Algorithmic Risk Assessment in Criminal Justice: Accuracy, Bias, and Fairness
Author: Ukrainian Psychological Hub · Published: October 11, 2026 · Editorial Policy
Algorithmic risk assessment in criminal justice refers to statistical or computational methods that estimate the likelihood of a specified future justice-system outcome using information available at an earlier decision point. The outcome may be a new arrest, a new conviction, a failure to appear in court, an alleged violation of supervision, or a narrower form of reoffending. The result may be expressed as a numerical score, a probability, or a category such as low, medium, or high risk. The defining question is whether a model is accurate, fair, clinically and legally appropriate for its actual use, and more beneficial than realistic alternatives.
The central scientific finding is that a risk algorithm can rank people with some predictive success while producing poorly calibrated probabilities, unequal error rates, or harmful decisions. A systematic review of commonly used sentencing instruments found generally poor-to-moderate external predictive performance and major gaps in calibration and error reporting. The review of 36 validation studies provides a more rigorous basis for evaluating these tools than claims based on a single accuracy percentage.
What is algorithmic risk assessment in criminal justice?
An algorithmic risk assessment combines information about people and their legal or institutional circumstances according to an explicit calculation. A traditional actuarial instrument may add weighted points for a small number of factors. A contemporary machine-learning model may estimate complex statistical relationships using a larger training dataset. Both are algorithms; artificial intelligence is not required. The important distinctions are the target outcome, data, model, threshold, validation population, decision context, and human use.
Criminal justice risk assessment has several separate purposes. In pretrial proceedings, a tool may estimate nonappearance or new-arrest outcomes during release. In sentencing, it may be introduced as contextual information, subject to relevant law. In corrections, it may help allocate supervision, programming, or reentry resources. In probation or parole, it may inform intensity of monitoring. Prediction of violence, general rearrest, technical violations, and court nonappearance are different targets; evidence supporting one does not establish accuracy for another.
A risk score is a statistical assessment result. It is not a diagnosis, a finding that a person committed an offense, an explanation of why an offense occurred, a measure of moral character, or a judicial finding of dangerousness. A legal decision requires independent application of applicable laws and procedural protections. For the broader forensic context, see Violence Risk Assessment: What It Can Predict and Where It Fails.
What do the algorithms actually predict?
The first evaluation question is the exact outcome definition. Rearrest, reconviction, self-reported behavior, a probation violation, a police contact, and failure to appear have distinct meanings. Rearrest measures a justice-system event shaped by enforcement, detection, reporting, and charging practices. An arrest may never lead to a conviction; a person who commits an offense may never be arrested. Calling rearrest a direct measure of underlying criminal behavior creates a measurement error before any model is trained.
Time horizons also matter. A predicted probability of a new arrest within six months cannot be directly compared with a five-year reconviction risk. A model developed for people released from prison may perform differently when applied to people awaiting trial, youth, women, people in another state, or people receiving different supervision. Changes in police practices, court reminders, housing support, or release policy can also alter measured outcomes without any underlying change in personal propensity.
Researchers therefore distinguish the prediction target from the legal question. A pretrial court may have to consider flight or public-safety concerns under its jurisdiction's statutes. A model trained to predict any rearrest may primarily recognize low-level enforcement outcomes, making it poorly aligned with a narrow serious-violence question. A risk score is relevant only to the extent its target, time horizon, population, and documented predictive properties match the authorized decision.
How are criminal justice risk scores calculated?
Models use combinations of historical and sometimes changeable variables, such as prior convictions, age at earlier legal contact, past supervision outcomes, employment or housing circumstances, and information from structured interviews. Instruments differ markedly. Prior convictions reflect both behavior and the legal system's past response. Socioeconomic and neighborhood information may reflect structural circumstances. The inclusion, definition, accuracy, and weighting of each variable require independent scrutiny.
In an additive score, the assigned points for each response are summed and compared with predefined ranges. In logistic regression, weighted input variables are transformed into an estimated outcome probability. More complex techniques include random forests and gradient-boosting models. These methods may capture nonlinear patterns, but added complexity is useful only if it improves independent performance, transportability, and decision outcomes without unacceptable tradeoffs.
An accurate algorithm requires an adequately defined outcome, representative development data, safeguards against data leakage, reliable measurement, sensible handling of missing data, temporal separation between prediction and outcome, and truly independent evaluation. Randomly splitting one historical dataset into training and test samples is weaker evidence of deployment readiness than evaluating a locked model on later cases from a different court or jurisdiction.
The model may output a probability or only a relative score. A label such as high risk has no universal percentage meaning: the threshold is chosen by developers or institutions. Decisions about thresholds determine how many people are flagged, how many are missed, and how much coercive burden is accepted. These are policy and legal judgments informed by statistics, rather than answers supplied by the algorithm itself.
How accurate are these tools? What the evidence shows
Accuracy is multidimensional. In a 2022 systematic review covering 11 common sentencing instruments, 36 studies and 597,665 participants were identified; 27 studies involving 177,711 participants were independent validations. Among independent studies with more than 500 participants, area-under-the-curve values generally ranged from 0.57 to 0.75. Studies frequently omitted calibration and false-positive/false-negative analyses, and methodological bias was common. These findings refer to the reviewed tools and settings, not to a universal performance ceiling for every algorithm (Fazel and colleagues, 2022).
A broader critical meta-review of criminogenic risk assessment examined 39 meta-analyses and systematic reviews and identified recurrent overinterpretation: confusing legally recorded outcomes with underlying offending, prediction with causal explanation, and predictive association with evidence that a particular intervention works. These distinctions are especially important when a numerical score appears scientifically precise.
An influential experimental comparison by Dressel and Farid (2018) found that COMPAS predictions in the dataset they examined were comparable to judgments of nonexpert participants, while a simpler model with two features achieved similar predictive performance. This study raised an important question about the added value of complexity. It does not demonstrate that every algorithm performs like a novice, nor does it constitute an evaluation of every version or deployment of COMPAS.
The correct comparative question is whether a specific tool, at a particular decision point, outperforms an appropriately defined baseline: an existing structured assessment, a simple statistical model, a judicial decision without the tool, or a less intrusive administrative policy. Evaluation must include the consequences of actual decisions, not merely a retrospective prediction contest.
Discrimination: AUC and ranking
Discrimination is the ability to place people with a recorded subsequent outcome above people without that outcome. The area under the receiver operating characteristic curve, or AUC, is the probability that a randomly selected person with the observed outcome receives a higher score than a randomly selected person without it, with ties handled conventionally. An AUC of 0.70 means this ranking succeeds in roughly 70 percent of such pairs; it does not mean an individual has a 70 percent probability of reoffending or that 70 percent of all individual predictions are correct.
AUC summarizes ranking over many possible thresholds. It does not reveal whether risk percentages are accurate, whether errors are concentrated in one group, whether a selected threshold is appropriate, or whether the tool improves real-world legal outcomes. The specific uncertainty interval and test-population characteristics also matter. AUC should be reported with calibration, confusion matrices, subgroup statistics, and useful alternatives.
Calibration: Are stated probabilities believable?
Calibration asks whether predicted probabilities correspond to observed outcome proportions. If people assigned a predicted risk of 20 percent experience the specified recorded outcome at approximately that frequency over the stated period, the score is calibrated around that range in that population. A model can discriminate reasonably well while systematically overestimating absolute risk.
Overall calibration is insufficient when groups have different calibration curves. Calibration should be examined by score band and relevant subgroups, with sample sizes and uncertainty. Predictive risk may change over time because of policy, population composition, or intervention. A score developed in one environment is not automatically calibrated in another.
The 2024 National Institute of Justice review of PATTERN, used in federal corrections under the First Step Act, demonstrates why ongoing evaluation matters. The report documented predictive analyses across racial and ethnic groups while continuing to identify differential prediction and overprediction relative to White individuals in particular subgroup comparisons. Federal revalidation is a governance process, not a claim that the same performance extends to pretrial courts or other instruments.
Sensitivity, specificity, false positives, and false negatives
Once a threshold is selected, a high-risk classification can be compared against the subsequently recorded outcome. A true positive is someone flagged who later has that outcome; a false positive is someone flagged who does not. A false negative is someone not flagged who later has the outcome; a true negative is someone not flagged who does not. These terms concern statistical classification against an observed label, not a verdict of innocence or guilt.
Sensitivity is the proportion of observed-outcome cases correctly flagged. Specificity is the proportion of observed non-outcome cases correctly not flagged. The false-positive rate is one minus specificity; the false-negative rate is one minus sensitivity. Positive predictive value is the proportion of flagged people who actually experience the recorded outcome during follow-up. These quantities answer different questions and should never be substituted for one another.
Consider an entirely hypothetical group of 1,000 people in which 100 experience the defined outcome. A classifier with 70 percent sensitivity and 80 percent specificity would correctly flag 70 people, miss 30, falsely flag 180 of the 900 people without an outcome, and correctly not flag 720. Of 250 people labeled high risk, only 70 would have the event: a positive predictive value of 28 percent. The same sensitivity and specificity generate a different positive predictive value when the underlying outcome rate changes. The illustration is arithmetic, not a reported result for any named instrument.
Rare serious outcomes pose special difficulties. Even a relatively discriminating model can produce large numbers of false alarms when base rates are low. Any claim that a system accurately predicts serious violence must include the event definition, follow-up, frequency, threshold, uncertainty, and practical consequences of false flags.
What does algorithmic bias mean?
Algorithmic bias can arise at multiple points. Historical records may reflect unequal exposure to policing. Labels may measure arrests rather than underlying conduct. Predictors can function as proxies for race, income, disability, geographic location, or other social circumstances. Underrepresentation during development can lead to poor subgroup performance. Scoring errors or institutional overrides can introduce additional disparities during use.
Racial bias is therefore a question about data generation, statistical performance, and decision consequences. The absence of an explicit race field does not guarantee racial neutrality. Other variables can carry correlated information, while records of prior justice-system contact may themselves contain structural differences. The National Institute of Justice's 2024 discussion of assessment best practices specifically warns against treating removal of race or proxy variables as a complete solution.
The term bias also has a narrower statistical meaning, such as differential prediction between groups. Researchers may find comparable AUC but different calibration, or similar calibration but different error rates. A legal or ethical claim of unfairness additionally requires context about which people are harmed, what rights are implicated, which comparator is relevant, and what a jurisdiction permits.
Fairness metrics: Why one definition does not settle the issue
Several mathematical criteria are used in criminal justice audits. Demographic parity asks whether different groups receive positive classifications at equal rates. Equal opportunity focuses on equal true-positive rates. Equalized odds calls for both true-positive and false-positive rates to be equal across groups. Predictive parity asks whether the positive predictive value of being flagged is equal. Calibration asks whether a given risk estimate corresponds to similar observed outcome rates across groups.
These criteria embody different choices about what should be equal. When groups have different observed base rates and prediction is imperfect, mathematical incompatibilities can arise. Chouldechova (2017) and Kleinberg, Mullainathan, and Raghavan (2017) formalized important tradeoffs between calibration-related and error-balance criteria. Their results are conditional mathematical statements; they do not imply that fairness work is futile or that existing disparities should be accepted.
Selecting one metric is an institutional value judgment. Equalizing false-positive rates matters when being falsely categorized high risk creates detention or burdensome supervision. Calibration matters when decision-makers interpret scores as probabilities. Equalizing release rates may matter under another stated policy objective. An audit should disclose the chosen criteria and analyze what happens to people under alternative thresholds and workflows.
A model can be statistically fair under one definition while enabling unfair treatment. Conversely, two groups may receive different average scores due to differences in measured outcomes even if one particular predictive fairness criterion is met. Fairness analysis must therefore examine procedural rights, accessible explanations, error correction, proportionality, and the effects of the entire decision system.
The COMPAS controversy and its actual lesson
COMPAS became widely known because investigators and researchers examined disparities in recidivism classifications in one county dataset. Subsequent academic work showed how apparently conflicting assessments of fairness can arise from different mathematical metrics. The debate is useful because it illustrates the difference between ranking performance, calibration, false-positive rates, and policy choices. It does not establish that a single analysis can speak for every version, population, use, or time horizon of a proprietary tool.
The Dressel–Farid study focused on a particular public dataset and a specific two-year rearrest outcome. Chouldechova's formal analysis addressed fairness definitions more generally. The broader audit question is whether vendors and agencies supply sufficient information to allow independent evaluation and meaningful challenge of a score.
The dedicated COMPAS article examines instrument-specific history and controversy. This article concentrates on cross-tool scientific and decision-making principles that also apply to public-sector tools and newer machine-learning systems.
Pretrial decisions, sentencing, and corrections require different safeguards
Pretrial assessment operates before a criminal conviction in the current case. Legal systems differ in the permissible reasons for detention, the procedural standard, and the weight attached to court attendance and public safety. The presumption of innocence and applicable procedural rights make the consequences of a false-positive classification particularly significant. A tool predicting any new arrest cannot automatically justify detention for a different statutory question.
Sentencing occurs under a distinct legal framework. The relevant evidence, permitted factors, limitations on individualized punishment, and appellate review vary by jurisdiction. The fact that a score predicts a recorded group outcome does not establish the punishment an individual deserves. Algorithms do not establish criminal responsibility, determine culpable mental state, or replace the court's legal analysis.
Post-sentencing systems often use risk-and-needs assessments to guide case planning and access to rehabilitative resources. A changeable need may be relevant to supportive intervention, whereas a historical predictor may improve ranking without being a treatment target. The NIJ Guidelines for Post-Sentencing Risk Assessment emphasize fairness, effectiveness, efficiency, and communication in the design and application of these tools. Here the question is whether assessment leads to better proportional, individualized services and outcomes.
For a wider explanation of recidivism and its predictors, see Recidivism: What Predicts Reoffending?. Predictive scores are one methodological component of that field; the development of desistance, rehabilitation, and broader social mechanisms is a separate scientific question.
Do risk tools improve real court decisions?
Improvement in prediction does not necessarily produce improvement in policy. Decision-makers may ignore valid recommendations, overrely on weak scores, use thresholds inconsistently, or add a restrictive condition to everyone labeled above a cutoff. A system may reduce detention in one jurisdiction and intensify supervision in another. Real-world evaluation must examine the complete workflow, including who receives an assessment, how judges interpret it, whether staff can override it, and what happens afterward.
A multi-site quasi-experimental study reported by NIJ compared assessed defendants with matched groups processed under prior or parallel practice. Risk assessment was associated with more nonfinancial pretrial release; structured guidelines were linked with greater release and less detention, alongside slightly higher rates of some nonviolent rearrest outcomes and no increase in failure to appear. These are results from the studied sites and design, not guaranteed causal effects in every court.
Another multi-site investigation documented differences in predictive validity for Black and White defendants while analyzing changes in release decisions. The combination illustrates why an algorithm can show subgroup limitations and still affect institutional outcomes differently than practices it replaces. Fairness of the score, fairness of judicial use, and fairness of resulting decisions must be audited separately.
Human judgment, automation bias, and accountability
Putting a human in the loop does not itself guarantee meaningful oversight. A judge may defer to a numerical score because it seems scientific; another may selectively disregard recommendations. An agency may have no usable explanation for particular outputs. Formal human review is effective only when reviewers understand what the score measures, can see relevant inputs, have authority to question the result, and document how it affected the legal decision.
A professional risk opinion is different from the legal judgment ultimately reached. The professional can interpret measurement limits and state the evidentiary basis. The court or authorized decision-maker applies the applicable legal standard. Clear separation of roles improves transparency and protects against turning a statistical indicator into a presumption of culpability.
NIST's Artificial Intelligence Risk Management Framework identifies validity, reliability, accountability, transparency, interpretability, privacy, and management of harmful bias as relevant to trustworthy AI. This is a voluntary general framework rather than a criminal-procedure statute or a certification that a particular risk model is lawful.
Explainability, proprietary models, and legal contestability
An explanation can describe the population used to train a model, each variable, overall feature influence, calibration, the meaning of a person's score, and available alternatives. These are not interchangeable forms of transparency. A readable explanation of a prediction does not prove it is accurate, while access to source code alone does not establish that a court's use is appropriate.
Proprietary secrecy creates additional governance problems when affected people cannot inspect key input records or identify an erroneous assessment. Corrections procedures should make relevant factual data accessible for challenge where required by law and establish how disputes are reviewed. Agencies should document model versions, vendor changes, recalibration, and the identity of people authorized to override assessments.
The due-process implications of a score depend on the jurisdiction, procedural setting, governing statutes, judicial decisions, and how heavily the score affects liberty. No single U.S. state ruling should be generalized to all states or countries. Scientific validity and legal admissibility are questions that intersect but are answered through different standards.
Population differences, cultural fairness, and model drift
A tool can lose accuracy when transported from adults to adolescents, from a prison-release cohort to pretrial defendants, across states or countries, or across language and cultural contexts. Variables can have different social meanings and different distributions. Binary sex-based models may provide incomplete information for some people; populations with low representation in the validation sample produce imprecise subgroup estimates.
Subgroup assessment should analyze outcomes, AUC, calibration, error rates, sample sizes, and uncertainty. Race, ethnicity, age, sex, disability, geography, and other dimensions may be important when ethically and legally appropriate. Merely dividing a small sample into many groups can create noisy estimates; the solution is better representative data and cautious interpretation, not assuming fairness where estimates are unavailable.
Drift occurs when a model's relationship with outcomes changes. The introduction of new court-notification systems, changes in supervision policy, altered policing intensity, or a new eligibility rule can change the observed outcome base rate. Periodic independent revalidation, clearly defined review triggers, and retirement of underperforming models are elements of responsible lifecycle management.
Is machine learning better than traditional actuarial assessment?
There is no universal winner. Conventional regression and point-based models can be transparent and competitive when the predictors and outcome are simple. Flexible machine-learning models can represent additional relationships but also overfit datasets or amplify measurement problems. A new model should be compared with simple benchmarks on the same people, horizon, outcome, and evaluation criteria.
A systematic review of machine-learning methods for recidivism identified studies with promising reported performance, while emphasizing the significance of transparent methods and reliable data. Results from selected machine-learning studies cannot be directly compared with the independent-sentencing validation review because their datasets, selection rules, and performance estimates differ. A higher reported accuracy in one paper does not establish superior performance in the same operational environment.
Model selection should prioritize demonstrated external calibration, decision-relevant accuracy, explainability adequate for the specific use, protection against harmful disparities, auditability, and continuous governance. A needlessly complex model offers little public benefit if a much simpler model performs similarly.
A minimum scientific and policy audit before deployment
A responsible audit begins by writing down the exact legal decision and prediction outcome. It then documents population eligibility, data provenance, missingness, subgroup representation, training/evaluation separation, baseline comparator, the full confusion matrix at contemplated thresholds, AUC, calibration curves, predicted versus observed risk, confidence intervals, and the meaning of each risk category.
The next stage evaluates who experiences false-positive and false-negative consequences, the availability of less restrictive alternatives, individual data correction, override practices, appeal routes, and downstream outcomes. Agencies should publish clear model documentation, allow independent replication to the extent compatible with privacy and legitimate security requirements, assign responsibility for monitoring, and provide a process for withdrawal when harms outweigh benefits.
The TRIPOD+AI reporting standard was developed for clinical prediction-model research, not specifically for criminal justice. Its emphasis on transparent reporting, evaluation populations, subgroup analysis, and reproducibility provides useful methodological guidance by analogy; it should not be presented as a binding forensic or legal standard. The NIJ post-sentencing guidelines address criminal-justice implementation more directly.
No high overall AUC can substitute for outcome definition, calibration, subgroup evaluation, policy impact analysis, and legal legitimacy. The decision whether to use a model should remain open to the possibility that an alternative with fewer liberty restrictions or lower administrative burden performs better.
Risks, ethical concerns, and practical implications
Potential benefits include more consistent scoring, explicit consideration of relevant factors, structured access to support services, and opportunities to measure variation in human decisions. Potential harms include unjustified detention, unequal supervision burdens, error propagation from inaccurate records, reduced procedural participation, stigmatizing labels, and diversion of resources away from effective interventions. Both possibilities must be demonstrated in context.
The most important ethical distinction concerns prediction and responsibility. A model may associate certain prior circumstances with recorded subsequent events without revealing the cause of those events. Statistical association cannot prove that a person will commit a crime, that a mental health diagnosis explains criminal behavior, or that someone deserves a more severe sanction. A risk factor is not an element of an offense.
Psychological and psychiatric diagnoses require separate clinical assessments and do not equate to criminality or future violence. A disability-related characteristic, communication difference, or life circumstance must never be treated as an automatic proxy for dangerousness. If assessment requires clinical expertise, qualified practitioners should interpret the evidence within the limits of the specific referral question.
For court professionals, the practical question is what decision the tool changes and at what cost. For journalists and researchers, it is whether numerical claims have meaningful denominators, thresholds, intervals, and subgroup comparisons. For individuals affected by an assessment, the relevant inquiries concern the recorded information, how the score was used, applicable contestation procedures, and legal assistance suited to the jurisdiction.
Common myths and misconceptions
A high-risk label means the person will commit a crime. In fact, risk categories organize uncertainty across a group; many flagged people will have no observed outcome.
A score with 70 percent AUC is 70 percent accurate about an individual. AUC measures pairwise ranking performance and cannot be read as an individual probability or categorical accuracy.
Removing race from the inputs removes racial bias. Historical patterns can remain in proxies, outcome measurement, and institutional use.
A model with equal average scores or equal calibration is fair in every sense. Distinct fairness criteria can disagree, and downstream legal burdens are separate from score properties.
Machine learning always improves on traditional assessment. Performance depends on target, dataset, validation, population, and decision context; additional complexity offers no inherent guarantee.
An algorithm makes sentencing objective. Decision objectives, thresholds, permitted factors, evidentiary requirements, and legitimate consequences remain normative and legal choices.
Frequently asked questions
What is an algorithmic risk assessment tool in criminal justice?
It is a rule-based statistical instrument or machine-learning system that uses recorded information to estimate a specified future justice-system outcome for a defined population and time horizon. It may aid a decision, but it does not establish guilt or replace legal standards.
Are criminal justice risk algorithms accurate?
Independent validation suggests variable and often modest discrimination for commonly used sentencing tools. Accuracy claims require AUC, calibration, false-positive and false-negative rates, subgroup results, confidence intervals, and evaluation in the intended setting. A score can rank reasonably well while misrepresenting absolute probabilities.
Why can an apparently accurate model produce many false positives?
The frequency of the target outcome matters. When an outcome is relatively uncommon, even an instrument with reasonable sensitivity and specificity can flag many people who do not later experience the outcome. Positive predictive value depends on the base rate as well as the model and threshold.
Are COMPAS and PATTERN the same system?
No. They are distinct instruments associated with different development and deployment histories. Scientific claims about one tool, version, outcome, or correctional population should not be transferred automatically to the other.
Is COMPAS racially biased?
Published analyses of particular COMPAS data have identified differences in false-positive classifications, alongside disagreement about which fairness metric should control an overall judgment. The answer depends on the precise population, target, model version, threshold, and fairness definition. Dedicated COMPAS analysis should examine those particulars rather than reducing a complex dispute to a slogan.
Can a court use a risk score to decide who is guilty?
A predictive risk score is not evidence that the charged act occurred. Criminal responsibility and guilt are determined according to the applicable substantive and procedural law. How a jurisdiction regulates predictive information at other stages is a distinct legal question.
Can algorithms replace a forensic psychologist?
A model can provide one kind of statistical information. Forensic evaluation may require collateral records, direct examination when appropriate, differential assessment, response-style analysis, and communication of uncertainty. The legal decision remains with the authorized decision-maker.
What is the difference between predictive fairness and legal fairness?
Predictive fairness concerns statistical properties such as calibration and group error rates. Legal fairness concerns rights, permissible factors, procedures, nondiscrimination, individual explanation, and proportionate decision-making under the applicable law. Both are essential to evaluating deployment.
How often should criminal justice algorithms be revalidated?
There is no single universal interval. Review frequency depends on the instrument, volume of decisions, drift, changes in population or legal policy, identified errors, and applicable regulatory requirements. A change in model, threshold, target population, or institutional procedure can justify earlier revalidation.
What should an affected person ask about a risk assessment?
Useful questions include which tool and version was used, what outcome and time horizon it predicts, which records were entered, whether those records can be corrected, the meaning and limits of the score, what role the score played in the decision, and which jurisdiction-specific review or legal procedures apply.
Related Articles
References
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.
Collins, G. S., et al. (2024). TRIPOD+AI statement: Updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ, 385, e078378. https://doi.org/10.1136/bmj-2023-078378.
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.
Fazel, S., Burghart, M., Fanshawe, T., Gil, S. D., Monahan, J., & Yu, R. (2022). The predictive performance of criminal risk assessment tools used at sentencing: Systematic review of validation studies. Journal of Criminal Justice, 81, 101902. https://doi.org/10.1016/j.jcrimjus.2022.101902.
Kleinberg, J., Mullainathan, S., & Raghavan, M. (2017). Inherent trade-offs in the fair determination of risk scores. Innovations in Theoretical Computer Science Conference. https://doi.org/10.4230/LIPIcs.ITCS.2017.43.
Lowder, E. M., Diaz, C. L., Grommon, E., & Ray, B. R. (2021). Effects of pretrial risk assessments on release decisions and misconduct outcomes relative to practice as usual. Journal of Criminal Justice, 73. NIJ publication record.
Lowder, E. M., Diaz, C. L., Grommon, E., & Ray, B. R. (2022). Differential prediction and disparate impact of pretrial risk assessments in practice: A multi-site evaluation. Journal of Experimental Criminology. NIJ publication record.
National Institute of Justice. (2021). Guidelines for post-sentencing risk assessment. https://nij.ojp.gov/library/publications/guidelines-post-sentencing-risk-assessment.
National Institute of Justice. (2024). 2023 review and revalidation of the First Step Act risk assessment tool. https://nij.ojp.gov/library/publications/2023-review-and-revalidation-first-step-act-risk-assessment-tool.
National Institute of Justice. (2024). Best practices for improving the use of criminal justice risk assessments: Insights from NIJ's Recidivism Forecasting Challenge Winners Symposium. https://nij.ojp.gov/topics/articles/best-practices-improving-use-criminal-justice-risk-assessments.
National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). https://www.nist.gov/itl/ai-risk-management-framework.
Prins, S. J., & Reich, A. (2021). Criminogenic risk assessment: A meta-review and critical analysis. PubMed Central full text.
Travaini, G. V., Pacchioni, F., Bellumore, S., Bosia, M., & De Micco, F. (2022). Machine learning and criminal justice: A systematic review of advanced methodology for recidivism risk prediction. PubMed Central full text.