BRENNEN

Industries / Government, education and community / Government

Benefits and payments

Deciding who qualifies for public money such as pensions, food aid and unemployment pay, and getting the right amount to the right person on time.

7

Milestones achieved, from 1909 to 2026.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1909

    Pensions paid from taxes

    On New Year's Day 1909 the first British old age pension paid for out of general taxation became payable: up to 5 shillings a week from age 70, means-tested, for people who had lived in the UK for 20 years. Applicants filled in a form at the Post Office, many needed help with it, and some Post Offices ran out of forms. The Pensions Archive Trust says well over 600,000 people were expected to qualify.

    pensionsarchive.org.uk

    Credit

    UK government

  2. 1984

    Food aid on a card

    The first Electronic Benefit Transfer (EBT) pilot, which put food stamp benefits on a debit card instead of paper coupons, began in Reading, Pennsylvania, in 1984. A 1996 law required every state to use EBT before October 2002, and by July 2004 all 50 states, the District of Columbia, the Virgin Islands and Guam issued benefits this way. The US Department of Agriculture says trafficking, mainly selling benefits for cash, fell from nearly 4% in the 1990s to around 1% once EBT was in full use, because every purchase left an electronic record.

    fns.usda.gov

    Credit

    US Department of Agriculture, Food and Nutrition Service

  3. 2013

    Payments straight to accounts

    India started Direct Benefit Transfer (DBT) on 1 January 2013 in 43 districts, paying welfare money directly into recipients' bank accounts to target it better and cut duplicate and false claims. It was extended across the country on 12 December 2014. Aadhaar, the national identity number, is preferred but not required. The government's DBT site lists 319 schemes and cumulative transfers of ₹54.3 lakh crore (₹54.3 trillion), its own figure, with no date given for the total.

    dbtbharat.gov.indbtbharat.gov.in

    Credit

    Government of India, DBT Mission

  4. 2015

    Debts by averaging

    From July 2015 to November 2019 Australia's welfare agency ran an automated system, known as Robodebt, that spread a person's yearly income from tax office records evenly across fortnights and raised a debt wherever that average was more than the person had reported. The Commonwealth later admitted it had no proper legal basis for those debts. Approving a A$112 million settlement in June 2021, Justice Murphy of the Federal Court found it had asserted at least A$1.763 billion in debts against about 433,000 people and recovered about A$751 million, and called the scheme "a massive failure of public administration".

    LessonAn average of a year's income says little about any one fortnight, and a debt needs a legal basis before the state collects it.

    judgments.fedcourt.gov.au

    Credit

    Federal Court of Australia, Prygodicz v Commonwealth of Australia (No 2)

  5. 2020

    Pandemic jobless fraud measured

    When US unemployment insurance paid millions of new claimants during the COVID-19 pandemic, fraud was likely between $100 billion and $135 billion, about 11% to 15% of all benefits paid from April 2020 to May 2023, according to a September 2023 estimate by the US Government Accountability Office (GAO).

    LessonChecks on who is claiming have to grow as fast as the number of claims.

    gao.gov

    Credit

    GAO

  6. 2021

    Nationality in a risk score

    In 2013 the Dutch Tax and Customs Administration, which pays childcare benefits, began using a self-learning algorithm to flag applications for possible fraud. Amnesty International found that non-Dutch nationals received higher risk scores, that the civil servants reviewing flagged cases were given no explanation for a score, and that tens of thousands of parents and caregivers, mostly from low-income families, were falsely accused of fraud. The scandal brought down the Dutch government in January 2021.

    LessonA risk score that nobody can explain, fed with inputs such as nationality, can turn checking into discrimination.

    amnesty.org

    Credit

    Amnesty International, for the analysis

  7. 2026

    Models tested on SNAP

    Vals AI, a testing firm, working with Code for America and the Center for Civic Futures, scored AI models on questions about SNAP, the US food aid program once called food stamps. It wrote 459 scenario questions covering all 50 states, Guam and the US Virgin Islands, from applying to renewing, reporting changes and overpayments, and published results on a 230-question test set. In its June 2026 study the best of 12 models scored 68.1%; on the leaderboard updated 12 October 2026 the top score was 76.93%, for Claude Opus 5. Vals concluded that no model is reliable enough to answer SNAP questions unsupervised, and that local procedures and recent rule changes were hardest.

    vals.ai

    Credit

    Vals AI (Meghana Kotcherlakota, Omar Almatov, Rayan Krishnan)

Watch for

  1. A policy assistant for caseworkers

    Code for America and Anthropic announced the SNAP Policy Navigator, built on Anthropic's Claude, which answers a caseworker's policy question in plain language with cited sources drawn from federal regulations, state manuals and policy directives. Code for America said the user "gets clarity on policy, not a decision on overall eligibility." The partners plan further tools to review eligibility documents and draft letters to recipients. No pilot states or results were named.

    Achieved whena state agency uses it in routine casework and an outside evaluation reports the accuracy of its answers or its effect on processing times.

    govexec.com

    Claimed by

    Code for America and Anthropic, May 2026

  2. Notices drafted with AI

    Public Policy Lab, a nonprofit design lab, says it is building an AI tool that drafts SNAP notice templates, the letters that tell people a decision or ask them for something, to meet the rules and be accessible, with staff approving each draft. It is planning the rollout with New Mexico and Oregon. Its case for the work: Arizona alone has 171 different SNAP notices and sent over 5 million of them from June 2024 to May 2025.

    Achieved whena state sends the new notices and an outside evaluation measures whether recipients understand them better or whether errors fall.

    publicpolicydesign.blog.gov.uk

    Claimed by

    Public Policy Lab, January 2026

Open

  1. 01

    Made-up proof, with answers

    A public set of made-up documents of the kind applicants send to prove income and costs, such as pay stubs, employer letters, benefit award letters, rent receipts and utility bills, in many layouts, some handwritten or photographed. Each comes with an answer key giving the amounts a caseworker should count. Then score AI document readers on it.

    The brief
    Credit · unclaimed
  2. 02

    Overpayment checks with known answers

    A public set of made-up claimant histories, with pay that changes from week to week, for which the correct overpayment, often zero, has been worked out by hand under one country's published rules. Any automated debt system, or any AI tool, can then be run on the same cases and scored.

    The brief
    Credit · unclaimed

Open 01 · the brief

Made-up proof, with answers

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of made-up documents of the kind applicants send to prove income and costs, such as pay stubs, employer letters, benefit award letters, rent receipts and utility bills, in many layouts, some handwritten or photographed. Each comes with an answer key giving the amounts a caseworker should count. Then score AI document readers on it.

Why it is still open

Code for America and Anthropic named document review as a planned tool in May 2026, not a shipped one. The main published test of AI on SNAP, Vals AI's Public Benefits Bench, asks policy questions rather than reading documents, and its question set is proprietary. I found no public test set of benefit verification documents; what exists for pay stubs is fraud detection sold to lenders and landlords.

govexec.comvals.ai

Built with what exists

a script that prints made-up documents in varied layouts and writes the answer key from the numbers it printed; a phone camera for photographed copies; AI models that read PDFs and photos, such as Claude, Gemini and GPT; Tesseract, an open-source text reader, as a baseline. (see the Note "Seven more flagged")

  1. First step

    Generate 50 made-up pay stubs in five layouts with known gross pay and pay dates, photograph ten of them, and score one AI model on all 50.

  2. Done looks like

    1,000 documents of 20 types published with answer keys, and at least four tools scored on the share of documents where every counted amount matches the key and on how often each tool says it cannot read a document instead of guessing.

Open 02 · the brief

Overpayment checks with known answers

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of made-up claimant histories, with pay that changes from week to week, for which the correct overpayment, often zero, has been worked out by hand under one country's published rules. Any automated debt system, or any AI tool, can then be run on the same cases and scored.

Why it is still open

Robodebt's debts came from averaging yearly income across fortnights, a method that Australia's Administrative Appeals Tribunal had rejected in about 29 decisions by May 2017, according to the Federal Court. Peter Whiteford of the Australian National University, analyzing data given to the Royal Commission, found that "almost nobody who received income had completely stable earnings" while Robodebt ran. In Michigan an automated unemployment fraud system wrongly accused about 40,000 people between 2013 and 2015. I found no public set of test cases for checking such calculations.

reporter.anu.edu.aujudgments.fedcourt.gov.austpp.fordschool.umich.edu

Built with what exists

the fortnightly income test in Australia's Social Security Act 1991, or the earnings rules of one US state's unemployment program; a spreadsheet or Python for the worked answers; AI models such as Claude, Gemini and GPT to test against the answers.

  1. First step

    Write 20 histories from the patterns the court and the ANU describe, such as casual work, work for part of a year and a single large payment, and work out each answer by hand.

  2. Done looks like

    300 histories, each answer agreed by two people working separately, published with the rules they follow, and the yearly-averaging method and at least three AI models scored on the share of debts they get right and the number of false debts they raise.