal decision rate may reflect explicit hostility, automatic category associations, statistical stereotyping, risk estimates based on imperfect information, institutional incentives, uneven legal resources, differential exposure, or a mixture of mechanisms. Competing explanations differ in observable implications. A study that varies category cues while holding evidence fixed can isolate one form of discrimination; a study that follows charges across stages can examine how earlier discretion shapes later outcomes; a randomized policy trial can test the consequences of changing institutional conditions.
Statistical discrimination is an economic decision framework in which group averages can influence decisions under uncertainty. In legal settings, the fact that an actor uses a purported predictive generalization neither makes it accurate nor establishes its legal acceptability. The accuracy of the generalization, its sources, the availability of individual information, protected-class constraints, and the resulting decision rule all require separate evaluation. Arnold, Dobbie, and Hull (2022) Arnold (2022) explicitly considered mechanisms including racial bias and statistical discrimination in their analysis of bail decisions.
Research on judges also complicates monocausal accounts. Cohen and Yang (2019) Cohen (2019) reported politically patterned variation under differing levels of discretion. Crow and Goulette (2022) Crow (2022) found offense-dependent associations involving district judicial composition. These results are most informative when the legal institution is analyzed at the correct level rather than treating judge characteristics, district composition, and social stereotypes as interchangeable causal variables.
The contribution of social psychology is to identify how categorical meanings become decision-relevant under uncertainty. The contribution of causal empirical legal research is to locate their possible effects within the legally structured sequence. The contribution of measurement science is to distinguish biased outcomes from biased instruments and incomplete records. The strongest synthesis connects all three without reducing institutional disparities to an individual attitude score.
Research Limitations and Boundaries of Interpretation
Much of the causal literature concerns a limited set of jurisdictions, observed administrative outcomes, and groups for which data are recorded. The underlying populations are selected at multiple institutional stages. Courts differ in statutes, discretion, sentencing guidelines, recordkeeping, and the definition of prohibited discrimination. Results should be qualified by legal context, period, and the populations to which the identification strategy applies.
A second limitation concerns uncertainty about mental mechanisms. Behavioral discrimination is empirically compatible with more than one psychological explanation. Professional actors may be influenced by stereotypes without conscious endorsement; they may also act under constraints that generate disparate outcomes despite intentions to treat everyone equally. Conversely, explicit prejudices may produce decisions that are obscured by aggregate models. Inferential discipline means examining what the research actually identifies.
A third limitation concerns the audit of algorithmic justice. Historical prediction datasets do not usually reveal all counterfactual outcomes. Evaluation may be affected by changing treatment after model deployment, group-varying labels, intervention-specific measurement, and distribution shift. Mathematical fairness metrics offer useful partial diagnostics, but cannot substitute for specifying a defensible legal policy and assessing its observed impact.
The source base for this monograph includes checked publisher and repository records, PubMed entries, original research articles where publicly accessible, and systematic-review abstracts where full article access is limited. Bibliographic and scientific conclusions are bounded by that access. No external peer review, human-subjects approval, original dataset access, or exhaustive systematic-search procedure is asserted.
Future Research: A Cumulative Decision-Path Agenda
Future studies should link legal decision stages while protecting the rights and privacy of affected individuals. Data systems ought to preserve the distinction between a person's conduct and institutional detection of that conduct. Prospective research could measure how changes in search, charge, detention, plea, and sentencing practices interact, rather than assessing each stage in isolation. Models should specify total and stage-specific causal effects and include the possibility that reforms in one stage displace discretion elsewhere.
A second agenda concerns calibrated information presentation. Studies can randomly vary uncertainty disclosures, risk-score displays, explanations, and decision defaults while measuring legal professionals' actual reliance and override behavior. This agenda is especially relevant to human–artificial legal workflows because an intervention can change both model performance and human interpretation. Causal tests should measure the combined decision policy.
A third agenda concerns affected persons' procedural experience. Structured assessments could connect substantive outcomes with perceptions of respect, opportunity to participate, access to explanation, and successful challenge of factual errors. Cross-cultural work should validate these constructs locally and test how the relationship between procedural legitimacy and distributional equality differs across institutions.
Finally, intervention research should prioritize durability. The key questions are whether improvements survive turnover, political change, organizational pressure, and distribution shift, whether they reduce consequential harm for affected populations, and whether transparent audit mechanisms allow independent correction when practices fail.
Conclusion
Stereotypes, disparities, and legal decision-making form a coherent research object because legal decisions are sequential, socially situated judgments made under uneven information and institutional constraints. Experimental psychology identifies processes through which categories affect attention, threat appraisal, and group deliberation. Empirical legal research demonstrates how police exposure, charging, pretrial release, and sentencing can produce unequal outcomes. Quasi-experimental studies strengthen causal inference about selected decision margins. Algorithmic research shows why predictive calibration, equal error rates, decision thresholds, and procedural accountability cannot be collapsed into a single fairness score.
The synthesis establishes a methodological requirement: every claimed disparity must name its decision point, denominator, comparison population, available evidence, inferential design, and counterfactual question. Psychological theories explain potential mechanisms; they acquire legal relevance through verified effects on decisions and institutions. Policy responses can then be evaluated against clearly defined outcomes rather than the mere presence of awareness training or a supposedly neutral model.
A durable science of fair legal decision-making will connect human categorization with institutional histories, causal identification, transparent measurement, and accountable procedures. Such a science treats human–artificial decision support as part of a legally governed institutional process whose effects must be observed, challenged, and independently validated.
References
- Arnold, D., Dobbie, W., & Hull, P. (2022). Measuring racial discrimination in bail decisions. American Economic Review, 112*(9), 2992–3038. https://doi.org/10.1257/aer.20201653
- Arnold, D., Dobbie, W., & Hull, P. (2025). Building nondiscriminatory algorithms in selected data. American Economic Review: Insights, 7*(2), 231–249. https://doi.org/10.1257/aeri.20240249
- Arnold, D., Dobbie, W., & Yang, C. S. (2018). Racial bias in bail decisions. The Quarterly Journal of Economics, 133*(4), 1885–1932. https://doi.org/10.1093/qje/qjy012
- Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data, 5*(2), 153–163. https://doi.org/10.1089/big.2016.0047
- Cohen, A., & Yang, C. S. (2019). Judicial politics and sentencing decisions. American Economic Journal: Economic Policy, 11*(1), 160–191. https://doi.org/10.1257/pol.20170329
- Correll, J., Park, B., Judd, C. M., & Wittenbrink, B. (2002). The police officer's dilemma: Using ethnicity to disambiguate potentially threatening individuals. Journal of Personality and Social Psychology, 83*(6), 1314–1329. https://doi.org/10.1037/0022-3514.83.6.1314
- Crow, M. S., & Goulette, N. (2022). Judicial diversity and sentencing disparity across U.S. District Courts. Journal of Criminal Justice, 82*, Article 101973. https://doi.org/10.1016/j.jcrimjus.2022.101973
- Dressel, J., & Farid, H. (2018). The accuracy, fairness, and limits of predicting recidivism. Science Advances, 4*(1), eaao5580. https://doi.org/10.1126/sciadv.aao5580
- Eberhardt, J. L., Davies, P. G., Purdie-Vaughns, V. J., & Johnson, S. L. (2006). Looking deathworthy: Perceived stereotypicality of Black defendants predicts capital-sentencing outcomes. Psychological Science, 17*(5), 383–386. https://doi.org/10.1111/j.1467-9280.2006.01716.x
- Eberhardt, J. L., Goff, P. A., Purdie, V. J., & Davies, P. G. (2004). Seeing Black: Race, crime, and visual processing. Journal of Personality and Social Psychology, 87*(6), 876–893. https://doi.org/10.1037/0022-3514.87.6.876
- Forscher, P. S., Lai, C. K., Axt, J. R., Ebersole, C. R., Herman, M., Devine, P. G., & Nosek, B. A. (2019). A meta-analysis of procedures to change implicit measures. Journal of Personality and Social Psychology, 117*(3), 522–559. https://doi.org/10.1037/pspa0000160
- Greenwald, A. G., McGhee, D. E., & Schwartz, J. L. K. (1998). Measuring individual differences in implicit cognition: The implicit association test. Journal of Personality and Social Psychology, 74*(6), 1464–1480. https://doi.org/10.1037/0022-3514.74.6.1464
- Kang, J., Bennett, M. W., Carbado, D. W., Casey, P., Dasgupta, N., Faigman, D. L., Godsil, R. D., Greenwald, A. G., Levinson, J. D., & Mnookin, J. (2012). Implicit bias in the courtroom. UCLA Law Review, 59*, 1124–1186. https://escholarship.org/uc/item/3sd1m815
- Kleinberg, J., Lakkaraju, H., Leskovec, J., Ludwig, J., & Mullainathan, S. (2018). Human decisions and machine predictions. The Quarterly Journal of Economics, 133*(1), 237–293. https://doi.org/10.1093/qje/qjx032
- Kurdi, B., Seitchik, A. E., Axt, J. R., Carroll, T. J., Karapetyan, A., Kaushik, N., Tomezsko, D., Greenwald, A. G., & Banaji, M. R. (2019). Relationship between the Implicit Association Test and intergroup behavior: A meta-analysis. American Psychologist, 74*(5), 569–586. https://doi.org/10.1037/amp0000364
- Lawson, S. G., Narkewicz, E. L., & Vincent, G. M. (2024). Disparate impact of risk assessment instruments: A systematic review. Law and Human Behavior, 48*(5–6), 427–440. https://doi.org/10.1037/lhb0000582
- Mitchell, T. L., Haw, R. M., Pfeifer, J. E., & Meissner, C. A. (2005). Racial bias in mock juror decision-making: A meta-analytic review of defendant treatment. Law and Human Behavior, 29*(6), 621–637. https://doi.org/10.1007/s10979-005-8122-9
- Pager, D. (2003). The mark of a criminal record. American Journal of Sociology, 108*(5), 937–975. https://doi.org/10.1086/374403
- Payne, B. K., Vuletich, H. A., & Lundberg, K. B. (2017). The bias of crowds: How implicit bias bridges personal and systemic prejudice. Psychological Inquiry, 28*(4), 233–248. https://doi.org/10.1080/1047840X.2017.1335568
- Pierson, E., Simoiu, C., Overgoor, J., Corbett-Davies, S., Jenson, D., Shoemaker, A., Ramachandran, V., Barghouty, P., Phillips, C., Shroff, R., & Goel, S. (2020). A large-scale analysis of racial disparities in police stops across the United States. Nature Human Behaviour, 4*, 736–745. https://doi.org/10.1038/s41562-020-0858-1
- Rachlinski, J. J., Johnson, S. L., Wistrich, A. J., & Guthrie, C. (2009). Does unconscious racial bias affect trial judges? Notre Dame Law Review, 84*(3), 1195–1246. https://scholarship.law.cornell.edu/facpub/786/
- Rehavi, M. M., & Starr, S. B. (2014). Racial disparity in federal criminal sentences. Journal of Political Economy, 122*(6), 1320–1354. https://doi.org/10.1086/677255
- Sommers, S. R. (2006). On racial diversity and group decision making: Identifying multiple effects of racial composition on jury deliberations. Journal of Personality and Social Psychology, 90*(4), 597–612. https://doi.org/10.1037/0022-3514.90.4.597
- Spohn, C., Mitchell, O., Oramas Mora, D., Moreira de Andrade, T., Fine, A., Montes, A., & White, M. (2026). Inequality in sentencing: A systematic review of the effects of race and ethnicity on sentence outcomes. Journal of Criminal Justice, 105*, Article 102706. https://doi.org/10.1016/j.jcrimjus.2026.102706
- Tyler, T. R. (2006). Why people obey the law*. Princeton University Press. https://doi.org/10.2307/j.ctv1j66769
- Wilson, J. P., Hugenberg, K., & Rule, N. O. (2017). Racial bias in judgments of physical size and formidability: From size to threat. Journal of Personality and Social Psychology, 113*(1), 59–80. https://doi.org/10.1037/pspi0000092