Research opportunity · Discretion curtailed, relocated or encoded?

What does being flagged by a fraud-risk model cost claimants before any decision?

Question

For benefit claimants selected for extra checks by a risk model, how large are the burdens imposed before any formal decision (payment delay, evidence requests, withdrawn or abandoned claims, repeat applications), how do they compare with claimants selected at random or not selected, and how are they distributed by nationality and age?

Not yet measuredOne PhD chapter or paperLow costPartly addressed

Why it matters

Selection for scrutiny is often treated as harmless because a human makes the final decision. If selection itself costs time, money or claims, then disparities in who is selected are disparities in treatment, even where final decisions are fair, and they are hard to challenge because no adverse decision exists yet. Current evidence is either a single median figure from the agency or qualitative testimony.

The debate lists this among what we still do not know: How common are selection-stage burdens that are difficult to challenge because no formal adverse decision has yet occurred?

Evidence that motivates it

Non-UK nationals were over-referred by the DWP Universal Credit Advances model in both audited periods; some older groups were over-referred in 2024–25, creating additional verification burden before the final decision.
Non-UK nationals and some older groups were over-referred by the DWP Advances model, creating extra verification burden; the claim does not say they were denied more often because of the model, so the study measures burden, not denial attributable to the model.
Algorithmically mediated selection can create material scrutiny and administrative burden before a final legal decision exists.
States that algorithmically mediated selection can create material burden before a final decision, without licensing a claim that every such burden is harmful; the study measures how large the burden is.
DUO’s risk-based selection for home visits indirectly discriminated against students with a migration background; the risk-profile selection was abandoned and current home-visit selection is random.
DUO replaced risk-profile selection for home visits with random selection, offering a comparison point where selection no longer depends on a risk profile.

A possible design

Two linked parts. (1) Administrative comparison: the DWP transparency record says referrals mix model-selected and control-group cases. Use FOI requests and parliamentary questions to obtain aggregate outcomes for three groups (model-referred, control-referred, automatically approved): distribution of payment delay (not only the median), share declined, share withdrawn or not pursued, and repeat requests, broken down by nationality and age band. Start from the published effectiveness assessment, which reports only a median delay, a model-versus-control effectiveness ratio and relative referral likelihoods by group, and request the missing breakdowns. Repeat for one or two other systems with published selection data (for example Dutch student-finance checks before and after the switch to random selection). (2) Claimant-side study: interviews or a short survey of claimants and welfare advisers recruited through advice charities, recording time spent, documents requested, missed payments and knowledge of why they were checked. Compare the agency figures with what claimants report.

Data needed

Aggregate FOI data from benefit agencies; published transparency records and effectiveness assessments; access to claimants through advice organisations; ethics approval for interviewing people in financial hardship.

Feasibility

One PhD chapter or paper. Low: FOI requests are free; interview incentives and transcription cost a few thousand pounds. Agencies may withhold figures on fraud-control grounds (the DWP record already withholds frequency and scale under FOI section 31). The control-group comparison only works if the agency releases figures by referral type.

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).

The DWP transparency record and its 2025-26 effectiveness assessment report only a one-day median delay for approved referred Advances, a model-versus-control effectiveness ratio and relative referral likelihoods by age and nationality; advocacy evidence to Parliament (November 2025) confirms that no delay distribution, maximum delay or evidence-request burden has been published. For Dutch student-finance checks, Algorithm Audit (2024) traced each control step by migration background and found non-European-background students had three times the chance of an unnecessary home visit, but it reports no delays, time costs or abandoned grants, and its random-sample groups were too small for group comparison. Journalistic and civil-society work on Rotterdam documents burden qualitatively. No study was found that measures pre-decision burden (delay distribution, evidence requests, withdrawn or abandoned claims) for model-selected versus randomly selected or unselected claimants, by nationality and age.

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: none of the 7 nearest works below is recorded as judged from an abstract only. 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; the Journal of Policy Analysis and Management article "Administrative Burden and the Access–Fraud Trade-Off" (Bell & Jilke, 2026) was checked via its Crossref abstract and is a survey experiment on public tolerance of fraud versus exclusion, not a study of flagged claimants. Google Scholar, SSRN and paywalled journals were not systematically searched, and Dutch- and Danish-language sources were not searched. Absence of a hit is not proof that the work does not exist. Literature check of 2026-10-13: The March 2024 Algorithm Audit main report (which holds the random-sample results) was not opened. Danish-language sources were not searched. Wiley pages returned errors. The later part of the search budget was exhausted, so no forward search on the DWP effectiveness assessment was 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

Department for Work and Pensions (UK): "DWP: Universal Credit Advances Model," UK Algorithmic Transparency Record (published 16 Dec 2025)
Agency's own record. Reports a one-day longer median payment delay for approved referred Advances, states that referrals mix high-risk and control-group cases, and withholds frequency and scale. Does not report delays for declined claims or by group.
Department for Work and Pensions (UK): "Effectiveness Assessment of Universal Credit Advances Model" (1 April 2025 to 31 March 2026; published 9 July 2026, updated 4 September 2026)
Agency assessment comparing model referrals with a stratified random control sample (reports the model as 2.5 times more effective, without counts) and giving relative referral likelihoods by age band and nationality. On burden it repeats the one-day median delay for approved referred Advances; it gives no delay distribution, no delays by group and no figures for withdrawn claims.
Lighthouse Reports, "Suspicion Machines" (6 Mar 2023)
Journalistic audit of Rotterdam's welfare fraud model, using code and a model file obtained through freedom-of-information requests and the threat of legal action; reports that the model discriminates by ethnicity, age, gender and parenthood. Burden on flagged people is described through individual cases; it is not measured.
Algorithm Audit, "Risk Profiling for Social Welfare Re-examination," Algoprudence problem statement AA:2023:02:P (May 2023)
Ethical review of Rotterdam's risk-based selection for re-examination; describes re-examination as stressful and document-heavy and includes a recipient testimony, but does not quantify burden.
Algorithm Audit, "Addendum Vooringenomenheid voorkomen" (Addendum: Preventing Bias), report for DUO on the control process for the uitwonendenbeurs, May 2024 (supplements the March 2024 report)
Opened. Uses CBS-linked aggregate data to follow students through each step of DUO's risk-based checks (risk profile, manual selection, home visit, finding of misuse, objection) by migration background. Students with a non-European background were 2.0 times as likely to be rated high risk, 6.2 times as likely to be selected in 2014, and had 3.0 times the chance of an unnecessary home visit. Random-sample populations from 2014 and 2017 exist but are too small for group comparison in the addendum. Partly covers burden by group for one system (home visits), but reports no delays, time costs, repayments or abandoned grants.
Public Law Project, written evidence FAE0006 to the House of Commons Public Accounts Committee (November 2025)
Opened. Advocacy evidence on DWP's machine-learning fraud tools. Notes that only a one-day median delay has been published for Universal Credit Advances referrals, that DWP has not confirmed the maximum delay, and that the burden of evidence requests is unknown; repeats DWP's relative referral likelihoods by nationality and age. Confirms the data gap the opportunity targets; it does not measure burden.
Gilman, "Code and Consequences: How Fraud Detection Algorithms Criminalize Welfare," Harvard Journal of Law & Technology 39(2) (Spring 2026)
Opened. Legal and normative analysis drawing on Michigan, the Netherlands and Australia; cites aggregate harm figures from other sources (for example Michigan's 93% error rate) but does not measure pre-decision burden on flagged claimants.

Literature check

Stands as first published. No study was found that measures pre-decision burden (payment delay distribution, evidence requests, withdrawn or abandoned claims) for model-selected versus randomly selected or unselected claimants. The closest new item is Algorithm Audit's May 2024 addendum on Dutch student-finance checks, which follows students through each control step by migration background and reports that students with a non-European background had 3.0 times the chance of an unnecessary home visit; it measures one burden (home visits) by group, not delays or abandonment, and says the random-sample population is too small for group comparison. Public Law Project's November 2025 evidence confirms that DWP has published only a one-day median delay for Universal Credit Advances referrals.

Searched: web (extended); web (standard); arxiv.org; ssrn.com; osf.io; academic.oup.com; onlinelibrary.wiley.com; cambridge.org; committees.parliament.uk; duo.nl; zoek.officielebekendmakingen.nl; jolt.law.harvard.edu; Dutch-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

Partly done already: related work covers some of this question (first-pass screen of 12 October 2026).

The DWP transparency record and its 2025–26 effectiveness assessment report that approved referred Advances take a median of one day longer than automatically approved ones, and the assessment compares model referrals with a random control sample and gives relative referral likelihoods by age and nationality; neither gives a delay distribution, burden by group, or figures for withdrawn claims. Journalistic and civil-society work on Rotterdam's welfare risk model documents stressful, repeated investigations qualitatively. No study was found that measures pre-decision burden for model-selected versus randomly selected or unselected claimants.

Search queries used

administrative burden of being flagged by algorithmic fraud risk model welfare claimants delays verification costs empirical study
Universal Credit Advances fraud model referral payment delays claimants evidence DWP fairness analysis
effect of being flagged by algorithmic fraud risk score on welfare claimants payment delay exit randomised audit selection comparison random selection study
regression discontinuity risk score threshold audit selection welfare benefits investigation effect on claimants take-up exit
study experiences of welfare recipients selected by risk profiling algorithm investigation Rotterdam OR Denmark OR Netherlands interviews burden before decision

Literature check, 13 October 2026:

effect of fraud investigation or benefit review on claimants exit take-up administrative data random audit welfare recipients
algorithmic risk-based verification housing benefit council tax reduction claimants delays evidence requests outcomes evaluation
welfare fraud risk profiling flagged recipients consequences quantitative [allowed_domains: arxiv.org, osf.io, ssrn.com, papers.ssrn.com]
administrative burden welfare verification checks claimants learning costs compliance costs algorithmic selection study [allowed_domains: academic.oup.com, onlinelibrary.wiley.com, journals.sagepub.com, cambridge.org, tandfonline.com]
DUO controle uitwonendenbeurs aselecte steekproef resultaten na stopzetten algoritme gevolgen studenten onderzoek
Universal Credit Advances model referrals delay claimants Work and Pensions Committee OR "Public Accounts Committee" OR NAO 2026 fairness analysis
Algorithm Audit DUO "controleproces uitwonendenbeurs" rapport 2024 migratieachtergrond hoorzitting huisbezoek
Gilman "Code and Consequences: How Fraud Detection Algorithms Criminalize Welfare"

Provenance

RO-LD05-03 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: Literature check added: a second, deeper search for existing work, with its queries, new nearest work and limits. The question and verdict stand. See v1 as published.

All research opportunities