Research opportunity · Discretion curtailed, relocated or encoded?
Outside corrections, how often do officials depart from algorithmic recommendations?
Question
Across deployed public-sector decision-support systems in benefits, employment services, child welfare, pretrial and immigration, how often do officials formally or informally depart from the system's recommendation, in which direction, how much of the variation between systems survives once departure is counted the same way, and how do these rates compare with the established override literature for offender risk instruments?
Not yet measuredMaster's thesisLow costPartly addressed
Why it matters
Override figures are quoted as evidence that discretion has or has not survived automation, but each comes from one system, with its own definition, denominator and period. A 0.58% profile-change rate in one country and a 21% screen-out rate in another are not the same quantity. Without a common way of counting, nobody can say whether nominal human review usually means real departure or rubber-stamping.
The debate lists this among what we still do not know: How often do public-sector workers actually depart from algorithmic recommendations across comparable systems?
Evidence that motivates it
- The presence of a human reviewer does not establish meaningful oversight; effective discretion depends on what the human can see, whether they can depart, how often they do so, and the organisational cost of disagreement.
Says effective discretion depends partly on how often humans actually depart; the claim does not license a universal oversight metric, which is why the study reports a comparable descriptive measure rather than a single score of oversight quality. - In Poland’s unemployment profiling system, formal profile changes were extremely rare—0.58% in the cited early-period data—while discretion persisted through interviewing, interpretation and data entry.
Poland's 0.58% formal profile-change rate is an early-period figure for one system, and the claim notes discretion persisted through interviewing and data entry; it is one data point the comparison would standardise, not a lifetime rate. - Allegheny supervisors screened out 21.0% of children marked as mandatory screen-ins; corrected analysis supports the conclusion that worker judgement reduced racial disparity relative to an AFST-only counterfactual.
Allegheny supervisors screened out 21.0% of mandatory screen-ins, a different kind of departure (a different margin and denominator) from Poland's, showing why a common coding scheme is needed.
A possible design
Build a coded dataset of override and deference figures for deployed public-sector decision-support systems (benefits, employment services, child welfare, pretrial, immigration). Sources: peer-reviewed studies, audit and inspection reports, court judgments, algorithmic transparency records, and freedom-of-information requests to agencies that run such systems. For each figure code: what counts as a departure (formal change of score or category, decision against recommendation, informal re-routing), the denominator, direction (more or less severe), period, whether the reviewer could see the score, and the cost of disagreeing (for example, a written justification requirement). Report the spread of rates under a harmonised definition and where harmonisation is impossible, and test whether recorded features such as justification requirements or score visibility go with lower departure rates. The analysis is descriptive; it does not claim that departures are good or bad. Use the override definitions and direction coding of the correctional override literature (see the 2025 scoping review below) as the starting codebook, and report correctional override rates as an external benchmark rather than compiling them again.
Data needed
Published studies and audits; algorithmic transparency records (for example the UK register); FOI responses on override counts; court documents. No individual-level data is required.
Feasibility
Master's thesis. Low: desk research and FOI requests, which are free in many jurisdictions. Many agencies will not hold override counts or will refuse FOI requests; recording which agencies cannot answer is itself a finding. Language coverage limits which countries can be included.
Has this been done?
Partly done already: related work covers some of this question (first-pass screen of 12 October 2026, checked again on 13 October 2026).
In offender risk assessment, override rates and their direction are routinely reported and have been compiled across tools and jurisdictions (a 2025 scoping review and meta-analyses of 21 studies on eight tools), with single-study rates ranging from under 10% to over 70%. Outside corrections, evidence is single-system (for example pretrial bail overrides, child welfare screening, Polish unemployment profiling) or qualitative (a 2024 systematic review of employment-service profiling reports no rates). No compilation was found that codes override and deference rates for benefits, employment services, child welfare, pretrial and immigration systems under one definition, or that compares them with the correctional benchmark.
How strong this evidence is
- Who searched and checked: Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert; an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Every check so far was done by AI; no person or domain expert has checked it yet.
- What was read: 5 of the 8 nearest works below were judged from an abstract, summary or search listing only, and 2 have details that could not be confirmed; each is marked. Not every entry says how much of the work was read.
- Where the search could not reach: Screened by web search on the date shown; Google Scholar, SSRN and paywalled journal pages were not systematically searched, and grey literature in languages other than English was not searched. Absence of a hit is not proof that the work does not exist. Literature check of 2026-10-13: The 2025 Psychology, Public Policy, and Law review was read only as an abstract digest; whether it pools override rates was not confirmed, and its authors come from a co-author's CV. The PMC article by Frechette & Lussier (2023) could not be opened (reCAPTCHA). Wiley and Taylor & Francis pages returned errors. No non-English searches beyond German were run.
- Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.
Nearest work found
- Angelova, Dobbie & Yang, "Algorithmic Recommendations and Human Discretion," NBER Working Paper 31747 (Sept 2023); published in Review of Economic Studies (advance article, 15 Sept 2025), doi:10.1093/restud/rdaf084 Judged from an abstract, summary or search listing only
Single-setting study of judges' overrides of a pretrial algorithm in bail decisions (NBER abstract read; the journal version was not opened): finds about 90% of judges do worse than the algorithm when they override, while about 10% do better on accuracy and fairness. Measures the quality of overrides in one system; does not compare override rates across systems. - Alon-Barkat & Busuioc, "Human–AI Interactions in Public Sector Decision Making: 'Automation Bias' and 'Selective Adherence' to Algorithmic Advice," Journal of Public Administration Research and Theory 33(1), 153–169 (2023)
Three Dutch survey experiments (two citizen samples and 1,345 civil servants) in a simulated contract-renewal task; no evidence of automation bias, and selective adherence to stereotype-congruent advice in one citizen study, not among civil servants. Measures adherence in a vignette, not override rates in deployed systems. - Selten, Robeer & Grimmelikhuijsen, "'Just like I thought': Street-level bureaucrats trust AI recommendations if they confirm their professional judgment," Public Administration Review 83(2), 263–278 (2023) Judged from an abstract, summary or search listing only
Factorial experiment with Dutch police officers using a mock predictive-policing system (abstract read via Crossref): 2 × 2 factorial design; officers trusted and followed recommendations that matched their own judgement, and explanations made no difference. Experimental, one setting; no field override rates. - Northcott, "Automating street-level discretion: A systematic literature review and research agenda," Policy Studies Journal (published online 10 Apr 2025), doi:10.1111/psj.70024 Judged from an abstract, summary or search listing only
Systematic review of how information and communication technologies change street-level discretion (removing face-to-face contact, standardisation, monitoring, decision support) and how frontline workers resist standardisation. Only the abstract was read (via Crossref); it synthesises mechanisms and does not report comparable override rates. - Oliver, Killen, Schmidt & Kroner, "A scoping review and meta-analyses of clinical override use in structured risk assessments: Clinical and training implications for field practice," Psychology, Public Policy, and Law 31(3), 217-233 (2025), doi:10.1037/law0000448 Judged from an abstract, summary or search listing onlySome details not confirmed
Read via an abstract digest (NYU Information for Practice); the APA page was not opened, and the authors are taken from co-author Kroner's CV. Scoping review of 21 studies on eight actuarial risk tools for adult and youth offenders, with meta-analyses showing predictive validity falls after override (AUC about .674 before, .593 after); the abstract says overrides are used frequently. It compiles override practice across tools and jurisdictions within corrections, which is the cross-system comparison the opportunity says is missing, but only for offender risk instruments and with a focus on accuracy rather than a harmonised rate. Whether it reports pooled override rates could not be confirmed from the abstract. - [Authors not confirmed; correspondence author Genevieve Parent], "Clinical Overrides with the YLS/CMI: Predictive Validity and Associated Factors," CrimRxiv preprint (posted 11 Nov 2022) Some details not confirmed
Opened. Single youth-justice study (n = 200) reporting a 33.5% override rate (30.3% upward, 3.2% downward; the abstract says 32.5%) and summarising prior youth studies with rates from 7.0% to 14.1% and up to 74.0% for youth who committed sexual offences. Shows that override rates in corrections are reported by direction and vary widely, which supports the opportunity's premise that definitions and denominators matter, and gives a benchmark sector. A journal version may exist (a Criminal Justice and Behavior DOI, 10.1177/00938548221131958, appeared in a search result) but was not confirmed. - Sirman-Winkler, Szafran & Wang, "Algorithmic Profiling in Public Employment Services: a Systematic Review on the Effects on Caseworkers and Jobseekers," EasyChair Preprint 13691 (June 2024)
Opened (full preprint). Systematic review of nine empirical studies of algorithmic profiling in public employment services; describes caseworker discretion qualitatively (for example advisers in France who ignore the tool, Danish caseworkers interpreting categories flexibly) but reports no override or deviation rates. Confirms the gap for employment services. - Bjerre-Nielsen, Rogers, Lassen, Chrisander, Raaschou-Pedersen, Grenet & Nielsen, "Trading off performance and human oversight in algorithmic policy: evidence from Danish college admissions," arXiv:2411.15348 (Nov 2024) Judged from an abstract, summary or search listing only
Read through a third-party summary page (pith.science), not the arXiv page itself. Does not measure overrides; it borrows an 18% override rate from bail decisions as a cost assumption. Illustrates the demand for comparable override figures outside the setting where they were measured, which is the problem the opportunity addresses.
Literature check
Narrowed to what the literature leaves open. The claim that no cross-system compilation of override rates exists is too broad. In offender risk assessment there is an established override literature that reports override rates and direction across tools and jurisdictions, now synthesised in a 2025 scoping review and meta-analyses of 21 studies covering eight tools (Psychology, Public Policy, and Law 31(3), doi:10.1037/law0000448), and single studies there report rates from about 7% to 74% with upward and downward overrides coded separately. No compilation was found for benefits, employment services, child welfare, pretrial or immigration systems, or across those sectors under one definition, so the opportunity should be narrowed to those sectors and use the correctional literature as a benchmark and coding template.
Searched: web (extended); web (standard); arxiv.org; osf.io; ssrn.com; academic.oup.com; onlinelibrary.wiley.com; journals.sagepub.com; tandfonline.com; sciencedirect.com; dl.acm.org; easychair.org; crimrxiv.com; ifp.nyu.edu (abstract digest); German-language web. Checked by Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert.
The screen before this check (verdict changed)
The search found no study that does the core design (first-pass screen of 12 October 2026).
No cross-system compilation of real override or deference rates under a common definition was found. Single-system quasi-experimental studies exist (for example pretrial bail overrides), systematic reviews synthesise how digital tools change discretion without compiling rates, and survey experiments test whether officials follow algorithmic advice in simulated tasks, but neither yields comparable rates across deployed systems.
Search queries used
override rates algorithmic risk assessment caseworkers comparison across systems public sector deferencesystematic review human override of algorithmic recommendations public administration frontline workers adherence rates"override rate" pretrial risk assessment OR child welfare screening tool OR unemployment profiling judges caseworkers empiricalcompilation of override and adherence rates across deployed algorithmic decision aids public sector review "adherence" "override" meta-analysis 2024 2025systematic review human discretion override algorithmic decision support street-level bureaucrats deviation rates meta-analysis 2025
Literature check, 13 October 2026:
caseworker deviation from algorithmic profiling recommendation share of cases public employment service empiricalhuman override algorithmic risk assessment rates review of studies pretrial child welfare probation "override" frequencystreet-level bureaucrats compliance with algorithmic advice field data override frequency [allowed_domains: arxiv.org, osf.io, ssrn.com, papers.ssrn.com]algorithmic decision support public administration discretion override rate empirical [allowed_domains: academic.oup.com, onlinelibrary.wiley.com, journals.sagepub.com, tandfonline.com, sciencedirect.com]Abweichung Sachbearbeiter algorithmische Empfehlung AMS-Algorithmus Arbeitsmarktchancen Korrektur Anteil"Algorithmic Recommendations and Human Discretion" Angelova Dobbie Yang extends 2026Frechette Lussier "Betting Against the Odds" clinical override risk assessment review override rates studiesmeta-analysis professional override structured risk assessment instruments override rate across studies Law and Human Behavior"A scoping review and meta-analyses of clinical override use in structured risk assessments"clinical override scoping review meta-analysis 2025 Psychology, Public Policy, and Law 31(3) 217 authorsAllegheny Family Screening Tool call screeners overrode recommendation rate disagreement study [allowed_domains: arxiv.org, dl.acm.org]
Provenance
RO-LD05-01 v2. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-05 debate on 2026-10-12. Checked by an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.
Changes
- v2, 13 October 2026: Narrowed after the literature check. The claim that no cross-system compilation of override rates exists is too broad. In offender risk assessment there is an established override literature that reports override rates and direction across tools and jurisdictions, now synthesised in a 2025 scoping review and meta-analyses of 21 studies covering eight tools (Psychology, Public Policy, and Law 31(3), doi:10.1037/law0000448), and single studies there report rates from about 7% to 74% with upward and downward overrides coded separately. No compilation was found for benefits, employment services, child welfare, pretrial or immigration systems, or across those sectors under one definition, so the opportunity should be narrowed to those sectors and use the correctional literature as a benchmark and coding template. The verdict changed from "No study found doing the core design" to "Related work exists". See v1 as published.
Earlier title: How often do officials depart from algorithmic recommendations, across systems?
Earlier question: Across deployed public-sector risk and profiling systems, how often do officials formally or informally depart from the system's recommendation, in which direction, and how much of the variation between systems survives once override is defined and counted the same way?