A letter arrives about your benefit claim. Perhaps a payment is delayed, an extra check is needed, or the answer is no. The letter will tell you what was decided. It will rarely tell you whether software helped decide who got a closer look. This guide explains how that happens, how to ask what shaped your case, and the deadlines that protect you. It is general information, not legal advice.
What one government chose to publish
The UK’s Department for Work and Pensions (DWP) offers advances to people waiting for their first Universal Credit payment: 1.3 million of them in 2025 to 2026, worth £700 million. Before paying, it runs requests through a machine learning model that flags those it considers high risk for fraud. Flagged requests go to a staff member for review. The department says there is “always a human intervention and decision,” with no automated decision by the model (DWP, 2025b).
Two details are worth noticing. The staff member reviewing a flagged request is not shown a risk score and is not told the model generated the referral. And for flagged requests later approved, the median payment came one day later than for those approved automatically (DWP, 2025b). The department’s own fairness assessment found notable differences in how often the model referred people of different ages and nationalities. It concluded there were “minimal concerns” of discrimination and said the model would be retrained (DWP, 2025a).
DWP deserves credit for publishing this. Most benefit systems publish nothing comparable. But notice what a claimant would see: a decision made by a person, with no mention of the model. The department says these decisions can be appealed as standard, with no separate process for the model (DWP, 2025b). That makes the ordinary appeal route, and your right to ask for your own information, the tools that matter.
Key idea: Triage is a decision too
Many public-sector AI tools do not say yes or no. They sort cases: which claims get paid quickly and which get checked. That sorting is called triage. A person still makes the final decision, so the system can truthfully say no decision is automated. But being picked for checks can mean delay, extra questions and stress, and if the sorting is skewed, some groups carry more of that burden. When you ask what decided your case, ask also what decided how your case was handled.
Your rights, depending on where you live
- United Kingdom. For most benefit decisions you first ask for a mandatory reconsideration, usually within one month of the decision; you can ask later with a good reason (GOV.UK, n.d.-b). After that, you can appeal to an independent tribunal within one month of the reconsideration decision (GOV.UK, n.d.-a). Separately, you can make a subject access request for the information held about you, usually answered within one month (Information Commissioner’s Office [ICO], n.d.). Where a significant decision is made solely by automated means, UK law adds safeguards including human review and a way to contest it (ICO, 2026).
- United States. If a state Medicaid agency acts on your case, its notice must give “a clear statement of the specific reasons” and the rules it relied on, and explain your right to a hearing (Content of notice, 2026). You have up to 90 days, depending on your state, to request a hearing, and the right to examine your case file and every document the agency will use (Request for hearing, 2026; Procedural rights, 2026). For SNAP food assistance, you can request a hearing on actions in the prior 90 days and examine the documents to be used (Fair hearings, 2026).
- European Union. Data protection law gives you a right to meaningful information about the logic of automated decisions. In a 2025 credit-scoring case, the EU’s top court said this means explaining the “procedure and principles actually applied” so you can understand which of your data were used and how; simply handing over an algorithm is not enough (Court of Justice of the European Union [CJEU], 2025). The AI Act lists systems that assess eligibility for public benefits as high-risk and gives affected people a right to clear explanations of the AI’s role in a decision (Regulation (EU) 2024/1689, art. 86, Annex III). According to K&L Gates, the main high-risk rules now apply from December 2027 (Pearce & Nietsch, 2026).
- Canada (federal programs). Federal institutions using automated decision systems must give notice that a decision is made or assisted by one, provide a meaningful explanation of how and why, inform you of recourse, and publish an algorithmic impact assessment (Treasury Board of Canada Secretariat, 2025). Provincial programs follow their own rules.
- Australia. For Centrelink decisions you can ask for a free explanation at any time, or a free formal review by an independent Authorised Review Officer. For most claim decisions, apply within 13 weeks to protect back pay. If you still disagree, you can go to the Administrative Review Tribunal (Services Australia, n.d.). From December 10, 2026, organisations covered by the Privacy Act must describe in their privacy policies the kinds of significant decisions made or substantially informed by computer programs (Office of the Australian Information Commissioner, 2026).
Five questions you can ask
- “What were the specific reasons for this decision, and which rules did you apply?”
- “Was any automated tool or risk model used in deciding my claim, or in selecting it for checks?”
- “Can I have a copy of my case file and all the information you hold about me?”
- “Did a person review my case, and what did they see?”
- “What is my deadline to ask for a review or appeal?” Ask this first, and diary it.
A sentence you can adapt for a written request: “Please send me a copy of the personal information you hold about me, including any risk scores, flags or outputs of automated tools used in my case, and an explanation of how they were used in the decision dated [date].” Keep a copy, and do not let a request for information delay your appeal.
Where to ask for review or complain
- UK: mandatory reconsideration, then appeal to a tribunal; data complaints to the ICO; free help from Citizens Advice.
- US: follow the hearing instructions on your notice from your state agency, and act within the deadline it gives.
- Canada: use the recourse route in your decision notice; privacy concerns about federal institutions go to the Privacy Commissioner of Canada.
- Australia: explanations and formal reviews from Services Australia; complaints about Australian Government agencies to the Commonwealth Ombudsman.
- EU: the review route in your decision, and your national data protection authority.
For agencies. If a model sorts claims in your service, could a reviewer reconstruct, for one claimant, what the model flagged and what the caseworker saw? Our free readiness assessment tests one workflow against eight control domains and points to your weakest one. The free book Operational AI Readiness goes deeper: Chapter 6 on the evidence spine, Chapter 9 on approval gates for high-impact actions, and Chapter 11 on what auditors and reviewers will ask to see. Read it free.
What the record does not show
DWP’s record does not say whether claimants are told the model was used, and much of it, including how often the model runs, is withheld for fraud-prevention reasons. Its fairness analysis measured age as the only protected characteristic. We do not know how many other benefit systems use similar tools, because most publish nothing. The EU ruling we cite concerned a private credit check, so how far it reaches public agencies will depend on the case. Deadlines differ by program and state. Always check the letter you received.
Follow what happens next. We will keep reading what governments publish about the systems behind benefit decisions, and what they leave out. The free Mercury Brief brings it to you with the record behind it. Get the free Mercury Brief.
FairByDesign is reader-funded. If this helped, get the free Mercury Brief or see ways to support. Support never buys editorial control.
References
Bakos, V. (2026). Operational AI readiness: A practitioner guide to governing AI that is already live. FairByDesign. https://fairbydesign.org/operational-ai-readiness/
Content of notice, 42 C.F.R. § 431.210 (2026). https://www.ecfr.gov/current/title-42/section-431.210
Court of Justice of the European Union. (2025, February 27). Automated credit assessment: The data subject is entitled to an explanation as to how the decision was taken in respect of him or her (Press release No 22/25, Case C-203/22, Dun & Bradstreet Austria). https://curia.europa.eu/jcms/upload/docs/application/pdf/2025-02/cp250022en.pdf
Department for Work and Pensions. (2025a, July 17). Universal Credit Advances model: Fairness assessment, 1 April 2024 to 31 March 2025. GOV.UK. https://www.gov.uk/government/publications/fairness-assessment-including-statistical-analysis-of-the-universal-credit-advances-machine-learning-model-1-april-2024-to-31-march-2025/universal-credit-advances-model-fairness-assessment
Department for Work and Pensions. (2025b, December 16). DWP: Universal Credit Advances model [Algorithmic transparency record]. GOV.UK. https://www.gov.uk/algorithmic-transparency-records/dwp-universal-credit-advances-model
Fair hearings, 7 C.F.R. § 273.15 (2026). https://www.ecfr.gov/current/title-7/section-273.15
GOV.UK. (n.d.-a). Appeal a benefit decision. https://www.gov.uk/appeal-benefit-decision
GOV.UK. (n.d.-b). Challenging a benefit decision (mandatory reconsideration). https://www.gov.uk/mandatory-reconsideration
Information Commissioner’s Office. (2026, March 31). What are the ADM safeguards? [Draft guidance]. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/automated-decision-making/what-are-the-adm-safeguards/
Information Commissioner’s Office. (n.d.). Getting copies of your information (SAR). https://ico.org.uk/for-the-public/getting-copies-of-your-information-subject-access-request/
Office of the Australian Information Commissioner. (2026, September 30). New resources on transparency for use of AI and automated decision-making. https://www.oaic.gov.au/news/media-centre/new-resources-on-transparency-for-use-of-ai-and-automated-decision-making
Pearce, S., & Nietsch, T. (2026, July 31). EU digital omnibus on AI enters into force. National Law Review (K&L Gates). https://natlawreview.com/article/eu-digital-omnibus-ai-enters-force
Procedural rights of the applicant or beneficiary, 42 C.F.R. § 431.242 (2026). https://www.ecfr.gov/current/title-42/section-431.242
Regulation (EU) 2024/1689 (Artificial Intelligence Act), art. 86 and annex III. European Commission AI Act Service Desk. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-86; https://ai-act-service-desk.ec.europa.eu/en/ai-act/annex-3
Request for hearing, 42 C.F.R. § 431.221 (2026). https://www.ecfr.gov/current/title-42/section-431.221
Services Australia. (n.d.). Explanations and formal reviews of a Centrelink decision. https://www.servicesaustralia.gov.au/explanations-and-formal-reviews-centrelink-decision
Treasury Board of Canada Secretariat. (2025). Directive on Automated Decision-Making (modified June 24, 2025). https://www.tbs-sct.canada.ca/pol/doc-eng.aspx?id=32592
Leave a Reply
You must be logged in to post a comment.