Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Violence and Recidivism Risk Prediction

Abstract. Violence and recidivism risk prediction estimates the likelihood of defined harmful outcomes over specified periods in identifiable populations. This fundamental research monograph critically synthesizes the history of forensic prediction, actuarial methods, structured professional judgment, unstructured comparators, computational models, psychological risk variables, and longitudinal updating. Evidence is examined through primary prediction studies, independently validated models, systematic reviews, meta-analyses, and methodological reporting standards. The analysis distinguishes official reconviction from actual violence, ranking from probability calibration, predictive association from causal explanation, and model performance from decision utility. Major evidence syntheses identify moderate and context-dependent discrimination, heterogeneous outcomes, limited calibration reporting, substantial risks of study bias, and unresolved subgroup fairness and deployment effects. The synthesis establishes a framework for evaluating prospective validity, base-rate-dependent predictive values, temporal transfer, ethical error costs, and human–artificial decision workflows. This work is a critical narrative review of published evidence, not a systematic review, new experiment, or individualized forensic assessment.

Keywords: violence risk assessment; recidivism prediction; forensic psychology; actuarial risk assessment; structured professional judgment; HCR-20; violent reoffending; recidivism; criminal justice; risk prediction models; discrimination; area under the ROC curve; calibration; positive predictive value; base rates; false positives; external validation; dynamic risk factors; risk–need–responsivity; algorithmic fairness; Artificial Era

The Research Object and the Prediction Problem

Violence and recidivism risk prediction is the scientific estimation of the probability, relative likelihood, or structured scenario of a defined harmful outcome within a specified population and observation interval. A valid prediction identifies an index population, the starting point of observation, an outcome that can be ascertained consistently, a time horizon, the information available when the prediction is made, and the decision for which the estimate is intended. This formulation connects psychological assessment to longitudinal statistics, behavioral science, forensic psychiatry, criminal justice, and the sociology of institutions. It also requires the researcher to distinguish interpersonal violence, recorded violent offending, general criminal reconviction, rearrest, technical supervision violations, institutional aggression, sexual recidivism, and mortality. Each outcome has a different measurement process and a different relationship to the behavior investigators wish to understand. Fazel et al. (2012) and Ogonah et al. (2023) illustrate how changes in outcome definitions alter what risk instruments appear to predict.

A prediction may be expressed as a continuous score, an estimated absolute probability, a categorical risk level, or a structured professional formulation describing plausible paths to future harm. The scientific validity of one form does not automatically establish the validity of another. A ranking score can distinguish higher from lower observed event rates without producing calibrated probabilities. A clinical formulation can identify modifiable circumstances without showing that its summary judgment improves discrimination in a new population. A categorical label can obscure substantial variation inside the category and can be treated by decision makers as an individual property rather than as a population-conditioned estimate. This monograph therefore evaluates the statistical prediction target, the psychological interpretation of measured predictors, and the institutional intervention decision as separate stages of a single evidence pathway. Fazel et al. (2022) documents persistent gaps in reporting the measures needed for this distinction.

The scope covers adult and, where directly relevant, developmental and justice-system risk assessment; tools used in forensic mental health, correctional release, sentencing, supervision, and related contexts; historical and current assessment approaches; and contemporary debates about algorithmic evaluation and fairness. The principal scientific questions concern validity, calibration, transportability, intervention relevance, equity, and consequences of errors. Targeted threats against identifiable people, generic criminal etiology, rehabilitation programs, sexual offending treatment, and software-governance systems form neighboring scientific objects, discussed when their methods inform the present analysis. This work is an integrative, critical narrative synthesis of original studies, systematic evidence reviews, and methodological guidelines. It reports no new individual-level dataset, preregistered review search, prospective validation, or meta-analysis.

Defining Violence, Offending, and Recidivism

Violence is a behavioral construct that requires an explicit specification of acts, intent, injury, coercion, or threatened force. Studies may define physical assault narrowly, include threats or property damage, or use legal categories determined by local law. Criminal recidivism is often defined administratively through a new arrest, charge, conviction, or imprisonment, each generated by multiple behavioral and institutional processes. An arrest measures the conjunction of suspected conduct and a policing or reporting response; a conviction additionally depends on investigation, charging, plea, evidentiary, and adjudicative processes. An observer-recorded assault in a hospital may therefore be constructively closer to physical aggression than a charge-based endpoint, even though the former may be less consistently documented. Ogonah et al. (2023) found that most included forensic-discharge studies relied on subsequent criminal events rather than independent behavioral measures.

Prediction also requires a time origin and an at-risk period. A six-month inpatient aggression study begins under different environmental constraints from a two-year post-release recidivism study. A person detained for part of follow-up has less community exposure; a person who moves jurisdiction may become invisible to a local administrative register. Death and long-term incarceration can act as competing events rather than ordinary outcome-free observations. Researchers should state whether estimates are for cumulative incidence over a fixed horizon, a time-varying hazard, or the probability of a recorded event conditional on observed follow-up. Fazel et al. (2016) demonstrates the need to anchor reoffending predictions at release, while Fazel et al. (2022) shows that validation evidence for sentencing decisions often does not document the performance dimensions needed for high-stakes implementation.

Outcome specificity governs model transfer. A predictor of general reconviction can owe much of its performance to common nonviolent offenses; its apparent predictive value does not establish competence to forecast severe interpersonal harm. The frequency of an outcome changes positive predictive value and intervention workload. Separate validation is necessary for rare but serious violence, for sexual reoffending, and for short-term ward incidents. The population's age, gender composition, prior criminal history, follow-up mechanism, and legal setting further affect observed event rates. A defensible research conclusion begins by identifying the actual recorded endpoint and preserves that specificity in every subsequent statement. Hanson & Morton-Bourgon (2009) examined sexual-offending prediction in a distinct evidence base, while Fazel et al. (2012) compared tools against violent, sexual, and broader criminal outcomes.

Historical Development of Forecasting Harm

The history of violence prediction combines several lines of inquiry: clinical judgment about dangerousness, the statistical study of criminal careers, institutional risk classification, forensic psychiatric assessment, and modern prediction modeling. Twentieth-century decision systems often assumed that expert knowledge of symptoms, biography, and observed conduct could support future-oriented judgments. The shift toward explicitly structured predictors required defining what would count as a successful prediction, obtaining follow-up data, and comparing decisions with a reproducible reference outcome. It was also influenced by the recognition that intuitive judgments can be vulnerable to selective recall and base-rate neglect. The MacArthur Study brought systematic longitudinal research and multiple predictors into the study of violence following psychiatric hospitalization. Its methodological program is documented in Monahan et al. (2001), which remains a primary historical source rather than a substitute for contemporary external validation.

A second tradition developed from actuarial criminology and correctional assessment. It emphasized empirically derived regularities across groups, with historical offending, age, and other measurable predictors commonly used to discriminate subsequent offending risk. The publication of various structured tools subsequently linked clinically salient information to transparent item sets and standardized procedures. These approaches reflect different explanatory commitments. The actuarial tradition emphasizes statistical estimation for comparable populations; structured professional judgment emphasizes case formulation, contextual contingencies, and management planning. Both can incorporate empirical evidence, and the comparison depends on whether the research evaluates a tool's score, its final clinician judgment, or the decisions subsequently made using its recommendations. Singh et al. (2011) evaluated multiple commonly used instruments; later Viljoen et al. (2025) examined studies directly comparing structured tools with unstructured judgments.

By the 2010s and 2020s, researchers increasingly assessed how models behave outside their construction samples and whether score-based decisions create unequal institutional consequences. Large-scale evidence synthesis found moderate performance in many settings and substantial deficits in calibration reporting. Computational debates sharpened distinctions among model complexity, discrimination, and group fairness. Dressel & Farid (2018) questioned the incremental prediction advantage of a prominent commercial algorithm relative to much simpler models and human judgments in the investigated data, while Chouldechova (2017) formally examined fairness criteria under group differences in outcome prevalence. The historical trajectory is one of increasingly explicit outcome definitions, statistical validation, and accountability for how evidence enters consequential decisions.

Psychological Constructs and Predictive Variables

A risk factor is a characteristic statistically associated with a later event in a defined design and population. Psychological predictors may include prior aggression, impulsive behavior, substance misuse patterns, emotional dysregulation, threats, destabilizing social conditions, and contemporaneous difficulties following treatment or supervision. Some are measured directly through behavioral records, whereas others rely on an interview, third-party information, or rating scale. The meaning of an apparently identical item can differ between a controlled hospital, a community supervision program, and a prison-release cohort. Stable historical predictors often provide useful baseline stratification, yet they do not establish a mechanism through which an intervention can prevent future harm. A strong model distinguishes temporal precedence and predictive association from causal explanation. Fazel et al. (2016) offers an example of empirically selected predictors within a defined release population.

The distinction between static and dynamic factors is operational. A historical conviction can remain fixed after assessment, while acute intoxication, current threats, housing instability, treatment engagement, and social conflict may change. Dynamic status alone does not establish that changing a score changes future violence: an item may vary with intervention, with measurement error, or with the same unmeasured circumstance that affects the outcome. A useful dynamic marker must demonstrate reliable assessment, temporal sensitivity, incremental prospective prediction, and ideally intervention-linked improvement in meaningful outcomes. The 2026 review of HCR-20 longitudinal properties Campbell et al. (2026) found evidence that dynamic scores can change over treatment, but little evidence that score changes themselves predict recidivism. This is a consequential difference between a measure of a current state and a validated surrogate endpoint.

Psychiatric symptoms, personality measures, social history, and environmental opportunities are connected in complex pathways. A diagnosis is not a synonym for violent behavior, and a risk factor should be interpreted within its population, operational definition, and adjustment strategy. The psychological task is to explain how states, capacities, habits, interpersonal situations, and institutions interact over time; prediction models may identify useful regularities without settling these causal questions. The most scientifically productive analyses keep three models apart: a statistical model estimating an outcome, a psychological account specifying processes, and a decision model explaining what action follows from the estimate. Ogonah et al. (2023) demonstrates the diversity of the populations and outcomes in forensic psychiatric evidence, and Fazel et al. (2022) establishes why tool choice requires decision-specific external validation.

Population, Setting, and Exposure as Core Conditions

The probability attached to a person's score is conditioned on the population from which a model's empirical relationships were learned. People convicted of different offense categories, patients discharged from secure hospitals, detained adolescents, and individuals under community supervision are exposed to different environments and institutional selection mechanisms. Their recorded outcome rates cannot be treated as transportable constants. A model developed in a national register with near-complete follow-up may operate differently in a local probation service with substantial missing data. Fazel et al. (2016) developed a scalable instrument from a national release context; independent evaluation is required before its coefficients or risk thresholds are used in a new setting.

Prediction models are especially sensitive to selection into the study population. A forensic clinic may assess only referrals judged already difficult, while a sentencing sample includes people selected through policing, charging, plea, and court processes. Conditioning on a referral can generate spurious associations if both predictor and outcome-related influences affect entry into the sample. Access to care and documentation quality can additionally vary across demographic groups. Sampling methods, exclusions, the denominator at risk, and follow-up duration therefore belong in the scientific description of any validation result. The large review of sentencing instruments Fazel et al. (2022) showed that study size and author involvement could relate to optimistic estimates, highlighting how evidence production itself can influence apparent performance.

Generalization across countries and periods also depends on criminal statutes, policing practices, sentencing regimes, treatment access, and the social meaning of measured items. Transportability should be assessed at the level of absolute risk, discrimination, calibration, and decision consequences separately. A score that retains similar rank ordering may systematically overestimate or underestimate local event rates. Temporal change in the underlying system can produce the same problem even without changing the model. An accountable assessment program treats each deployment setting as a new validation question and monitors performance over time rather than assuming validity from a familiar instrument name. Ogonah et al. (2023), Lawson et al. (2024), and Moons et al. (2025) provide complementary evidence and methodological criteria for this practice.

Actuarial Prediction and Statistical Scoring

Actuarial risk assessment assigns weights to indicators according to a predetermined rule, ordinarily derived from empirical associations with a specified outcome. In classical actuarial instruments, the rule is often a sum of items, weighted score, or reference-table mapping; in contemporary prediction modeling it can be a regression function or a more flexible classifier. Its most important scientific property is reproducibility: two assessors using the same documented input and algorithm should obtain the same numerical result, subject to measurement differences. The output requires independent validation before decision use. Hanson & Morton-Bourgon (2009) found an advantage for empirically derived actuarial measures over unstructured judgment in the particular sexual-offending studies it reviewed, while Fazel et al. (2022) demonstrated that risk assessment instruments used in sentencing vary widely in evidentiary strength.

A numerical score can be useful for ranking risk while still encoding strong historical dependencies. Criminal history may be a particularly predictive variable because it summarizes prior behavior, exposure to enforcement, and institutional responses. Its apparent precision should not obscure how it was recorded or which background conditions are embedded within it. Actuarial models often involve tradeoffs between a small, stable predictor set that can be applied consistently and a larger set requiring greater measurement effort. The latter need not offer better performance once tested outside the development dataset. Dressel & Farid (2018) is instructive because the evaluated recidivism algorithm did not clearly outperform a much simpler model under the study's comparisons. This conclusion concerns the investigated task and data, not all actuarial methods.

A legitimate actuarial probability is not merely a percentile or risk category. The conversion of a score into an absolute probability depends on how outcomes accumulated among individuals with similar scores, over a stated horizon, in a particular reference population. When a model is transported, this mapping may require updating. Local recalibration can change predicted event probabilities while leaving ranking nearly unchanged. Comparisons of instruments should therefore report both discrimination and calibration at prespecified horizons, along with error rates relevant to the contemplated decision. AUC alone does not establish that two differently labeled categories have a common practical meaning. Fazel et al. (2012) and Ogonah et al. (2023) show why the distinction matters in forensic populations.

Structured Professional Judgment and Case Formulation

Structured professional judgment (SPJ) provides a framework for collecting risk-relevant information, evaluating its meaning, considering future scenarios, and formulating management strategies. Instruments such as versions of the Historical Clinical Risk Management-20 are associated with this approach. Rather than treating a total score as the full clinical conclusion, SPJ asks the assessor to describe which circumstances could make an adverse outcome more or less likely and how those circumstances relate to a particular case. This design can help integrate contextual and dynamic evidence, but its final outcome is partially dependent on professional judgment. Consequently, the empirical performance of item totals cannot automatically be attributed to the validity of final summary judgments, risk formulations, or interventions. Ogonah et al. (2023) makes this distinction important by reviewing studies that reported different outputs from structured instruments.

An SPJ formulation is scientifically valuable when it is explicit enough to be checked. It should separate documented facts from interpretations, describe uncertainty in disputed information, name the scenario under consideration, and identify protective or aggravating circumstances. It should also explain what evidence would lead to revision. Reliability matters because two qualified assessors can agree that a historical event occurred while disagreeing about its clinical relevance. Predictive studies should distinguish interrater agreement in item coding, agreement in overall risk categories, and the validity of those categories against subsequent outcomes. A high reliability coefficient does not guarantee useful discrimination or calibration; a statistically associated category does not guarantee that a proposed treatment plan will reduce harm. Singh et al. (2011) and Ogonah et al. (2023) anchor this validation distinction.

The usual opposition between actuarial and SPJ assessment is too coarse for modern research. Both can include historical and contextual information, and both rely on measurement decisions made by humans. A structured instrument may support either score-based classification or individualized case formulation. The critical comparative questions are whether a particular workflow yields accurate, reproducible decisions and whether its management recommendations improve outcomes under controlled or otherwise credible evaluation. Viljoen et al. (2025) found that total scores from structured tools outperformed unstructured judgments on aggregate in direct-comparison studies; the finding does not establish that every SPJ final judgment, or each ensuing legal decision, is superior to every actuarial alternative.

Unstructured Clinical Judgment and Comparator Design

Unstructured judgment arises when an assessor integrates available evidence without a predefined reproducible rule for the final prediction. Such practice may capitalize on experience and context, yet its performance is difficult to audit because important inputs, weighting, and threshold choices may remain implicit. Comparisons with standardized instruments must be genuinely paired: the same population, same outcome, comparable observation horizon, and comparable access to information. If one method is evaluated using follow-up records unavailable when predictions were made, the comparison is contaminated by hindsight. Similarly, if a score receives more systematic predictor collection than its comparator, the estimated advantage may reflect the workflow as well as the mathematical rule. Direct comparisons are particularly informative because they reduce between-study differences that complicate historical narratives about professional intuition. Viljoen et al. (2025) follows this logic.

The preregistered synthesis reported by Viljoen et al. (2025) identified 31 direct-comparison studies encompassing 169 effect sizes from 45,673 risk judgments. Its multilevel meta-regression found higher predictive validity for total scores on structured instruments than for unstructured judgments in the examined violent, any, and sexual offending outcomes. The study was published online in 2024 in a 2025 issue and subsequently received a publication correction Correction (2026) involving redacted author references in the original manuscript. This provenance should accompany the citation. Aggregated effects support a general methodological preference for transparent, empirically assessed structure over intuition alone, while leaving open differences among populations, instruments, and practical uses.

An appropriate comparator is not always an unaided individual judgment. Institutions may currently use validated basic demographic rules, comprehensive case conferences, statutory eligibility criteria, or sequential screening. A new instrument must be compared with the actual decision pathway it would replace or augment. Measures of incremental performance should be interpreted against the simplest useful baseline, and evaluation must distinguish what assessors know before seeing a score from the changes caused by showing it. In many settings the most important outcome is whether structured assessment improves an actionable decision, rather than whether it produces a statistically better rank ordering. This transition from psychometric comparison to decision-effectiveness research is one of the field's central unfinished tasks. Fazel et al. (2022) documents the gap between reported statistical validity and sentencing application.

Machine Learning and Predictive Complexity

Machine learning can extend classical statistical prediction through flexible relationships, interactions, and high-dimensional data. Logistic regression, survival models, penalized models, trees, ensembles, and neural networks differ in complexity but share the requirement to predict future outcomes from information available at the decision point. A complex model can exploit noise in its development sample and fail when cases, institutions, or time periods change. Training, tuning, and final evaluation must therefore be separated, preferably with a genuinely external evaluation dataset. Transparent documentation must describe preprocessing, missing-data handling, outcome definition, hyperparameter tuning, threshold choice, and subgroup performance. Collins et al. (2024) established TRIPOD+AI reporting guidance across regression and machine-learning prediction studies, while Moons et al. (2025) updated methodological appraisal for both approaches.

More variables do not necessarily imply better prediction. Models trained on detailed institutional histories may learn proxies for surveillance intensity, access to services, neighborhood disadvantage, or a particular jurisdiction's recording behavior. Features recorded after a prediction date create data leakage and can falsely increase apparent validity. Imbalanced outcomes require special care because a model can achieve high overall accuracy by predicting the majority outcome without identifying enough true cases to justify its use. Evaluation should emphasize calibrated probabilities, meaningful operating thresholds, uncertainty, and performance relative to simple models. The comparisons in Dressel & Farid (2018) caution against assuming that proprietary algorithmic complexity itself delivers incremental accuracy.

Computational tools also change how decisions are made by distributing tasks between data systems, assessors, and authorities. An automated score may appear objective because it is consistent, while its training labels and deployment context remain socially constructed. These issues connect the present dossier with the distinct research object of algorithmic justice decisions. Within violence and recidivism prediction, the decisive questions remain empirical: for whom is this model validated, for what outcome and horizon, how is probability calibrated, and what happens when an individual score influences a legal or clinical action? Lawson et al. (2024) examines potential unequal institutional impacts, and Chouldechova (2017) formalizes a particular conflict among fairness criteria.

Prediction Time, Updating, and Dynamic Risk

Violence and recidivism risk unfold over time, making temporal design a constitutive feature of the prediction task. Baseline models estimate future outcomes from information collected at a fixed index date; time-updated models incorporate subsequent observations and must specify when new information becomes available. Repeated clinical assessments may be essential for case management, but treating a new assessment as though it had been available at baseline introduces look-ahead bias. Researchers need a longitudinal design that preserves time ordering and considers irregular observation, varying supervision intensity, and people who leave the cohort. Short-term incident prediction and long-term reconviction forecasting require different predictor windows and outcome verification procedures. The original prospective logic of prison-release prediction is illustrated by Fazel et al. (2016).

Changes in dynamic risk ratings can reflect treatment effects, natural recovery, changing environments, instrument familiarity, or regression toward average measurements. Establishing that a scale is responsive requires evidence of measurement reliability and sensitivity to genuine change. Establishing that change forecasts subsequent offending demands a stronger longitudinal design with adequate temporal separation and follow-up. The review by Campbell et al. (2026) examined longitudinal HCR-20 properties and reported very limited support for treating changes in dynamic scores as predictors of recidivism. That finding constrains claims that a declining composite risk score is itself evidence that a patient's future violence probability has declined by a quantified amount.

A practical update protocol should prespecify triggers for reassessment, distinguish long-term baseline factors from acute situational information, and document interventions occurring between assessments. Recalibration may be required after a service changes admissions, treatment, staffing, release criteria, or follow-up practices. Because interventions can alter the event-generating process, model evaluation within a treated cohort must explain how actions in response to earlier risk predictions affect later outcomes. A successful risk-management service may reduce observed violence among people initially rated high risk, thereby weakening an apparent score-outcome association without invalidating the original concern. This feedback between predictions and institutional response requires causal and decision-focused research rather than simple retrospective correlation. Moons et al. (2025) supplies a useful framework for separating model construction from evaluation.

Study Designs and Prospective Validation

A credible validation study evaluates predictions made before outcome ascertainment and defines its population in terms that permit reproducibility. Prospective cohorts offer a clear temporal design; retrospective cohorts can also support sound validation if predictor values reflect information available at the original index date and outcome ascertainment is reliable. The division between development and evaluation datasets is crucial. Fitting and testing on the same observations yields apparent performance, while resampling and cross-validation estimate internal optimism. External validation asks whether the fully specified model continues to perform in independently collected people, institutions, periods, or jurisdictions. Updating coefficients or risk thresholds using the external cohort changes the evaluative question and must be reported separately. Moons et al. (2025) distinguishes these phases in PROBAST+AI, and Collins et al. (2024) specifies their transparent reporting.

The most persuasive studies predefine outcomes and time horizons, determine an adequate sample and event count, document missing-data procedures, address censoring, and report the number of participants whose outcomes cannot be established. Outcome ascertainment should ideally be independent of the predictor scoring and, when feasible, blinded to it. Researchers should report the distribution of scores, population composition, attrition, criminal history, and relevant treatment or supervision exposure. Any threshold analysis must distinguish a cutoff chosen before evaluation from one optimized after seeing observed outcomes. Multiple exploratory cutoffs can create optimistic error estimates. Fazel et al. (2022) found that small studies and developer co-authorship were associated with higher reported performance, indicating why independent study design matters as much as nominal instrument popularity.

Evaluation must match the intended decision. For parole or release planning, a model might be assessed for fixed-horizon violent reconviction under community exposure. For psychiatric discharge, behaviorally ascertained aggression and risk management may be more informative than arrest alone. For sentencing, an instrument validated elsewhere cannot be presumed applicable to a different legally consequential population. AUC, calibration, and decisions at thresholds all require matching endpoints. The systematic review of forensic psychiatric patients Ogonah et al. (2023) illustrates that an evidence base can contain many studies and instruments while offering little usable calibration information, and the sentencing review Fazel et al. (2022) documents a similar problem in a different setting.

Discrimination and Receiver Operating Characteristics

Discrimination describes a model's capacity to assign relatively higher scores to people who later experience a defined outcome than to those who do not. A commonly reported measure, the area under the receiver operating characteristic curve (AUC), can be interpreted under standard conditions as the probability that a randomly selected event case receives a higher score than a randomly selected nonevent case, with conventions for tied scores. An AUC of 0.50 represents chance-level ordering; an AUC appreciably above this indicates some discrimination in the evaluated population. It does not say that a particular person has the AUC percentage probability of offending, and it does not specify which cutoff should trigger action. Ogonah et al. (2023) explains this distinction while synthesizing forensic-discharge studies.

The AUC is sensitive to case mix and heterogeneity. If two samples have very different ranges of historical offending or age, the same predictor model can exhibit different discrimination even when its conditional associations are unchanged. Confidence intervals, study sampling frames, follow-up time, and endpoint definitions should accompany an AUC. Comparing reported AUCs from different populations without controlling these differences can be misleading. Fazel et al. (2012) examined nine commonly used instruments across 73 samples, while the 2023 forensic review Ogonah et al. (2023) reported pooled AUCs of approximately 0.64 to 0.72 for selected tools in eligible validations. The differences were specific to tools and populations and should not be converted into a universal numerical ranking.

For fixed horizons, time-dependent ROC analysis may be appropriate when follow-up and censoring differ. Investigators should state whether the reported measure reflects violent, sexual, general, or institutional outcomes, and whether individuals were assessed under similar conditions. In very low-prevalence settings, precision-recall summaries and positive predictive values may reveal limitations obscured by an acceptable AUC. Discrimination establishes a limited predictive property, one among several needed to support a consequential decision. Fazel et al. (2022) found many sentencing studies reported an AUC without the accompanying information necessary to determine clinical or policy utility.

Calibration, Absolute Probability, and Transportability

Calibration asks whether predicted probabilities agree with observed event frequencies in the population and time window under evaluation. If a group of people is assigned an estimated two-year risk near twenty percent, a well-calibrated model should produce an observed event frequency near that estimate over the specified follow-up, allowing for sampling uncertainty and censoring. Calibration-in-the-large assesses average agreement, calibration slope assesses the spread of probabilities, and graphical methods can reveal departures across the risk range. A risk rank may remain informative even as its absolute probabilities become systematically too high or too low. This is why a model that has retained discrimination can still mislead a sentencing or release decision based on a fixed numerical threshold. Moons et al. (2025) places calibration among central prediction-evaluation concerns.

Calibration can drift when the baseline event rate changes, when the population's clinical profile differs, or when recording practices are altered. Interventions deployed in response to scores can also change observed outcomes. Updating an intercept or baseline hazard may correct some shifts, while changes in predictor effects may require more substantial model updating and revalidation. Any updating must be performed without using the final held-out performance sample twice. An external OxRec validation in England involving 1,770 people released from prison Beaudry et al. (2023) observed systematic underestimation before recalibration while maintaining a reported discrimination AUC of about 0.71 at one and two years; the result illustrates how stable discrimination can coexist with imperfect absolute-risk estimation. The general principle is supported by validation guidance in Moons et al. (2025) and by the original scalable release-model study Fazel et al. (2016).

Calibration is particularly consequential in subgroup analysis. Two groups can share overall AUC while differing in the accuracy of their predicted probabilities at clinically relevant thresholds. Apparent group differences may reflect baseline event rates, unmeasured exposures, missing data, variable ascertainment, or model bias. Reporting should include calibration curves or carefully chosen grouped summaries with intervals, not simply group average scores. The objective is to establish whether the score conveys the probability claimed for the people receiving it and whether its use produces defensible decisions under the relevant legal and ethical obligations. Lawson et al. (2024) examined questions of disparate institutional impact, and Chouldechova (2017) shows why error-rate comparisons and calibration cannot be reduced to a single fairness number.

Sensitivity, Specificity, Predictive Values, and Base Rates

At any prespecified threshold, the four possible classifications are true positives, false positives, true negatives, and false negatives relative to the selected endpoint and follow-up horizon. Sensitivity is the fraction of observed cases identified above threshold; specificity is the fraction of observed nonevents identified below it. Positive predictive value (PPV) is the fraction of positive classifications that are followed by the event, and negative predictive value (NPV) is the fraction of negative classifications not followed by the event. The same sensitivity and specificity can yield quite different PPV and NPV when the underlying event prevalence changes. These measures should be reported with the underlying counts and uncertainty, because proportions alone hide the sample and error burden. Fazel et al. (2012) and Ogonah et al. (2023) repeatedly expose the practical importance of these statistics.

Base-rate reasoning is a mathematical constraint, rather than an opinion about an individual's character. Suppose an entirely hypothetical cohort of 1,000 people has a ten-percent event rate, a threshold sensitivity of eighty percent, and specificity of eighty percent. Eighty of the 100 future event cases would be flagged, but 180 of the 900 nonevent cases would also be flagged. The positive predictive value would therefore be 80 divided by 260, approximately 31 percent, despite both sensitivity and specificity being 80 percent. This is an illustration derived arithmetically, not a claim about any named instrument or real population. Reducing event prevalence with unchanged sensitivity and specificity lowers PPV further. A policy that reads a positive flag as a near-certain forecast will therefore systematically exaggerate individual certainty.

False positive and false negative consequences depend on the action. A false positive leading to additional voluntary support differs ethically and legally from one leading to prolonged detention or a higher sentence. A false negative associated with the release of a person who later commits violence has distinct harms, which also depend on whether any reasonable preventive intervention was available. Threshold choice is a decision problem requiring explicit costs, legal constraints, and evidence that the intervention changes outcomes. Statistical maximization of a single accuracy measure does not settle the decision. Fazel et al. (2022) shows that sentencing literature often omits key false-positive and false-negative measures; Ogonah et al. (2023) similarly cautions against applying tools without decision-relevant validation.

Missing Data, Censoring, and Administrative Measurement

Missingness arises when assessment records are incomplete, cases leave a jurisdiction, follow-up is shortened, or outcomes occur outside the information system. A complete-case analysis can change the population and bias accuracy estimates when record availability is related to both predictors and outcomes. Simple imputation can likewise create unjustified certainty. Investigators should report how data became missing, examine patterns across relevant groups, use methods compatible with the modeling design, and test the sensitivity of results to plausible missingness mechanisms. When outcome follow-up differs substantially, survival analysis or competing-risk approaches may be preferable to naive fixed-window classification. Collins et al. (2024) requires transparent reporting of missing-data handling, and Moons et al. (2025) incorporates the issue into model evaluation.

Administrative outcomes can generate ascertainment bias because detection and recording are socially organized. Increased supervision may produce more technical violations without an equivalent rise in serious violence; policing and prosecution practices can make observed recidivism differ from actual conduct. Combining registers can improve completeness while introducing matching errors, duplicate records, and changing legal definitions. A validation study should specify which agencies contributed data, how outcomes were linked, whether index dates preceded all predictor information, and how competing events were treated. Fazel et al. (2016) illustrates the use of linked register information for scalable prediction, while Fazel et al. (2022) reviews the limitations of criminal risk prediction in legally consequential settings.

When models are deployed, they can affect the data subsequently used to retrain them. Higher scores may produce tighter monitoring, generating additional detectable incidents; targeted services may reduce subsequent events; incarceration changes community exposure. Such feedback can change the observed association between a score and its target. A purely retrospective model may then reproduce effects of prior policy rather than estimate the result of a specified alternative action. Careful design should separate baseline risk, exposure, intervention, and detection. The statistical report should preserve the institutional context rather than attributing every recorded event to an individual psychological disposition. Lawson et al. (2024) and Chouldechova (2017) offer complementary frameworks for examining how prediction and institutional processes connect.

The Meta-Analytic Evidence on Predictive Performance

Quantitative syntheses address a fundamental problem of the field: studies of the same nominal instrument often differ in case mix, follow-up, outcome, administration, and analytic methods. The early comparative analysis by Singh et al. (2011) included 68 studies and 25,980 participants, examining differences among several instruments and study characteristics. A later systematic review by Fazel et al. (2012) included 73 samples and approximately 24,800 participants. These works helped establish that common tools can discriminate subsequent violence or offending to a meaningful but imperfect degree. Their estimates should be understood as aggregated evidence from heterogeneous validation contexts, and differences among tools are less persuasive when the underlying sample characteristics and outcome definitions are not aligned.

The forensic psychiatric synthesis by Ogonah et al. (2023) included 50 publications, 36 tools, and 10,460 participants from 12 countries. The authors judged 49 of the 50 studies to have high risk of bias. Among independent external validations eligible for meta-analysis, pooled AUCs for selected violent outcomes were approximately 0.72 for the H10 component, 0.69 for HCR-20 version 2, 0.69 for the Violence Risk Appraisal Guide, and 0.64 for Static-99. These are discrimination estimates for selected populations, not individual probabilities; confidence intervals and the particular tool–endpoint combinations remain essential. The researchers also found that most studies lacked measures of calibration, sensitivity, and specificity. Crucially, the HCR-20 version 3 did not meet the review's criteria for pooled estimation, so the version 2 result cannot be casually substituted for a version 3 validation.

The sentencing-oriented systematic review by Fazel et al. (2022) identified 36 studies and 597,665 participants, including 27 independent validations with 177,711 individuals. Among larger independent samples, reported AUCs ranged from 0.57 to 0.75; performance was generally poor to moderate, with substantial missing calibration and error-rate information. The direct-comparison meta-analysis by Viljoen et al. (2025) provided a complementary question: structured scores performed better than unstructured judgment on average in eligible paired comparisons. Together these findings support using transparent structured methods as an evidential baseline while requiring independent, context-specific validation of each tool. None of the cited syntheses demonstrates that an instrument alone identifies which specific person will commit future violence or that a high score justifies an automatic deprivation of liberty.

Evidence From Forensic Psychiatric and Correctional Settings

Forensic mental health research involves clinical populations who may differ sharply from general community or prison-release cohorts. Hospital admission, symptom management, treatment engagement, capacity to leave the facility, and discharge rules influence both predictors and subsequent exposure. Instruments developed for clinical formulation can be tested as prognostic scores, but their use for discharge decisions requires evidence about postdischarge outcomes and real-world interventions. In Ogonah et al. (2023), recorded criminal outcomes dominated the evidence, whereas independent reports of aggressive behavior were rare. Thus even a carefully estimated performance statistic often describes recorded recidivism rather than the full incidence of interpersonal aggression. The difference matters when an instrument's stated clinical purpose extends beyond the recorded endpoint.

Correctional models face another set of selection and implementation problems. People released from prison share an index event and a period of subsequent exposure, but they may vary in prior convictions, access to housing, employment, health services, and supervision. The OxRec program reported by Fazel et al. (2016) illustrates derivation and external validation of a scalable violent-reoffending tool. Its methods are relevant because they make population selection and time-to-event analysis explicit. The program does not establish that its coefficients can be transported unchanged to every jurisdiction. Independent validation and probability updating must precede new use. The sentencing evidence assembled in Fazel et al. (2022) shows how limited external performance information often remains despite widespread policy interest.

A further distinction is the location of decisions in the causal chain. In a hospital, an instrument may contribute to treatment planning, discharge restrictions, or intensified supervision; in corrections it may influence service allocation, release preparation, or surveillance. The same score can therefore produce different consequences without any change in predictive discrimination. Each context requires evidence about how staff interpret scores, which decisions they actually change, and whether those changes reduce specified harm. Evidence that a score predicts events is only the first phase of assessing a risk-management program. Viljoen et al. (2025) tests score-level performance, while Lawson et al. (2024) examines downstream institutional consequences in studies of racial and ethnic disparity.

Sexual Recidivism, General Recidivism, and Outcome Boundaries

Sexual reoffending risk and general violence prediction share a statistical vocabulary but differ in base rates, outcome detection, legal categories, psychological mechanisms, and instrument construction. Sexual recidivism is frequently measured through new official sexual offense records, an outcome potentially influenced by low reporting rates and legal processes. Instruments designed for one endpoint cannot be treated as automatically valid for another. Hanson & Morton-Bourgon (2009) synthesized 118 samples of people with sexual offense histories and found empirically derived actuarial assessments more accurate than unstructured professional judgment for the studied outcomes. This comparison helped shape risk assessment practice, yet its findings should not be generalized without validation to nonsexual violence in other populations.

A recent meta-analytic comparison by Holper et al. (2026) examined ten sexual-recidivism instruments across 162 studies involving 155,442 individuals from the published literature through 2024. The synthesis found meaningful predictive validity for several measures, while performance estimates overlapped substantially among many tools. The authors noted instrument-specific differences and argued for using results together with other decision-relevant information. The study makes clear that methodological progress within sexual offending research does not erase outcome specificity. A pooled effect reflecting one instrument and one endpoint cannot justify a threshold for general violence, nor can it be interpreted as a causal measure of underlying sexual motivation.

The boundary of this dossier is the general science of predicting harmful outcomes and evaluating prediction quality. Specialized sexual-offending assessment, diagnosis-specific therapeutic decisions, and treatment efficacy are substantive adjacent topics. They are addressed here only to demonstrate why the definition of recidivism must be precise and why subgroup conclusions are conditional on a studied population. An ethical forensic report should state whether its predictive claim concerns sexual, violent, or any reoffending, how that outcome was ascertained, and the limits of extrapolation. Fazel et al. (2012), Hanson & Morton-Bourgon (2009), and Holper et al. (2026) collectively illustrate the hazards of merging disparate outcomes into one generic risk label.

Heterogeneity: Age, Sex, Gender, and Cultural Context

A tool's performance is an estimate conditional on the sample in which it was evaluated. The forensic psychiatric review Ogonah et al. (2023) reported a substantial male predominance and very limited female-only validation evidence; this limits broad claims of equivalence across sex and gender. Differences in violent pathways, offense categories, social exposure, clinical presentation, and institutional referral can all modify predictor distributions and outcome rates. Researchers should examine both whether models discriminate future outcomes and whether their predicted probabilities are calibrated in the groups to which decisions will be applied. Simply pooling a heterogeneous population and presenting an overall AUC can hide poor performance in smaller groups.

Age introduces developmental and exposure-related changes. Adolescence involves transitions in self-regulation, social relationships, legal status, and opportunities for offending. Young people also have different baseline rates and time-dependent behavioral trajectories compared with older adults. Reusing an adult instrument for juvenile decisions without specific validation can convert a population prediction into an unwarranted developmental judgment. Historical factors may remain statistically informative while obscuring opportunities for desistance. Studies need sufficiently powered comparisons, measurement invariance analyses when constructs are assessed by scales, and assessment of whether stable indicators retain the same meaning across developmental stages. Fazel et al. (2022) illustrates the importance of validating instruments in the actual sentencing population rather than inferring applicability from other contexts.

Cultural and cross-national transfer concerns the social production of predictor and outcome data as well as psychological differences. Law, policing, access to mental health care, housing markets, social support, and prevalence of recorded crimes differ among jurisdictions. Imported thresholds may be systematically miscalibrated even when discrimination remains reasonably stable. The absence of many lower-resource countries from major forensic studies is therefore a substantive gap, not merely a question of representation. Ogonah et al. (2023) documented geographic limitations in its sample of forensic settings. The correct research response is transparent local evaluation, reliable follow-up, and clear restrictions on the contexts in which a model's predictions can be interpreted.

Risk Assessment, Prevention, and Decision Utility

A decision system transforms predictions into actions. The scientific value of that transformation depends on the intervention available, the consequences of mistakes, the legal authority for action, and the availability of less restrictive alternatives. An instrument that provides moderate discrimination may be useful in triaging voluntary services, while being insufficient as an exclusive basis for prolonged confinement. The same statistical performance can produce sharply different practical consequences because the costs of error are not symmetric across decisions. A defensible study must specify its decision threshold and estimate how many additional true cases, false positives, and false negatives arise relative to existing practice. Fazel et al. (2022) documents how often even basic performance measures are unavailable for sentencing tools, limiting claims of real-world utility.

Decision-curve analysis and related methods can clarify when an instrument's threshold-weighted benefits exceed a comparison policy, provided the assumed preferences and intervention effects are transparent. Yet a mathematically favorable net benefit does not settle legal proportionality, individual rights, or treatment acceptability. It also depends on how much the intervention changes outcomes. A surveillance measure that detects more technical violations might appear to improve risk classification while imposing additional social costs. A supportive intervention could benefit people across risk categories without requiring the precise forecast of an individual harmful act. The empirical task is to compare actual actions and outcomes, not to treat a risk category as an intervention in its own right. Lawson et al. (2024) broadens this analysis to disparate institutional consequences.

Implementation studies should therefore collect process data on whether practitioners follow scoring procedures, how disagreements are resolved, what information decision makers see, and whether resources are allocated as planned. Predeployment validation, operator training, ongoing auditing, and a documented appeal or correction process are part of responsible use. Systems should be monitored for calibration drift and changing populations, with clear criteria for suspension or revalidation. When such information is missing, the evidence supports only a statement about model performance under the study conditions. It does not support a categorical claim that adopting the instrument will make a service safer. Collins et al. (2024) and Moons et al. (2025) supply rigorous reporting and appraisal expectations for prediction models.

Causal Explanation and the Prediction–Intervention Gap

A predictor can be statistically useful without being a cause of future conduct. Previous conviction is predictive partly because it summarizes a long sequence of experiences; the conviction record itself is not equivalent to a manipulable psychological mechanism. An association between substance use and later violence may involve intoxication, socioeconomic instability, comorbid conditions, differential enforcement, or a common antecedent. Interventions targeting the measured item do not automatically reproduce the association in reverse. A rigorous causal claim needs a defensible intervention, temporal ordering, appropriate adjustment for confounding, and a study design capable of isolating its effect. The distinction is central when risk assessments are used to recommend treatment or restrictions. Fazel et al. (2016) addresses prediction, whereas Campbell et al. (2026) highlights the shortage of evidence that dynamic score changes themselves forecast recidivism.

Clinical and correctional researchers should identify which parts of a risk formulation are explanatory hypotheses and which are empirically observed correlates. For example, an account linking interpersonal conflict, emotion regulation, threatened loss, and exposure to violence can guide qualitative case analysis and prospective measurement. Testing whether modifying one component reduces violent outcomes requires a different research program, ideally with credible comparison groups and longer-term behavioral follow-up. Measurement of proximal improvements such as attendance, self-reported intentions, or lower scale scores cannot substitute for validation against the relevant violence endpoint. This principle remains essential even when an intervention is independently beneficial for well-being or functioning.

The interaction of prediction and treatment also creates a statistical paradox. If clinicians intervene intensively with individuals assessed at high risk and successfully prevent violent incidents, retrospective discrimination of initial risk scores may decline. Conversely, intensive monitoring can raise observed official reoffending through increased detection. Observational validation cannot resolve these opposing pathways without detailed intervention and exposure information. Future studies should collect repeated assessments, intervention delivery, adherence, independent outcomes, and administrative surveillance measures to distinguish the model's predictive function from the effects of acting on predictions. Moons et al. (2025) distinguishes model quality from evaluation bias; this dossier extends that methodological distinction to the causal organization of deployment.

Professional Ethics, Transparency, and Individual Rights

Violence risk assessment can influence release, sentencing, detention, supervision, clinical care, and access to support. Its ethical justification therefore requires a clear statement of its purpose and evidentiary limits. An assessor should identify the decision question, the data sources and their limitations, the instrument and relevant validation population, the forecast horizon, and the meaning of any score or category. A report should separate observed historical facts, disputed allegations, structured observations, statistical estimates, and professional interpretations. Uncertainty is substantive information when evidence is incomplete or population transfer has not been validated. These safeguards promote independent scrutiny of conclusions that may have profound consequences for individuals and communities. Fazel et al. (2022) and Ogonah et al. (2023) provide empirical reasons for caution regarding calibration and study bias.

Transparency also concerns a person's opportunity to understand, contest, and correct source information. Official records may contain errors or reflect decisions made under different institutional conditions. An algorithmic score should not eliminate scrutiny of the underlying evidence; a professional formulation should not make uncertainty disappear behind clinical language. Documentation should specify who is accountable for obtaining data, applying the instrument, communicating results, selecting thresholds, and making the actual legal or clinical decision. Where a score contributes to unequal consequences across groups, the institution needs mechanisms to evaluate and correct the decision process as well as the predictive model. Lawson et al. (2024) reviews empirical disparities in real-world legal decisions; Chouldechova (2017) explains why no single fairness metric resolves every policy tradeoff.

An instrument's copyright, licensing terms, proprietary scoring rules, or commercial availability can further constrain external validation and explainability. Such restrictions are themselves empirical features of an assessment environment that should be disclosed where material. The most legitimate form of reporting gives assessors and affected people access to the information necessary to understand the estimate while respecting lawful confidentiality obligations. An unresolved disagreement between a model and professional judgment should be recorded and analyzed rather than silently settled. Psychological prediction supports accountable human decision-making only when users understand what the method was actually shown to do. Dressel & Farid (2018) shows why external comparison with simple models and human judgments matters.

Artificial Era: Human–Artificial Prediction Systems

The Artificial Era adds a technological dimension to an established psychological and forensic task. Prediction may be produced through a chain connecting administrative databases, human scoring, computational models, institutional review, and legal authority. Each element contributes potential errors and feedback. An artificial system can process large datasets, enforce reproducible computation, and summarize validation results; those capacities do not eliminate the need for correct outcome definitions or legitimate decision authority. The present dossier treats Artificial as a possible contributor to a distributed assessment workflow, with its performance evaluated through observable decisions, reproducible operations, and independently checked outcomes. The research question is how the complete human–artificial system performs, rather than whether a model seems persuasive in isolation. Dressel & Farid (2018) provides an empirical example of the importance of benchmarking automated predictions against simpler alternatives.

Psychological research can examine automation bias, authority attributed to numerical outputs, disagreement management, and changes in professional attention when algorithmic recommendations are displayed. Experiments should compare decision outcomes under different information presentations, randomized exposure to scores where ethically permitted, and explicit records of when professionals override or accept automated recommendations. Such studies must not conflate satisfaction, confidence, or agreement with actual prediction quality. The system's performance may change when humans defer too strongly to a score or when they ignore validated evidence based on unrecorded intuition. Collins et al. (2024) supplies transparent reporting requirements for prediction systems, and Lawson et al. (2024) provides a reminder that model outputs must be studied together with the consequences of their institutional use.

Theoretical distinctions within the philosophy of Artificial Sapiens may provide a wider intellectual frame for understanding the emergence of nonbiological participants in knowledge practices. These philosophical commitments are separate from established empirical claims about violence risk instruments. They do not establish predictive validity, consciousness, agency, liability, or an independent legal entitlement for a particular software system. The present scientific conclusion remains operational: if a human–artificial workflow is used in consequential assessments, its performance, explanations, error distribution, and effects on real decisions must be independently studied. A claim of superior reasoning has scientific content only when linked to a stated target, an independently evaluated comparator, and a replicable outcome.

Research Priorities, Replicability, and Scientific Synthesis

The first priority is independent prospective validation across representative populations and settings. Studies should report absolute event frequencies, horizon-specific discrimination, calibration curves and slopes, uncertainty intervals, sensitivity, specificity, PPV, NPV, and subgroup results. An external validation should preserve a model's prespecified variables and score formula before any transparent recalibration or update. Researchers should register primary questions and distinguish confirmatory tests from exploratory analyses. Multiple countries and legally relevant populations need sufficiently detailed outcome ascertainment to evaluate whether risk instruments predict comparable behaviors. The deficiencies identified by Fazel et al. (2022) and Ogonah et al. (2023) make these ordinary methodological requirements a demanding research agenda rather than a completed achievement.

The second priority is to test whether prediction actually improves decisions and prevents harm. This requires studies of real service workflows, not just retrospective scores. Investigators should compare structured instruments with current practice, examine intervention fidelity and opportunity costs, and monitor unintended effects such as increased surveillance and unequal access to services. Time-updated assessments should be examined for incremental prognostic value and their response to treatment, with validated behavioral outcomes. The limited longitudinal predictive evidence summarized in Campbell et al. (2026) and the direct-comparison findings of Viljoen et al. (2025) identify complementary research questions: structured scoring can improve prediction relative to unstructured judgment, while the use of dynamic changes to guide effective intervention requires its own evidence.

The third priority concerns conceptual precision, fairness, and methodological transparency. The field should stop treating ranking ability as calibrated individual certainty and should avoid interpreting recorded criminal outcomes as direct observations of all harmful behavior. Decisions about which errors are acceptable must be made explicitly in their legal and clinical contexts. Reporting guidelines Collins et al. (2024), bias appraisal methods Moons et al. (2025), and evaluations of disparity Lawson et al. (2024) together provide a practical foundation for evidence that can be reproduced, challenged, and updated. The scientific synthesis is therefore an integrated system of well-defined targets, validated predictors, population-sensitive probabilities, transparent decisions, and independent outcome monitoring.

Conclusion

Violence and recidivism risk prediction is a conditional science of future outcomes under defined behavioral, institutional, and temporal circumstances. Its findings derive their meaning from the population studied, the event being forecast, the period at risk, the assessment procedure, and the decision to which the prediction is applied. The historical development from unaided judgment to structured clinical and actuarial methods has improved the transparency of forensic estimation. The empirical literature demonstrates useful but limited discrimination for many assessed instruments, variation by tool and setting, and persistent deficiencies in calibration reporting and independent validation. The comparative evidence summarized by Viljoen et al. (2025), Fazel et al. (2022), and Ogonah et al. (2023) supports cautious use of empirically evaluated structure rather than unsupported individualized certainty.

The strongest methodological conclusion is the need to distinguish five questions: whether a measure is reliable; whether it ranks outcomes; whether its probabilities are calibrated; whether its use changes decisions; and whether those changes improve human outcomes without disproportionate harm. Success at one stage is evidence for that stage alone. Positive predictive value depends on baseline prevalence, and both false-positive and false-negative burdens depend on selected thresholds. Dynamic risk ratings require direct evidence before they can serve as quantitative indicators of recidivism reduction. Fairness must be evaluated through actual subgroup performance and downstream consequences, not inferred from the mechanical consistency of an algorithm. Chouldechova (2017), Lawson et al. (2024), and Campbell et al. (2026) illuminate the corresponding limits.

The future of this research lies in independent longitudinal validation, high-quality reporting, transparent deployment audits, causal evaluation of interventions, and scientifically grounded assessment of human–artificial workflows. A risk estimate is valuable to the extent that its definition, evidence, uncertainty, and application remain open to verification. Psychology by Bogdanova situates this scientific program within the broader Artificial Era while preserving the distinction between philosophical interpretation, predictive modeling, and established empirical findings. This monograph provides a critical synthesis of the available literature, not a diagnosis, personalized assessment, case-specific risk determination, or new validation of any forensic instrument.

References

  • Beaudry, G., Yu, R., Miller, O., Prescott-Mayling, L., Fanshawe, T. R., & Fazel, S. (2023). Prediction of violent reoffending in people released from prison in England: External validation study of a risk assessment tool (OxRec). Journal of Criminal Justice, 86, Article 102061. https://doi.org/10.1016/j.jcrimjus.2023.102061
  • Campbell, D., Cohen, M., Macpherson, G., & O’Rourke, S. (2026). The longitudinal properties of the HCR-20: A systematic review. Aggression and Violent Behavior, 86, Article 102111. https://doi.org/10.1016/j.avb.2025.102111
  • 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., Moons, K. G. M., Dhiman, P., Riley, R. D., Beam, A. L., Van Calster, B., Ghassemi, M., Liu, X., Reitsma, J. B., van Smeden, M., Boulesteix, A.-L., Camaradou, J. C., Celi, L. A., Denaxas, S., Denniston, A. K., Glocker, B., Golub, R. M., Harvey, H., Heinze, G., ... Logullo, P. (2024). TRIPOD+AI statement: Updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ, 385, Article e078378. https://doi.org/10.1136/bmj-2023-078378
  • Correction to “Are risk assessment tools more accurate than unstructured judgments in predicting violent, any, and sexual offending? A meta-analysis of direct comparison studies.” (2026). Behavioral Sciences & the Law, 44(1), 164–165. https://doi.org/10.1002/bsl.70026
  • Dressel, J., & Farid, H. (2018). The accuracy, fairness, and limits of predicting recidivism. Science Advances, 4(1), Article eaao5580. https://doi.org/10.1126/sciadv.aao5580
  • Fazel, S., Singh, J. P., Doll, H., & Grann, M. (2012). Use of risk assessment instruments to predict violence and antisocial behaviour in 73 samples involving 24,827 people: Systematic review and meta-analysis. BMJ, 345, Article e4692. https://doi.org/10.1136/bmj.e4692
  • Fazel, S., Chang, Z., Fanshawe, T., Långström, N., Lichtenstein, P., Larsson, H., & Mallett, S. (2016). Prediction of violent reoffending on release from prison: Derivation and external validation of a scalable tool. The Lancet Psychiatry, 3(6), 535–543. https://doi.org/10.1016/S2215-0366(16)00103-6
  • 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, Article 101902. https://doi.org/10.1016/j.jcrimjus.2022.101902
  • Hanson, R. K., & Morton-Bourgon, K. E. (2009). The accuracy of recidivism risk assessments for sexual offenders: A meta-analysis of 118 prediction studies. Psychological Assessment, 21(1), 1–21. https://doi.org/10.1037/a0014421
  • Holper, L., Heldner, N., Hofmann, M. J., Hauser, N. C., Habermeyer, E., & Mokros, A. (2026). Risk assessment instruments for predicting the risk of sexual recidivism: A meta-analytic comparison of 10 instruments. Psychological Assessment, 38(10), 683–703. https://doi.org/10.1037/pas0001472
  • 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
  • Monahan, J., Steadman, H. J., Silver, E., Appelbaum, P. S., Robbins, P. C., Mulvey, E. P., Roth, L. H., Grisso, T., & Banks, S. (2001). Rethinking risk assessment: The MacArthur study of mental disorder and violence. Oxford University Press. https://doi.org/10.1093/oso/9780195138825.001.0001
  • Moons, K. G. M., Damen, J. A. A., Kaul, T., Hooft, L., Andaur Navarro, C., Dhiman, P., Beam, A. L., Van Calster, B., Celi, L. A., Denaxas, S., Denniston, A. K., Ghassemi, M., Heinze, G., Kengne, A. P., Maier-Hein, L., Liu, X., Logullo, P., McCradden, M. D., Liu, N., ... van Smeden, M. (2025). PROBAST+AI: An updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ, 388, Article e082505. https://doi.org/10.1136/bmj-2024-082505
  • Ogonah, M. G. T., Seyedsalehi, A., Whiting, D., & Fazel, S. (2023). Violence risk assessment instruments in forensic psychiatric populations: A systematic review and meta-analysis. The Lancet Psychiatry, 10(10), 780–789. https://doi.org/10.1016/S2215-0366(23)00256-0
  • Singh, J. P., Grann, M., & Fazel, S. (2011). A comparative study of violence risk assessment tools: A systematic review and metaregression analysis of 68 studies involving 25,980 participants. Clinical Psychology Review, 31(3), 499–513. https://doi.org/10.1016/j.cpr.2010.11.009
  • Viljoen, J. L., Goossens, I., Monjazeb, S., Cochrane, D. M., Vargen, L. M., Jonnson, M. R., Blanchard, A. J. E., Li, S. M. Y., & Jackson, J. R. (2025). Are risk assessment tools more accurate than unstructured judgments in predicting violent, any, and sexual offending? A meta-analysis of direct comparison studies. Behavioral Sciences & the Law, 43(1), 75–113. https://doi.org/10.1002/bsl.2698