Forks over knives: Predictive inconsistency in criminal justice algorithmic risk assessment tools
Open accessAbstract
Big data and algorithmic risk prediction tools promise to improve criminal justice systems by reducing human biases and inconsistencies in decision-making. Yet different, equally justifiable choices when developing, testing and deploying these socio-technical tools can lead to disparate predicted risk scores for the same individual. Synthesising diverse perspectives from machine learning, statistics, sociology, criminology, law, philosophy and economics, we conceptualise this phenomenon as predictive inconsistency. We describe sources of predictive inconsistency at different stages of algorithmic risk assessment tool development and deployment and consider how future technological developments may amplify predictive inconsistency. We argue, however, that in a diverse and pluralistic society we should not expect to completely eliminate predictive inconsistency. Instead, to bolster the legal, political and scientific legitimacy of algorithmic risk prediction tools, we propose identifying and documenting relevant and reasonable ‘forking paths’ to enable quantifiable, reproducible multiverse and specification curve analyses of predictive inconsistency at the individual level.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
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).”
From Scoping review on the integration of artificial intelligence and machine learning with structured instruments to predict juvenile recidivism (2026) · cites this paper as Greene et al., 2022
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 measuredComparing 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).”
From Scoping review on the integration of artificial intelligence and machine learning with structured instruments to predict juvenile recidivism (2026) · cites this paper as Greene et al., 2022
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).”
From Scoping review on the integration of artificial intelligence and machine learning with structured instruments to predict juvenile recidivism (2026) · cites this paper as Greene et al., 2022
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Cites
Dissecting racial bias in an algorithm used to manage the health of populations
Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019) · Science · doi:10.1126/science.aax2342 · Cited by 7,137 · cited 1× · see citing works
“For instance, algorithm audits in the health care domain have uncovered evidence of racial bias due to spurious correlations stemming from construct invalidity (Obermeyer et al., 2019).”
BackgroundThe ethics of algorithms: Mapping the debate
Mittelstadt, B. D., Allo, P., Taddeo, M., Wachter, S., & Floridi, L. (2016) · Big Data & Society · doi:10.1177/2053951716679679 · Cited by 3,432 · see citing works
Model Cards for Model Reporting
Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, I. D., & Gebru, T. (2019) · Proceedings of the Conference on Fairness, Accountability, and Transparency · doi:10.1145/3287560.3287596 · Cited by 2,681 · cited 1× · see citing works
“Model-level ‘predictive multiplicity’ metrics (Marx et al., 2020) can be combined with auditing tools such as data sheets (Gebru et al., 2021) and model cards (Mitchell et al., 2019) to help document data and modelling choices relevant to a specific prediction.”
BackgroundGebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., Daumé, H., & Crawford, K. (2018) · arXiv · doi:10.48550/arxiv.1803.09010 · Cited by 2,474
Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., III, H. D., & Crawford, K. (2021) · Communications of the ACM · doi:10.1145/3458723 · Cited by 2,469 · see citing works
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 · see citing works
Groups of diverse problem solvers can outperform groups of high-ability problem solvers
Hong, L., & Page, S. E. (2004) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.0403723101 · Cited by 2,103 · see citing works
Principles alone cannot guarantee ethical AI
Mittelstadt, B. (2019) · Nature Machine Intelligence · doi:10.1038/s42256-019-0114-4 · Cited by 1,886 · cited 1× · see citing works
“One implication is that ARAI designers cannot evade the ethical issues of oper-ationalising ‘essentially contested’ constructs and values such as fairness (Friedler et al., 2021; Jacobs & Wallach, 2021; Mittelstadt, 2019).”
BackgroundD'Ignazio, C., & Klein, L. F. (2020) · doi:10.7551/mitpress/11805.001.0001 · Cited by 1,738 · see citing works
Hill Collins, P. (2002) · Routledge · doi:10.4324/9780203900055 · Cited by 1,532 · see citing works
Mass imprisonment and the life course: Race and class inequality in U.S. incarceration
Pettit, B., & Western, B. (2004) · American Sociological Review · doi:10.1177/000312240406900201 · Cited by 1,499 · see citing works
Validity: on the meaningful interpretation of assessment data
Downing, S. M. (2003) · Medical Education · doi:10.1046/j.1365-2923.2003.01594.x · Cited by 1,434 · see citing works
The accuracy, fairness, and limits of predicting recidivism
Dressel, J., & Farid, H. (2018) · Science Advances · doi:10.1126/sciadv.aao5580 · Cited by 1,071 · see citing works
Fairness in Criminal Justice Risk Assessments: The State of the Art
Berk, R., Heidari, H., Jabbari, S., Kearns, M., & Roth, A. (2021) · Sociological Methods & Research · doi:10.1177/0049124118782533 · Cited by 988 · see citing works
Prisons Do Not Reduce Recidivism: The High Cost of Ignoring Science
Cullen, F. T., Jonson, C. L., & Nagin, D. S. (2011) · The Prison Journal · doi:10.1177/0032885511415224 · Cited by 407 · see citing works
Jacobs, A. Z., & Wallach, H. (2021) · Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency · doi:10.1145/3442188.3445901 · Cited by 362 · see citing works
Friedler, S. A., Scheidegger, C., & Venkatasubramanian, S. (2021) · Communications of the ACM · doi:10.1145/3433949 · Cited by 262 · see citing works
Assessing Risk Assessment in Action
Stevenson, M. T. (2017) · SSRN Electronic Journal · doi:10.2139/ssrn.3016088 · Cited by 195 · see citing works
Procurement As Policy: Administrative Process for Machine Learning
Mulligan, D. K., & Bamberger, K. A. (2019) · SSRN Electronic Journal · doi:10.2139/ssrn.3464203 · Cited by 75 · see citing works
Flagged first, then by citations
Research works printed but not linked: 9
- Ackrill, J. L. (1988) A new Aristotle reader . Princeton University Press.
- (1990) Race and the probability of pleading guilty. Journal of Quantitative Criminology , 6 , 315-334.
- (1992) Charge reduction: An analysis of prosecutorial discretion in burglary and robbery cases. Journal of Quantitative Criminology , 8 , 317-333.
- Ashworth, A. (2005) Sentencing and criminal justice . Cambridge University Press.
- Ashworth, A. and Blake, M. (1996) The presumption of innocence in english criminal law. Criminal Law Review , 306-317.
- Berk, R. (2019) Machine learning risk assessments in criminal justice settings . Springer.
- Berk, R. and Bleich, J. (2014) Forecasts of violence to inform sentencing decisions. Journal of Quantitative Criminology , 30 , 79-96.
- -(2018b) Fairness in machine learning: Lessons from political philosophy. In Conference on Fairness, Accountability and Transparency , 149-159. PMLR.
- -(2020) On the apparent conflict between individual and group fairness. In Proceedings of the 2020 conference on fairness, accountability, and transparency , 514-524.
Cited by
Sorted by citations
Argudo Palacios, A., & Mechraoui, A. (2026) · Cuadernos de RES PUBLICA en derecho y criminología · doi:10.46661/respublica.13244
“…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).”
- Authors
- Greene, T., Shmueli, G., Fell, J., Lin, C.-F., & Liu, H.-W.
- Affiliations
- National Tsing Hua University (TW)
- Institute of Service Science National Tsing Hua University , Hsinchu , Taiwan
- Monash University (AU)
- Corresponding author
- Greene, Travis
- Venue
- Journal of the Royal Statistical Society Series A: Statistics in Society
- Publisher
- Oxford University Press (OUP)
- Published
- 2022-11-30
- Volume
- 185
- Issue
- Supplement_2
- Pages
- S692-S723
- ISSN
- 0964-1998 (ISSN-L)
- 1467-985X
- OpenAlex
- W4296938656
- Cited by
- 15
- References
- 19
- Citations by year
- 2026: 8 · 2025: 1 · 2024: 2 · 2023: 2 · 2022: 1 · 2021: 1
- FWCI
- 1.56
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- academic.oup.com/journals/pages/open_access/funder_policies/chorus/standard_publication_model (vor)
- Locations
- Journal of the Royal Statistical Society Series A (Statistics in Society) (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- Singapore Management University Institutional Knowledge (InK) (Singapore Management University) (submittedVersion, other-oa)
- arXiv (Cornell University) (cc-by)
- Funders
- Taiwan National Science and Technology Council (108-2410-H-007-091-MY3, 111-2628- H-007-001)
- Topics
- Social Sciences › Safety Research › Ethics and Social Impacts of AI
- Social Sciences › Political Science and International Relations › Artificial Intelligence in Law
- Social Sciences › Sociology and Political Science › Criminal Justice and Corrections Analysis
- Keywords
- algorithmic risk assessment, criminal justice, risk prediction tools, specification curve analysis, multiverse analysis, algorithmic bias, reproducibility
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.
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
