Scoping review on the integration of artificial intelligence and machine learning with structured instruments to predict juvenile recidivism
Open accessAbstract
Predicting the risk of recidivism in juvenile justice has traditionally relied on structured instruments, while the incorporation of AI/ML models has been proposed as a means of improving their performance. However, the actual added value of these models, their integration with structured tools, and their implications for practice remain a subject of debate. This scoping review evaluates whether integrating AI/ML models with structured risk assessment tools in juvenile justice meaningfully enhances recidivism prediction and practice. Five studies that met these criteria were included. Overall, the results show a consistent pattern of moderate discrimination and generally marginal improvements as model complexity increases, suggesting the existence of a certain empirical limit to predictive capacity in this field. Improvements are more likely when the comparator is strictly linear or when contrasted with unstructured judgment, suggesting that the added value of AI/ML may lie in modeling interactions between variables rather than in a substantial increase in accuracy. Nevertheless, aggregate performance may mask relevant differences among subgroups, including disparities associated with sociodemographic variables, which raises implications regarding equity and institutional legitimacy. Added to this are the challenges arising from their implementation, including model traceability, the interaction between professionals and these automated systems, and the need for frameworks that ensure oversight and auditing.
AI snapshot
Written from the paper’s full text.
TL;DR
This scoping review maps five empirical studies that link AI/ML models to structured juvenile risk assessment instruments, finding consistently moderate discrimination (AUC 0.69–0.75) and only marginal gains as model complexity increases. The authors argue the added value of AI/ML lies mainly in modelling non-linear interactions rather than substantially improving accuracy, while equity, traceability, and professional-oversight concerns remain unresolved.
Key findings
- Across the five included studies, models achieved AUC values from 0.69 to 0.75, with incremental value over comparators ranging from 0.00 to 0.09 AUC points.
- Improvements ranged from none (0.74 RF vs. 0.74 LR in Sheppard et al., 2024) to moderate (0.75 GBDT vs. 0.67 MJCA in Mori et al., 2024; 0.73 ML vs. 0.64 SAVRY in Miron et al., 2021; 0.69 RF vs. 0.64 LR in Ting et al., 2018); Kigerl et al. (2022) found differences between model families of less than 0.01 (0.72 stacking vs. 0.71 LR).
- Gains appear more likely when the comparator is strictly linear or an unstructured human score, suggesting AI/ML captures non-linear relationships between risk factors rather than raising accuracy substantially.
- Aggregate performance can mask subgroup disparities: Miron et al. (2021) reported increased false positives penalising vulnerable subgroups such as young foreigners, with disparities by nationality and gender.
Methods
- Scoping review following Peters et al. (2020), Arksey and O'Malley (2005), and reported per PRISMA-ScR; guiding question concerned empirical evidence on AI/ML for juvenile recidivism prediction and its comparison with or integration into structured instruments.
- Searches in Web of Science, Scopus, PsycINFO, and PubMed (last search 2 February 2026) yielded 28 records; after removing 5 duplicates, 23 were screened, 8 reviewed in full text, and 5 studies included. Inclusion required peer-reviewed empirical articles from 2016 onward on youth in the juvenile justice system with an explicit link to structured risk instruments.
- Included studies came from the United States (n=2), Spain (n=1), Japan (n=1), and Singapore (n=1), with samples from N=855 to N=227,451, using WSJCA, SAVRY, MJCA, PAT, and YLS/CMI 2.0. Data extraction coded usage strategy (integration, comparison, optimization) and summarised performance via AUC.
Limitations
- Stated by the authors: only five studies were included, limiting statistical generalizability; high heterogeneity in modelling techniques, comparators, and jurisdictions prevented quantitative synthesis; publication bias cannot be ruled out given variable transparency in validation, calibration, and subgroup analysis.
- Evident from the text: explicit assessment of group equity was rare in the reviewed literature, and the review reports no calibration, external validation, or temporal-stability metrics for the included models, which the authors themselves argue are necessary alongside AUC.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
11 of 70 cited sources checked · 5 supported · 0 flagged · accuracy not yet measured
- Premisecitation
When there are historical differences in baseline rates of criminal prevalence between socio-demographic groups, it is mathematically difficult to satisfy all statistical notions of equity.
“As demonstrated in the specialist literature, when there are historical differences in baseline rates of criminal prevalence between socio-demographic groups, it is mathematically difficult to satisfy all statistical notions of equity (Chouldechova, 2017;Hardt et al., 2016;Kleinberg et al., 2017).”
Quote in the paperSubstantial claim1 of 3 sources checked- cites Chouldechova, 2017 source
Section 2.1 (p. 2) states that when recidivism prevalence differs between groups, a test-fair score cannot have equal false-positive and false-negative rates across those groups. This supports the claimed mathematical trade-off, but the paper does not characterize those differences as historical or establish a result about all statistical notions of equity. (LLM, conf=0.86)
Experimental — accuracy not yet measured - cites Hardt et al., 2016
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Kleinberg et al., 2017
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationreasoning
The actual impact of integrating AI/ML with structured instruments depends not only on discrimination, but also on decision-making behavior and the framework for use, including training, protocols, thresholds, documentation, and continuous auditing.
“Therefore, the actual impact of integrating AI/ML with structured instruments depends not only on discrimination, but also on decision-making behavior and the framework for use, including training, protocols, thresholds, documentation, and continuous auditing (Doleac & Stevenson, 2024;Green & Chen, 2021;Greene et al., 2022).”
Quote in the paperSubstantial claim1 of 3 sources checked- cites Doleac & Stevenson, 2024
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Green & Chen, 2021
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Greene et al., 2022 source
The discussion of judicial overrides (p. 30) says final decisions can diverge from algorithmic recommendations, while §4.6 (p. 28) discusses how risk thresholds and score presentation affect use. Section 6 (p. 33) proposes documenting modelling choices to enable auditing. These passages support the broader point that predictive performance alone does not determine an ARAI’s impact, but they do not establish the claim’s specific requirements for practitioner training, protocols, or continuous auditing when integrating AI/ML with structured instruments. (LLM, conf=0.84)
Experimental — accuracy not yet measured
- Interpretationcitation
The pattern of marginal improvements is consistent with recent methodological critiques in criminal justice that describe predictive inconsistency among algorithmic tools and question the widespread superiority of opaque models when more interpretable and auditable alternatives exist.
“This pattern is consistent with recent methodological critiques in criminal justice that describe predictive inconsistency among algorithmic tools and question the widespread superiority of opaque models when more interpretable and auditable alternatives exist (Greene et al., 2022;Rudin, 2019).”
Quote in the paperNot a substantial claim1 of 2 sources checked- cites Greene et al., 2022 source
- cites Rudin, 2019
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Premisecitation
Comparing results across models is not straightforward because there is a wide variety of methods used, ranging from the type of artificial intelligence chosen to how exactly recidivism is defined or at what stage of the trial the prediction is made.
“However, recent literature warns that comparing results across models is not straightforward.This is because there is a wide variety of methods used, ranging from the type of artificial intelligence chosen to how exactly recidivism is defined or at what stage of the trial the prediction is made (Eaglin, 2016;Engel & Grgić-Hlača, 2021;Greene et al., 2022;Kigerl et al., 2022;Travaini et al., 2022).”
Quote in the paperSubstantial claim2 of 3 sources checked- cites Eaglin, 2016 source
- cites Engel & Grgić-Hlača, 2021 source
The Introduction distinguishes general from violent recidivism, and the additional details (p. 333) distinguish predictions used in sentencing from those used in bail decisions. These passages support variation in the outcome predicted and the decision stage, but the paper does not establish the claimed difficulty of comparing results across AI models or attribute it to a variety of AI methods. (LLM, conf=0.83)
Experimental — accuracy not yet measured - cites Greene et al., 2022 source
- Findingcitation
In certain assessment scenarios, improvements in performance were found compared to linear approaches.
“In this context, improvements in performance were found compared to linear approaches in certain assessment scenarios (Ting et al., 2018).”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Ting et al., 2018
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Findingreasoning
When evaluated against a direct comparator, the incremental value ranged from 0.00 to 0.09 AUC points.
“When evaluated against a direct comparator, the incremental value ranged from 0.00 to 0.09 AUC points.”
Quote in the paperSubstantial claim- Rests on an assumption: The AUC values reported in the included studies are comparable across studies despite differences in methods, definitions, and prediction stages.
- Rests on an assumption: The five included studies are sufficient to map the available empirical evidence on the use of AI/ML in predicting juvenile recidivism with structured instruments.
- Findingnone given
Explicit assessment of group equity was rare in the reviewed literature.
“Explicit assessment of group equity was rare in the reviewed literature.”
Quote in the paperSubstantial claim - Findingown data
Overall, the results show generally marginal improvements as model complexity increases.
“Overall, the results show a consistent pattern of moderate discrimination and generally marginal improvements as model complexity increases”
Quote in the paperSubstantial claim- Rests on an assumption: The AUC values reported in the included studies are comparable across studies despite differences in methods, definitions, and prediction stages.
- Rests on an assumption: The five included studies are sufficient to map the available empirical evidence on the use of AI/ML in predicting juvenile recidivism with structured instruments.
- Findingown data
The methodological link between AI/ML and these tools is grouped into three categories: Integration, Comparison, and Optimization.
“Regarding the methodological link between AI/ML and these tools, the evidence is grouped into three categories:”
Quote in the paperNot a substantial claim - Findingown data
The five selected studies were conducted in the United States (n=2), Spain (n=1), Japan (n=1), and Singapore (n=1).
“The five selected studies were conducted in the United States (n= 2), Spain (n= 1), Japan (n= 1), and Singapore (n= 1), with sample sizes ranging from N= 855 to N= 227,451.”
Quote in the paperNot a substantial claim - Findingcitation
Virtually identical performance was documented between complex techniques and traditional approaches such as logistic regression, especially when the sample size is large, or the evaluated scenario reduces predictive variability.
“Virtually identical performance was documented between complex techniques and traditional approaches such as logistic regression, especially when the sample size is large, or the evaluated scenario reduces predictive variability (Kigerl et al., 2022;Sheppard et al., 2024).”
Quote in the paperSubstantial claim0 of 2 sources checked- Rests on an assumption: The AUC values reported in the included studies are comparable across studies despite differences in methods, definitions, and prediction stages.
- Rests on an assumption: The five included studies are sufficient to map the available empirical evidence on the use of AI/ML in predicting juvenile recidivism with structured instruments.
- Findingcitation
Under certain methodological and comparative conditions, AI/ML models can outperform traditional assessments.
“Despite the previously described asymptotic limit, the review also confirms that, under certain methodological and comparative conditions, AI/ML models can outperform traditional assessments.”
Quote in the paperSubstantial claim - Findingown data
Overall, the models achieved AUC values ranging from 0.69 to 0.75.
“Overall, the models achieved AUC values ranging from 0.69 to 0.75.”
Quote in the paperSubstantial claim- Rests on an assumption: The AUC values reported in the included studies are comparable across studies despite differences in methods, definitions, and prediction stages.
- Rests on an assumption: The five included studies are sufficient to map the available empirical evidence on the use of AI/ML in predicting juvenile recidivism with structured instruments.
- Findingcitation
The study by Miron et al. (2021) shows that integrating AI/ML models with the SAVRY instrument can achieve an overall improvement in discriminatory power compared to traditional scoring.
“The study by Miron et al. (2021) shows that by integrating AI/ML models with the SAVRY instrument, it is possible to achieve an overall improvement in discriminatory power compared to traditional scoring.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Miron et al., 2021
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Findingown data
Overall, the results show a consistent pattern of moderate discrimination.
“Overall, the results show a consistent pattern of moderate discrimination and generally marginal improvements as model complexity increases”
Quote in the paperSubstantial claim- Rests on an assumption: The AUC values reported in the included studies are comparable across studies despite differences in methods, definitions, and prediction stages.
- Rests on an assumption: The five included studies are sufficient to map the available empirical evidence on the use of AI/ML in predicting juvenile recidivism with structured instruments.
- Findingcitation
In several studies, performance differences between model families were marginal.
“In several studies, performance differences between model families were marginal.”
Quote in the paperNot a substantial claim- Rests on an assumption: The AUC values reported in the included studies are comparable across studies despite differences in methods, definitions, and prediction stages.
- Rests on an assumption: The five included studies are sufficient to map the available empirical evidence on the use of AI/ML in predicting juvenile recidivism with structured instruments.
- Findingown data
The analysis of the five included studies provides evidence on the actual capabilities and limitations of AI/ML in clinical and judicial practice.
“The analysis of the five included studies provides evidence on the actual capabilities and limitations of these technologies in clinical and judicial practice.”
Quote in the paperSubstantial claim- Rests on an assumption: The five included studies are sufficiently methodologically sound for their reported predictive performance to be synthesized as evidence on AI/ML capabilities.
- Findingcitation
In the included sample, maximizing predictive performance can introduce disparities based on nationality and sex.
“In the included sample, it was documented that maximizing predictive performance can introduce disparities based on nationality and sex.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Miron et al., 2021
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Findingcitation
The transition from traditional statistical models to complex AI/ML architectures yields only marginal improvements in studies with large samples and preventative approaches.
“The results of studies with large samples (Kigerl et al., 2022) and preventative approaches (Sheppard et al., 2024) demonstrate that the transition from traditional statistical models to complex AI/ML architectures yields only marginal improvements.”
Quote in the paperSubstantial claim0 of 2 sources checked - Findingcitation
Ting et al. (2018) found that the model captures non-linear relationships between risk factors, thereby improving screening compared to traditional statistical methods.
“The model captures non-linear relationsh ips between risk factors, thereby improvin g screening compared to traditional statistical methods.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Ting et al., 2018
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Findingcitation
In studies that incorporate structured information from the instrument as predictor variables, the results are consistent with non-linear relationships between risk factors.
“In studies that incorporate structured information from the instrument as predictor variables, the results are consistent with the identification of non-linear relationships between risk factors.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Ting et al., 2018
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Findingown data
Sample sizes ranged from N=855 to N=227,451.
“The five selected studies were conducted in the United States (n= 2), Spain (n= 1), Japan (n= 1), and Singapore (n= 1), with sample sizes ranging from N= 855 to N= 227,451.”
Quote in the paperNot a substantial claim - Findingcitation
Mori et al. (2024) found that the ML model outperforms the MJCA score in terms of discrimination (ΔAUC = 0.08).
“The ML model outperforms the MJCA score in terms of discriminati on (ΔAUC = 0.08).”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Mori et al., 2024 source
- Findingcitation
Sheppard et al. (2024) found similar performance across techniques.
“Similar performa nce across technique s.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Sheppard et al., 2024 source
- Findingown data
The structured instruments used were: WSJCA, SAVRY, MJCA, PAT, and YLS/CMI 2.0.
“The structured instruments used were: WSJCA (Barnoski, 2004), SAVRY (Borum et al., 2006), MJCA (Correction Bureau, Ministry of Justice of Japan; see UNAFEI, 2019), PAT (Florida Department of Juvenile Justice, 2014), and YLS/CMI 2.0 (Hoge & Andrews, 2011).”
Quote in the paperSubstantial claim - Findingcitation
When directly comparing computational performance against the score assigned by the human evaluator or against strictly linear comparators, moderate improvements in discriminative capacity are observed.
“Empirical results demonstrate that, when directly comparing computational performance against the score assigned by the human evaluator (Mori et al., 2024) or against strictly linear comparators (Ting et al., 2018), moderate improvements in discriminative capacity are observed.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites Mori et al., 2024 source
- cites Ting et al., 2018
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Findingcitation
When comparing AI/ML models with the traditional SAVRY instrument, an increase in false positives was observed that specifically penalizes vulnerable subgroups such as young foreigners.
“Specifically, when comparing the models with the traditional instrument (SAVRY), an increase in false positives was observed, which specifically penalizes vulnerable subgroups, such as young foreigners (Miron et al., 2021).”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Miron et al., 2021
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Findingcitation
Kigerl et al. (2022) found that sample size has a greater influence than algorithmic complexity.
“Sample size has a greater influence than algorithm ic complexit y.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Kigerl et al., 2022 source
- Findingcitation
Kigerl et al. (2022) found that differences between model families are marginal (<0.01).
“The difference s between model families are marginal (<0.01).”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Kigerl et al., 2022 source
- Interpretationcitation
Reliance on baseline recidivism rates and their variability across time periods and jurisdictions, coupled with data quality limitations, imposes a methodological barrier that AI/ML, on its own, cannot overcome.
“Reliance on baseline recidivism rates and their variability across time periods and jurisdictions, coupled with data quality limitations, imposes a methodological barrier that AI/ML, on its own, cannot overcome (Berk, 2021;Steyerberg, 2019;Weber et al., 2025).”
Quote in the paperSubstantial claim0 of 3 sources checked- cites Berk, 2021 source
- cites Steyerberg, 2019
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Weber et al., 2025
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Study design and conductreasoning
Publication bias cannot be ruled out, particularly in the literature on predictive models, where transparency in validation, calibration and subgroup analysis is variable.
“Third, publication bias cannot be ruled out, particularly in the literature on predictive models, where transparency in validation, calibration and subgroup analysis is variable.”
Quote in the paperNot a substantial claim - Study design and conductown data
High heterogeneity in modelling techniques, comparators used, and jurisdictional contexts prevented quantitative synthesis from being carried out.
“Second, the high heterogeneity in modelling techniques, comparators used, and jurisdictional contexts prevented quantitative synthesis from being carried out.”
Quote in the paperSubstantial claim - Premisecitation
The introduction of automated recommendations may alter cognitive decision-making processes.
“Although the studies included in this review focus on statistical performance, research warns that the introduction of automated recommendations may alter cognitive decision-making processes.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites Doleac & Stevenson, 2024
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Green & Chen, 2021
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Study design and conductown data
The small number of included studies (five) limits the statistical generalizability of the results.
“First, the small number of included studies (five) limits the statistical generalizability of the results, although this number reflects the nascent state of the empirical literature that specifically examines the intersection of AI/ML with structured instruments in youth populations.”
Quote in the paperSubstantial claim - Premisecitation
Juvenile criminal behaviour is subject to fluctuations linked to dynamic risk factors, maturation processes, and contextual contingencies, which may limit the temporal stability of predictions based on one-off measurements.
“Juvenile criminal behaviour is subject to fluctuations linked to dynamic risk factors, maturation processes, and contextual contingencies, which may limit the temporal stability of predictions based on one-off measurements (Cuevas et al., 2019).”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Cuevas et al., 2019 source
- Interpretationcitation
Aggregate performance may mask relevant differences among subgroups, including disparities associated with sociodemographic variables.
“Nevertheless, aggregate performance may mask relevant differences among subgroups, including disparities associated with sociodemographic variables”
Quote in the paperSubstantial claim- Rests on an assumption: The AUC values reported in the included studies are comparable across studies despite differences in methods, definitions, and prediction stages.
- Rests on an assumption: The five included studies are sufficient to map the available empirical evidence on the use of AI/ML in predicting juvenile recidivism with structured instruments.
- Premisecitation
AI/ML models applied to criminology cannot extract a strong predictive signal if it does not exist in the administrative and historical data sets.
“In this regard, as Berk (2021) warns, AI/ML models applied to criminology cannot extract a strong predictive signal if it does not exist in the administrative and historical data sets.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Berk, 2021 source
- Interpretationreasoning
There is a certain empirical limit to predictive capacity in this field.
“suggesting the existence of a certain empirical limit to predictive capacity in this field”
Quote in the paperSubstantial claim- Rests on an assumption: The AUC values reported in the included studies are comparable across studies despite differences in methods, definitions, and prediction stages.
- Rests on an assumption: The five included studies are sufficient to map the available empirical evidence on the use of AI/ML in predicting juvenile recidivism with structured instruments.
- Interpretationcitation
The argument that the implementation of AI/ML will lead to a substantial improvement in accuracy within the juvenile justice system does not appear to be consistently supported by the currently available evidence.
“Consequently, the argument that the implementation of AI/ML will lead to a substantial improvement in accuracy within the juvenile justice system does not appear to be consistently supported by the currently available evidence (Kigerl et al., 2022;Sheppard et al., 2024).”
Quote in the paperSubstantial claim0 of 2 sources checked - Interpretationreasoning
The added value of AI/ML may lie in modeling interactions between variables rather than in a substantial increase in accuracy.
“suggesting that the added value of AI/ML may lie in modeling interactions between variables rather than in a substantial increase in accuracy”
Quote in the paperSubstantial claim- Rests on an assumption: The AUC values reported in the included studies are comparable across studies despite differences in methods, definitions, and prediction stages.
- Rests on an assumption: The five included studies are sufficient to map the available empirical evidence on the use of AI/ML in predicting juvenile recidivism with structured instruments.
- Premisecitation
The deployment of AI/ML in juvenile justice raises challenges regarding traceability.
“As an extension of the demands for fairness, the deployment of AI/ML in juvenile justice raises challenges regarding traceability.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites Wang et al., 2023 source
- cites Završnik, 2021
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationcitation
The predictive ceiling suggests that the performance of risk assessment tools is more dependent on the quality, validity, and structure of the available data than on the sophistication of the algorithm.
“This predictive ceiling suggests that the performance of risk assessment tools is more dependent on the quality, validity, and structure of the available data than on the sophistication of the algorithm (Berk, 2021;Steyerberg, 2019).”
Quote in the paperSubstantial claim0 of 2 sources checked- cites Berk, 2021 source
- cites Steyerberg, 2019
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Classified as probably not a substantial claim. 17
- May combine two claims. 21
Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).
PremisecitationThere is a clear limit to technological reformism in justice systems.
“As Ugwudike (2022) points out, there is a clear limit to technological reformism in justice systems.”
Quote in the paperNot a substantial claim1 of 1 sources checked- cites Ugwudike, 2022 source
Probably not a substantial claim: may not pass S2 (checkable).
PremisecitationEven if there is a moderate increase in discrimination, methodological evaluation cannot be limited to reporting AUC, given that applied utility depends on how the prediction is used to make concrete decisions.
“However, even if there is a moderate increase in discrimination, methodological evaluation cannot be limited to reporting AUC, given that applied utility depends on how the prediction is used to make concrete decisions (Monahan & Skeem, 2016;Steyerberg, 2019).”
Quote in the paperNot a substantial claim1 of 2 sources checked- cites Monahan & Skeem, 2016 source
- cites Steyerberg, 2019
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
May combine two claims.
PremisecitationThe deployment of risk scores in criminal and juvenile justice systems has distributive consequences and directly affects institutional legitimacy.
“This finding shifts the debate to the realm of algorithmic equity, given that the deployment of risk scores in criminal and juvenile justice systems has distributive consequences and directly affects institutional legitimacy (Garrett & Monahan, 2020;Završnik, 2020Završnik, , 2021)).”
Quote in the paperSubstantial claim1 of 3 sources checked- cites Garrett & Monahan, 2020
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Završnik, 2020 source
Sections 3.1 and 3.2.1 connect criminal-justice risk assessments and predictive systems to consequential decisions and unequal treatment. The Introduction says opaque AI decision-making can undermine legitimacy, but the copy does not establish that risk scores directly affect institutional legitimacy or address their deployment in juvenile justice. (LLM, conf=0.84)
Experimental — accuracy not yet measured - cites Završnik, 2021
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Study design and conductcitationThe study design and implementation followed the methodological guidelines outlined in the chapter on scoping reviews by Peters et al. (2020), supplemented by the classic methodological framework of Arksey and O'Malley (2005).
“The study design and implementation followed the methodological guidelines outlined in the chapter on scoping reviews by Peters et al. (2020), supplemented by the classic methodological framework of Arksey and O'Malley (2005).”
Quote in the paperNot a substantial claim1 of 2 sources checkedMay combine two claims.
PremisecitationGeneric policies on human oversight do not guarantee effective control unless they define when and how one may override the system, what justifications are required, and what audits are applied; otherwise, oversight may become a mere formality.
“Furthermore, generic policies on human oversight do not guarantee effective control unless they define when and how one may override the system, what justifications are required, and what audits are applied; otherwise, oversight may become a mere formality (Green, 2022;Engel & Grgić-Hlača, 2021).”
Quote in the paperSubstantial claim1 of 2 sources checked- cites Green, 2022
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Engel & Grgić-Hlača, 2021 source
Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.
Experimental — accuracy not yet measured
May combine two claims.
PremisecitationWithin the juvenile justice system, the true value of an AI/ML model depends on the specific clinical or judicial decision it aims to support and on how a continuous score translates into operational thresholds with asymmetric consequences between false positives and false negatives.
“Within the context of the juvenile justice system, the true value of an AI/ML model depends on the specific clinical or judicial decision it aims to support (e.g., the intensity of supervision, the prioritization of socio-educational programs, or referral to specialist resources), and on how a continuous score translates into operational thresholds with asymmetric consequences between false PositI”
Quote in the paperSubstantial claim1 of 3 sources checked- cites Eaglin, 2017
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Garrett & Monahan, 2020
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Monahan & Skeem, 2016 source
Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.
Experimental — accuracy not yet measured
May combine two claims.
FindingcitationAggregate performance can hide relevant differences between subgroups, including disparities associated with sociodemographic variables, which introduces implications for equity and institutional legitimacy.
“Sin embargo, el rendimiento agregado puede ocultar diferencias relevantes entre subgrupos, incluyendo disparidades asociadas a variables sociodemográficas, lo que introduce implicaciones en términos de equidad y legitimidad institucional.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Findingown dataFollowing deduplication, 5 duplicate records were removed, leaving 23 unique records for screening.
“Following deduplication, 5 duplicate records were removed, leaving 23 unique records for screening.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingown dataEight reports were reviewed in full text.
“Eight reports were reviewed in full text, of which three were excluded: two for not having an explicit link to a structured risk assessment instrument and one for being additional publication due to double-counting, thus avoiding the duplication of evidence.”
Quote in the paperNot a substantial claimMay combine two claims.
FindingreasoningThere is an empirical limit in the predictive capacity of AI/ML in juvenile recidivism, with results showing a consistent pattern of moderate predictive discrimination and generally marginal improvements when model complexity increases.
“The first cross-sectional finding of this review is the existence of an empirical limit in the predictive capacity of AI/ML, showing stagnation at levels of moderate predictive discrimination.”
- Also stated in: § 4.5 Limitations and Conclusion · hedged
“En conjunto, los resultados muestran un patrón consistente de discriminación moderada y mejoras generalmente marginales al incrementar la complejidad de los modelos, lo que sugiere la existencia de cierto límite empírico en la capacidad predictiva en este ámbito.”
Quote in the paperSubstantial claim- Rests on an assumption: The five included studies are sufficiently methodologically sound for their reported predictive performance to be synthesized as evidence on AI/ML capabilities.
- Also stated in: § 4.5 Limitations and Conclusion · hedged
Probably not a substantial claim: may not pass S3 (load-bearing).
Findingown dataThe literature search identified 28 records.
“The literature search identified 28 records, distributed across PsycINFO (n= 7), PubMed (n= 2), Web of Science (n= 8), and Scopus (n= 11).”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingown dataDuring the title and abstract selection phase, 15 records were excluded for not meeting the eligibility criteria.
“During the title and abstract selection phase, 15 records were excluded for not meeting the eligibility criteria.”
Quote in the paperNot a substantial claimMay combine two claims.
FindingcitationIn the study by Miron et al. (2021), the average gain in discriminatory power coexists with significant predictive disparities based on the nationality and gender of the minors assessed, so overall improvement does not guarantee a uniform impact across subgroups.
“However, this study introduces a nuance: this average gain coexists with significant predictive disparities based on the nationality and gender of the minors assessed, underscoring that the overall improvement in performance does not guarantee a uniform impact across subgroups.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Miron et al., 2021
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Classified as probably not a substantial claim.
Findingown dataFive studies that met the criteria were included in the scoping review.
“Five studies that met these criteria were included.”
- Also stated in: § 3.1 Study Selection · plain
“Consequently, five studies were ultimately included.”
Quote in the paperSubstantial claim- Rests on an assumption: The five included studies are sufficient to map the available empirical evidence on the use of AI/ML in predicting juvenile recidivism with structured instruments.
- Also stated in: § 3.1 Study Selection · plain
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Findingown dataThree reports were excluded: two for not having an explicit link to a structured risk assessment instrument and one for being additional publication due to double-counting.
“Eight reports were reviewed in full text, of which three were excluded: two for not having an explicit link to a structured risk assessment instrument and one for being additional publication due to double-counting, thus avoiding the duplication of evidence.”
Quote in the paperNot a substantial claimMay combine two claims.
FindingreasoningImprovements are more likely when the comparator is strictly linear or when contrasted with unstructured judgment.
“Improvements are more likely when the comparator is strictly linear or when contrasted with unstructured judgment”
Quote in the paperSubstantial claim- Rests on an assumption: The AUC values reported in the included studies are comparable across studies despite differences in methods, definitions, and prediction stages.
- Rests on an assumption: The five included studies are sufficient to map the available empirical evidence on the use of AI/ML in predicting juvenile recidivism with structured instruments.
May combine two claims.
FindingcitationMiron et al. (2021) found that machine learning improves predictive accuracy, but introduces disparities based on nationality and gender.
“Machine learning improves predictive accuracy, but introduce s disparities based on nationalit y and gender.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Miron et al., 2021
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
May combine two claims.
FindingcitationThe variability ranged from no predictive differences (0.74 vs. 0.74 in Sheppard et al., 2024) to moderate improvements (0.75 vs. 0.67 in Mori et al., 2024; 0.73 vs. 0.64 in Miron et al., 2021; and 0.69 vs. 0.64 in Ting et al., 2018).
“This variability ranged from no predictive differences (0.74 vs. 0.74 in Sheppard et al., 2024) to moderate improvements (0.75 vs. 0.67 in Mori et al., 2024;0.73 vs. 0.64 in Miron et al., 2021;and 0.69 vs. 0.64 in Ting et al., 2018).”
Quote in the paperSubstantial claim0 of 3 sources checked- cites Sheppard et al., 2024 source
- cites Mori et al., 2024 source
- cites Miron et al., 2021
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Rests on an assumption: The AUC values reported in the included studies are comparable across studies despite differences in methods, definitions, and prediction stages.
- Rests on an assumption: The five included studies are sufficient to map the available empirical evidence on the use of AI/ML in predicting juvenile recidivism with structured instruments.
May combine two claims.
PremisecitationAn AI/ML system may rank subjects relatively well and yet still produce poorly calibrated probabilities or deviate due to changes in data distribution across different contexts and time periods, which seriously compromises its usefulness.
“An AI/ML system may rank subjects relatively well and yet still produce poorly calibrated probabilities or deviate due to changes in data distribution across different contexts and time periods, which seriously compromises its usefulness (Moons et al., 2019;Steyerberg, 2019;Wolff et al., 2019).”
Quote in the paperSubstantial claim0 of 3 sources checkedProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductcitationThe drafting of the final report was structured in accordance with the PRISMA-ScR statement checklist (Tricco et al., 2018).
“The drafting of the final report was structured in accordance with the PRISMA-ScR statement checklist (Tricco et al., 2018).”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Tricco et al., 2018
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
May combine two claims.
PremisecitationProcessing architectures can hinder the legal rationale behind decisions and the detection of underlying flaws or biases.
“Since processing architectures can hinder the legal rationale behind decisions and the detection of underlying flaws or biases (Wang et al., 2023), criminology argues that systems must be designed to ensure institutional inspection and control (Završnik, 2021).”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Wang et al., 2023 source
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Study design and conductnone givenThe selection of studies was carried out independently by two reviewers in two consecutive stages.
“The selection of studies was carried out independently by two reviewers in two consecutive stages.”
Quote in the paperNot a substantial claimMay combine two claims.
PremisereasoningEven when technical transparency is ensured, practical feasibility depends on how professionals integrate the scores into their decision-making and which organizational rules they apply.
“However, even when technical transparency is ensured, practical feasibility depends on how professionals integrate the scores into their decision-making and which organizational rules they apply.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S2 (checkable). May combine two claims.
InterpretationcitationThe cost of complexity may outweigh the benefit if it reduces the capacity for auditing and practical explanation, especially when the incremental gain in performance is small.
“This requirement is perhaps most relevant when the incremental gain in performance is small, because the cost of complexity may outweigh the benefit if it reduces the capacity for auditing and practical explanation (Rudin, 2019;Wang et al., 2023).”
Quote in the paperNot a substantial claim0 of 2 sources checked- cites Rudin, 2019
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Wang et al., 2023 source
Probably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone givenAny disagreements regarding the decision to include a study were resolved through discussion and consensus among the authors.
“Any disagreements regarding the decision to include a study were resolved through discussion and consensus among the authors.”
Quote in the paperNot a substantial claimMay combine two claims.
InterpretationcitationThe potential added value of AI/ML, when observed, may lie in capturing nonlinear interactions and relationships between variables, especially when the comparator is linear or when the human score does not explicitly incorporate such combinations; improvements are more likely when the comparator is strictly linear or when compared with unstructured judgment.
“These findings are consistent with the idea that the potential added value of AI/ML, when observed, may lie in capturing nonlinear interactions and relationships between variables, especially when the comparator is linear or when the human score does not explicitly incorporate such combinations (Mori et al., 2024;Steyerberg, 2019;Ting et al., 2018).”
- Also stated in: § 4.5 Limitations and Conclusion · hedged
“Las mejoras son más probables cuando el comparador es estrictamente lineal o cuando se contrasta con juicio no estructurado, lo que apunta a que el valor añadido de la IA/ML puede residir en modelar interacciones entre variables más que en un incremento sustancial de la precisión.”
Quote in the paperSubstantial claim0 of 3 sources checked- cites Mori et al., 2024 source
- cites Steyerberg, 2019
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Ting et al., 2018
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Also stated in: § 4.5 Limitations and Conclusion · hedged
May combine two claims.
InterpretationreasoningThe synthesized evidence suggests that AI/ML can provide flexible modeling and represent nonlinear relationships, but its improvement in predictive performance in juvenile justice appears limited given the current state of data, evaluation designs, and institutional definitions of recidivism as an outcome variable.
“In conclusion, the synthesized evidence suggests that AI/ML can provide flexible modeling and represent nonlinear relationships, but its improvement in predictive performance in juvenile justice appears limited given the current state of data, evaluation designs, and institutional definitions of recidivism as an outcome variable.”
Quote in the paperSubstantial claim- Rests on an assumption: The five included studies are sufficiently methodologically sound for their reported predictive performance to be synthesized as evidence on AI/ML capabilities.
- Rests on an assumption: The recidivism outcome definitions used across the included studies are sufficiently comparable to allow synthesis.
Probably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone givenData extraction was carried out using a standardized form specifically developed for the purpose of this review.
“Data extraction was carried out using a standardized form specifically developed for the purpose of this review.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone givenThe inclusion and exclusion criteria were established in accordance with the PCC (Population, Concept, and Context) framework.
“The inclusion and exclusion criteria were established in accordance with the PCC (Population, Concept, and Context) framework.”
Quote in the paperNot a substantial claimMay combine two claims.
PremisecitationPhenomena such as anchoring bias against machine scoring, overconfidence in technology, and the displacement of responsibility to the algorithm are well-documented.
“Phenomena such as anchoring bias against machine scoring, overconfidence in technology, and the displacement of responsibility to the algorithm are well-documented (Doleac & Stevenson, 2024;Green & Chen, 2021).”
Quote in the paperSubstantial claim0 of 2 sources checked- cites Doleac & Stevenson, 2024
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Green & Chen, 2021
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone givenThe deduplication process combined an initial automated phase using the platform's tools with a subsequent manual verification to ensure the integrity of the database.
“The deduplication process combined an initial automated phase using the platform's tools with a subsequent manual verification to ensure the integrity of the database.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim. May combine two claims.
Interpretationnone givenChallenges derived from implementation include traceability of models, interaction between professionals and automated systems, and the need for frameworks that guarantee supervision and auditing.
“A ello se suman los desafíos derivados de su implementación, incluyendo la trazabilidad de los modelos, la interacción entre profesionales y sistemas automatizados, y la necesidad de marcos que garanticen supervisión y auditoría.”
Quote in the paperSubstantial claim
The paper relies on this without support.
The AUC values reported in the included studies are comparable across studies despite differences in methods, definitions, and prediction stages.
“However, recent literature warns that comparing results across models is not straightforward.This is because there is a wide variety of methods used, ranging from the type of artificial intelligence chosen to how exactly recidivism is defined or at what stage of the trial the prediction is made (Eaglin, 2016;Engel & Grgić-Hlača, 2021;Greene et al., 2022;Kigerl et al., 2022;Travaini et al., 2022).”
“Overall, the models achieved AUC values ranging from 0.69 to 0.75.”
The paper relies on this without support.
The five included studies are sufficient to map the available empirical evidence on the use of AI/ML in predicting juvenile recidivism with structured instruments.
“Five studies that met these criteria were included.”
“Consequently, five studies were ultimately included.”
The paper relies on this without support.
The five included studies are sufficiently methodologically sound for their reported predictive performance to be synthesized as evidence on AI/ML capabilities.
“The analysis of the five included studies provides evidence on the actual capabilities and limitations of these technologies in clinical and judicial practice.”
The paper relies on this without support.
The recidivism outcome definitions used across the included studies are sufficiently comparable to allow synthesis.
“In conclusion, the synthesized evidence suggests that AI/ML can provide flexible modeling and represent nonlinear relationships, but its improvement in predictive performance in juvenile justice appears limited given the current state of data, evaluation designs, and institutional definitions of recidivism as an outcome variable.”
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
SAVRY study assigned conflicting usage strategies
MinorThe text identifies SAVRY as an example of optimization, defined as rescaling items or deriving new indices. Table 3 instead classifies the SAVRY study as a mix of item comparison and integration.
“Optimization: where AI/ML is used to rescale items or derive new indices from the original structure (e.g., SAVRY; PAT).”
“Miron et al. (2021) Spain (N = 855) SAVRY Logit, MLP, kNN, SVM, DT, NB, RF.Mixed (item comparison and integration)”
Cites
PRISMA Extension for Scoping Reviews (PRISMA-ScR): Checklist and Explanation
Tricco, A. C., Lillie, E., Zarin, W., O'Brien, K. K., Colquhoun, H., Levac, D., Moher, D., Peters, M. D. J., Horsley, T., Weeks, L., et al. (2018) · Annals of Internal Medicine · doi:10.7326/m18-0850 · Cited by 43,565 · see citing works
Scoping studies: towards a methodological framework
Arksey, H., & O'Malley, L. (2005) · International Journal of Social Research Methodology · doi:10.1080/1364557032000119616 · Cited by 36,742 · cited 1× · see citing works
“The study design and implementation followed the methodological guidelines outlined in the chapter on scoping reviews by Peters et al. (2020), supplemented by the classic methodological framework of Arksey and O’Malley (2005).”
An introduction to statistical learning with applications in R
Sohil, F., Sohali, M. U., & Shabbir, J. (2021) · Statistical Theory and Related Fields · doi:10.1080/24754269.2021.1980261 · Cited by 5,145 · cited 1× · see citing works
“…of AI that trains algorithms using millions of previously observed cases to learn which combinations of variables are associated with a relevant outcome, such as recidivism or re-engagement with the system, and to estimate that risk in new cases (Berk, 2021; James et al., 2021; Wang et al., 2023).”
PROBAST: A Tool to Assess the Risk of Bias and Applicability of Prediction Model Studies
Wolff, R. F., Moons, K. G. M., Riley, R. D., Whiting, P. F., Westwood, M., Collins, G. S., Reitsma, J. B., Kleijnen, J., Mallett, S., & for the PROBAST Group† (2019) · Annals of Internal Medicine · doi:10.7326/m18-1376 · Cited by 2,984 · cited 1× · see citing works
“An AI/ML system may rank subjects relatively well and yet still produce poorly calibrated probabilities or deviate due to changes in data distribution across different contexts and time periods, which seriously compromises its usefulness (Moons et al., 2019; Steyerberg, 2019; Wolff et al., 2019).”
Peters, M., Godfrey, C., McInerney, P., Munn, Z., Tricco, A., & Khalil, H. (2019) · JBI Reviewer's Manual · doi:10.46658/jbirm-20-01 · Cited by 2,222 · cited 1× · see citing works
“The study design and implementation followed the methodological guidelines outlined in the chapter on scoping reviews by Peters et al. (2020), supplemented by the classic methodological framework of Arksey and O’Malley (2005).”
Fair prediction with disparate impact: a study of bias in recidivism prediction instruments
Chouldechova, A. (2017) · Big Data · doi:10.1089/big.2016.0047 · Cited by 2,159 · cited 1× · see citing works
“As demonstrated in the specialist literature, when there are historical differences in baseline rates of criminal prevalence between socio-demographic groups, it is mathematically difficult to satisfy all statistical notions of equity (Chouldechova, 2017; Hardt et al., 2016; Kleinberg et al., 2017).”
Moons, K. G. M., Wolff, R. F., Riley, R. D., Whiting, P. F., Westwood, M., Collins, G. S., Reitsma, J. B., Kleijnen, J., & Mallett, S. (2019) · Annals of Internal Medicine · doi:10.7326/m18-1377 · Cited by 1,669 · see citing works
The Psychology of Criminal Conduct
Bonta, J., & Andrews, D. A. (2023) · Routledge eBooks · doi:10.4324/9781003292128 · Cited by 760 · see citing works
Youth Level of Service/Case Management Inventory
Hoge, R. (2011) · doi:10.4324/9780203843666-11 · Cited by 348 · cited 1×
“One of the international benchmarks for these tools is the Youth Level of Service/Case Management Inventory (YLS/CMI) developed by Hoge and Andrews (2002), designed to guide the entire process for young people, from assessment through to case closure.”
Risk Assessment in Criminal Sentencing
Monahan, J., & Skeem, J. L. (2016) · Annual Review of Clinical Psychology · doi:10.1146/annurev-clinpsy-021815-092945 · Cited by 256 · cited 2× · see citing works
“However, even if there is a moderate increase in discrimination, methodological evaluation cannot be limited to reporting AUC, given that applied utility depends on how the prediction is used to make concrete decisions (Monahan & Skeem, 2016; Steyerberg, 2019).”
Criminal justice, artificial intelligence systems, and human rights
Završnik, A. (2020) · ERA Forum · doi:10.1007/s12027-020-00602-0 · Cited by 231 · see citing works
Artificial Intelligence, Predictive Policing, and Risk Assessment for Law Enforcement
Berk, R. A. (2021) · Annual Review of Criminology · doi:10.1146/annurev-criminol-051520-012342 · Cited by 213 · cited 6× · see citing works
“This predictive ceiling suggests that the performance of risk assessment tools is more dependent on the quality, validity, and structure of the available data than on the sophistication of the algorithm (Berk, 2021; Steyerberg, 2019).”
InfluentialPrimer on Risk Assessment and the Statistics Used to Evaluate Its Accuracy
Helmus, L., & Babchishin, K. (2017) · Criminal Justice and Behavior · doi:10.1177/0093854816678898 · Cited by 125 · cited 1× · see citing works
“…approach allowed for flexibility, it suffered from serious reliability issues; for example, two evaluators could reach opposing conclusions about the same minor, making the judicial measure more dependent on who was evaluating than on the characteristics of the case (Helmus & Babchishin, 2017).”
Machine Learning for Brain Disorders
Colliot, O. (2023) · Neuromethods · doi:10.1007/978-1-0716-3195-9 · Cited by 120 · see citing works
Parole decision-making: A salient factor score
Hoffman, P. B., & Beck, J. L. (1974) · Journal of Criminal Justice · doi:10.1016/0047-2352(74)90031-2 · Cited by 103 · cited 1× · see citing works
“One example was the Salient Factor Score (Hoffman & Beck, 1974), used to objectify parole decisions.”
In Pursuit of Interpretable, Fair and Accurate Machine Learning for Criminal Recidivism Prediction
Wang, C., Han, B., Patel, B., & Rudin, C. (2023) · Journal of Quantitative Criminology · doi:10.1007/s10940-022-09545-w · Cited by 101 · cited 3× · see citing works
“Since processing architectures can hinder the legal rationale behind decisions and the detection of underlying flaws or biases (Wang et al., 2023), criminology argues that systems must be designed to ensure institutional inspection and control (Završnik, 2021).”
Travaini, G. V., Pacchioni, F., Bellumore, S., Bosia, M., & Micco, F. D. (2022) · International Journal of Environmental Research and Public Health · doi:10.3390/ijerph191710594 · Cited by 82 · cited 2× · see citing works
“…there is a wide variety of methods used, ranging from the type of artificial intelligence chosen to how exactly recidivism is defined or at what stage of the trial the prediction is made (Eaglin, 2016; Engel & Grgić-Hlača, 2021; Greene et al., 2022; Kigerl et al., 2022; Travaini et al., 2022).”
Youth Level of Service/Case Management Inventory 2.0
Hoge, R. D., & Andrews, D. A. (2012) · PsycTESTS Dataset · doi:10.1037/t05078-000 · Cited by 72 · see citing works
Algorithmic Risk Assessment in the Hands of Humans
Stevenson, M., & Doleac, J. L. (2019) · doi:10.2139/ssrn.3489440 · Cited by 60 · cited 2× · see citing works
“Phenomena such as anchoring bias against machine scoring, overconfidence in technology, and the displacement of responsibility to the algorithm are well-documented (Doleac & Stevenson, 2024; Green & Chen, 2021).”
Campos, C. R., Guedes, C., Diniz, C., et al. (2018) · Arte e ensaios · doi:10.37235/ae.n36.22487 · Cited by 57
Youth Level of Service/Case Management Inventory 2.0 (YLS/CMI 2.0)
(2019) · The SAGE Encyclopedia of Criminal Psychology · doi:10.4135/9781483392240.n550 · Cited by 56 · cited 3×
“The structured instruments used were: WSJCA (Barnoski, 2004), SAVRY (Borum et al., 2006), MJCA (Correction Bureau, Ministry of Justice of Japan; see UNAFEI, 2019), PAT (Florida Department of Juvenile Justice, 2014), and YLS/CMI 2.0 (Hoge & Andrews, 2011).”
InfluentialEvaluating causes of algorithmic bias in juvenile criminal recidivism
Miron, M., Tolan, S., Gómez, E., & Castillo, C. (2020) · Artificial Intelligence and Law · doi:10.1007/s10506-020-09268-y · Cited by 51 · cited 4× · see citing works
“The study by Miron et al. (2021) shows that by integrating AI/ML models with the SAVRY instrument, it is possible to achieve an overall improvement in discriminatory power compared to traditional scoring.”
InfluentialConstructing Recidivism Risk for Sentencing
Eaglin, J. (2016) · SSRN Electronic Journal · doi:10.2139/ssrn.2821136 · Cited by 45 · see citing works
Ghasemi, M., Anvari, D., Atapour, M., Wormith, J., Stockdale, K. C., & Spiteri, R. (2020) · Criminal Justice and Behavior · doi:10.1177/0093854820969753 · Cited by 39 · cited 1× · see citing works
“…language of intervention, allowing us to assess what the technology contributes to addressing the same problem and how it can be integrated without losing the technical justification for decisions (Ghasemi et al., 2021; Raiche et al., 2023; Sheppard et al., 2024; Tasharrofi & Barnes, 2022).”
Risk-Need-Responsivity (RNR) classification models: Still evolving
Taxman, F. S., & Smith, L. (2020) · Aggression and Violent Behavior · doi:10.1016/j.avb.2020.101459 · Cited by 37 · cited 1× · see citing works
“Finally, the responsiveness principle adapts the method, requiring that strategies be tailored to the minor's learning capacity and motivation to ensure the effectiveness of the treatment (Taxman & Smith, 2021).”
Dynamic risk factors and timing of recidivism for youth in residential placement
Cuevas, C., Wolff, K. T., & Baglivio, M. T. (2019) · Journal of Criminal Justice · doi:10.1016/j.jcrimjus.2018.10.003 · Cited by 35 · cited 2× · see citing works
“An AI/ML system may rank subjects relatively well and yet still produce poorly calibrated probabilities or deviate due to changes in data distribution across different contexts and time periods, which seriously compromises its usefulness (Moons et al., 2019; Steyerberg, 2019; Wolff et al., 2019).”
Wormith, J. S., & Zidenberg, A. M. (2018) · Springer eBooks · doi:10.1007/978-3-030-01030-0_2 · Cited by 30 · cited 1× · see citing works
“The need principle, for its part, determines what to treat, focusing on direct criminogenic factors (such as antisocial attitudes or substance use) rather than nonspecific clinical aspects (Wormith & Zidenberg, 2018).”
Ting, M. H., Chu, C. M., Zeng, G., Li, D., & Chng, G. S. (2017) · Journal of Social Work · doi:10.1177/1468017317743137 · Cited by 29 · cited 4× · see citing works
“In this context, improvements in performance were found compared to linear approaches in certain assessment scenarios (Ting et al., 2018).”
InfluentialPredictive Algorithms in Justice Systems and the Limits of Tech-Reformism
Ugwudike, P. (2022) · International Journal for Crime, Justice and Social Democracy · doi:10.5204/ijcjsd.2189 · Cited by 29 · cited 1× · see citing works
“As Ugwudike (2022) points out, there a clear limit to technological reformism in justice systems.”
Ley Orgánica 5/2000 de 12 de enero, reguladora de la responsabilidad penal de los menores
Generales, E. C. (2000) · doi:10.14198/altern2000.8.12 · Cited by 26
Algorithmic Risk Assessment in the Hands of Humans
Stevenson, M. T., & Doleac, J. L. (2024) · American Economic Journal Economic Policy · doi:10.1257/pol.20220620 · Cited by 24 · see citing works
Factors influencing acceptance and trust of chatbots in juvenile offenders’ risk assessment training
Raiche, A.-P., Dauphinais, L., Duval, M., Luca, G. D., Rivest‐Hénault, D., Vaughan, T., Proulx, C., & Guay, J. (2023) · Frontiers in Psychology · doi:10.3389/fpsyg.2023.1184016 · Cited by 20 · cited 1× · see citing works
“…language of intervention, allowing us to assess what the technology contributes to addressing the same problem and how it can be integrated without losing the technical justification for decisions (Ghasemi et al., 2021; Raiche et al., 2023; Sheppard et al., 2024; Tasharrofi & Barnes, 2022).”
Washington State Juvenile Court Assessment
Barnoski, R. (2004) · PsycTESTS Dataset · doi:10.1037/t04488-000 · Cited by 20 · see citing works
Engel, C., & Grgić-Hlača, N. (2021) · Journal of Legal Analysis · doi:10.1093/jla/laab001 · Cited by 16 · cited 2× · see citing works
“Furthermore, generic policies on human oversight do not guarantee effective control unless they define when and how one may override the system, what justifications are required, and what audits are applied; otherwise, oversight may become a mere formality (Green, 2022; Engel & Grgić-Hlača, 2021).”
Forks over knives: Predictive inconsistency in criminal justice algorithmic risk assessment tools
Greene, T., Shmueli, G., Fell, J., Lin, C.-F., & Liu, H.-W. (2022) · Journal of the Royal Statistical Society Series A: Statistics in Society · doi:10.1111/rssa.12966 · Cited by 15 · cited 3× · see citing works
“…there is a wide variety of methods used, ranging from the type of artificial intelligence chosen to how exactly recidivism is defined or at what stage of the trial the prediction is made (Eaglin, 2016; Engel & Grgić-Hlača, 2021; Greene et al., 2022; Kigerl et al., 2022; Travaini et al., 2022).”
Hilterman, E., Vermunt, J., Nicholls, T. L., et al. (2019) · International Journal of Forensic Mental Health · doi:10.1080/14999013.2019.1580803 · Cited by 11
Predicción de la reincidencia con delincuentes juveniles: adaptación del IGI-J
Genovés, V., Martín, E., & Doménech, M. (2017) · doi:10.4995/reinad.2017.6484 · Cited by 10 · cited 1×
“The assessment focuses on identifying deficits in areas such as parental supervision, educational performance or the nature of their relationships with peer groups-variables that recent research confirms as predictors of juvenile recidivism (Garrido et al., 2017).”
Andrés-Pueyo, A. (2017) · doi:10.1016/j.apj.2017.02.003 · Cited by 10 · cited 2×
“These instruments, based on the Risk-Need model, represented a conceptual revolution by evaluating not only static (unchangeable) history, but also the subject's current and modifiable situation (Loinaz, 2017).”
The great methods bake-off: Comparing performance of machine learning algorithms
Kigerl, A., Hamilton, Z., Kowalski, M., & Mei, X. (2022) · Journal of Criminal Justice · doi:10.1016/j.jcrimjus.2022.101946 · Cited by 8 · cited 4× · see citing works
“…there is a wide variety of methods used, ranging from the type of artificial intelligence chosen to how exactly recidivism is defined or at what stage of the trial the prediction is made (Eaglin, 2016; Engel & Grgić-Hlača, 2021; Greene et al., 2022; Kigerl et al., 2022; Travaini et al., 2022).”
InfluentialHistory of the Juvenile Justice System
Thompson, K., & Morris, R. (2016) · doi:10.1007/978-3-319-29343-1_5 · Cited by 6 · cited 1×
“In the United States, the origin of the juvenile justice system dates back to 1899 with the enactment of the Illinois Juvenile Court Act legislation, which led to the creation of the first specialized court in Chicago (Thompson & Morris, 2016).”
A fairness scale for real-time recidivism forecasts using a national database of convicted offenders
Verrey, J., Neyroud, P., Sherman, L., & Ariel, B. (2025) · Neural Computing and Applications · doi:10.1007/s00521-025-11478-x · Cited by 5 · cited 1× · see citing works
“Unlike a traditional structured instrument, where criteria and their weighting are predefined, a machine learning model automatically adjusts its parameters during training by comparing its predictions with actual outcomes to minimize error (Colliot, 2023; Verrey et al., 2025).”
Mori, T., Harada, K., Yamaki, M., & Yamaguchi, M. (2024) · Psychology, Public Policy, and Law · doi:10.1037/law0000421 · Cited by 5 · see citing works
Sheppard, K. G., Talaugon, A. R., & Hernandez, J. L. (2024) · Journal of Criminal Justice · doi:10.1016/j.jcrimjus.2024.102262 · Cited by 4 · cited 5× · see citing works
“Virtually identical performance was documented between complex techniques and traditional approaches such as logistic regression, especially when the sample size is large, or the evaluated scenario reduces predictive variability (Kigerl et al., 2022; Sheppard et al., 2024).”
InfluentialYukhnenko, D., Blackwood, N., Lichtenstein, P., et al. (2025) · Law and Human Behavior · doi:10.1037/lhb0000641 · Cited by 3 · cited 1×
“This shift does not imply the disappearance of standardized instruments or case management logic, but rather a change in focus towards models trained on large databases and longitudinal records (Travaini et al., 2022; Yukhnenko et al., 2025).”
Optimizing Community Supervision Practices with the Elo-rating System: A Proof-of-Concept
Tasharrofi, S., & Barnes, J. C. (2022) · Justice Evaluation Journal · doi:10.1080/24751979.2022.2052343 · Cited by 2 · cited 1×
“…language of intervention, allowing us to assess what the technology contributes to addressing the same problem and how it can be integrated without losing the technical justification for decisions (Ghasemi et al., 2021; Raiche et al., 2023; Sheppard et al., 2024; Tasharrofi & Barnes, 2022).”
Flagged first, then by citations
7 of the 52 references couldn’t be matched to a known paper.
Research works printed but not linked: 5
- Individualización de las medidas de sanción en la persona adolescente. En Manual de justicia penal para adolescentes (pp. 417- 456).
- BONTA, J., & ANDREWS, Donald. A. (2023). The psychology of criminal conduct (7a ed.). Routledge.
- PMCid:PMC10312240
- STEYERBERG, Ewout. W. (2019). Clinical prediction models: A practical approach to development, validation, and updating. Springer International Publishing. https://doi.org/10.1007/978-3-030-16399- · DOI
- PMCid:PMC12401775
- Authors
- Argudo Palacios, A., & Mechraoui, A.
- Venue
- Cuadernos de RES PUBLICA en derecho y criminología
- Publisher
- Universidad Pablo de Olavide
- Published
- 2026-05-25
- Pages
- 1-16
- ISSN
- 2990-0697 (ISSN-L)
- OpenAlex
- W7162437235
- Semantic Scholar
- bc4e3a9db17ad9b414372f2ec6ddc5e8bb0d1442
- Cited by
- 0
- References
- 52
- FWCI
- 0.00
- Type
- journal article
- Language
- en
- Open access
- diamond
- doi.org
- Licence
- Locations
- Cuadernos de RES PUBLICA en derecho y criminología (publishedVersion, cc-by-nc)
- RIO (Universidad Pablo de Olavide) (submittedVersion, cc-by-nc)
- Topics
- Social Sciences › Safety Research › Ethics and Social Impacts of AI
- Social Sciences › Psychology › Clinical Psychology › Psychopathy, Forensic Psychiatry, Sexual Offending
- Social Sciences › Political Science and International Relations › Artificial Intelligence in Law
- Keywords
- artificial intelligence, machine learning, juvenile justice, algorithmic discrimination, algorithmic fairness, sociodemographic disparities, model traceability, algorithmic auditing
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-28.
Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-28.
Cited sources are checked by gpt-6-sol.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
