The sources we found most useful for working out how governments use algorithms: how to find them, what to ask for and what to do when an agency says no.
22 June 2026 · guide · key source
A framework that splits AI into four stages (data, compute, models, applications) and shows what kind of story each stage can yield, with worked examples from several countries.
2026 · video
A masterclass video on spotting government processes that may rely on automation, such as benefits decisions or risk assessments, requesting documents and data, filing targeted FOI requests, and building sources inside institutions.
2024-2025 · website toolkit
Short plain-language guides on algorithmic accountability reporting: what to ask, who to ask, and seven common framing pitfalls. A second set covers platform gatekeeping such as shadow bans and includes a do-it-yourself algorithm audit.
curriculum (videos and slides)
Free course modules with videos and downloadable slide decks in three tracks: an introduction to AI reporting, an intensive for reporters covering AI, and an editor's guide. Includes lessons on investigating bias and on public records requests.
December 2025 · toolkit · key source
A how-to guide drawing on Amnesty's investigations in Denmark, Sweden, Serbia, France, the Netherlands, the UK and elsewhere. It covers scoping a system, FOI requests with templates, testing a system empirically, and routes to accountability such as litigation and regulators.
Investigations that exposed a government system, with the evidence they used and how they got it.
NL · 2023
Rotterdam ranked every welfare recipient by fraud risk with a machine-learning model, and journalists did something that almost never happens: they got hold of the model itself and took it apart. They found it scored parents, young people, women and people with weaker Dutch higher, and that it was only about 50 percent more accurate than random selection. The city had already put it on hold, and later agreed the Dutch-language variables were an unacceptable proxy for migrant background.
SE · 2024
In Sweden the Social Insurance Agency refused almost everything for three years, and the investigation went ahead anyway by going through the agency's overseer. It showed that a risk model for parental benefit selected women more than 1.5 times as often as men, and people with a foreign background about 2.5 times as often. Amnesty International called for the system to be discontinued, and in 2025 the agency took it out of use.
NL · 2025
Amsterdam did nearly everything the responsible-AI playbook asks: it opened its records, tested for bias, consulted citizens and built a model that could be explained. The journalists still found it unfair. In the live pilot the bias reversed, with Dutch applicants and women wrongly flagged more often, and the model did no better than caseworkers. The city stopped the pilot in November 2023.
FR · 2023
In France, journalists used access-to-information law to get the scoring code of the national family benefits fund. The model gives every household on the benefit rolls a risk score, and the authors say it pushes scores up for single parents, low-income households and people with disabilities. The disclosure fed a court case that is still pending, and in January 2026 the fund published the source code of its current model.
NL · 2023
Reporters could not get into the Dutch student-finance agency's risk system, so they asked lawyers for their case counts instead. Of 376 objection and appeal cases the 32 lawyers had handled since 2012, 367 involved students with a migration or bicultural background. The education minister suspended algorithm-based selection in 2023, the data protection authority later found the checks unlawfully discriminatory, and the state accepted liability for the damage.
NL
The Dutch tax authority scored childcare benefit applications with a self-learning risk model, and Dutch citizenship was one of its inputs. Tens of thousands of parents were wrongly treated as fraudsters, and migrant and low-income families were hit hardest. Years of reporting by Trouw and RTL Nieuws, pressure from MPs and a data protection investigation established the use of nationality in July 2020. The government resigned in January 2021, and the tax authority was fined €2.75 million. The model was one part of a wider failure.
NL
SyRI let Dutch government bodies combine personal data to predict who might commit benefits or tax fraud, and it was used mainly in poorer neighbourhoods. Almost everything about how it worked was secret. A coalition of civil-society groups and two writers, later joined by the FNV union, took the state to court. In February 2020 a court in The Hague ruled that the law behind it breached the right to private life, because its use could not be understood or checked. The government did not appeal.
DK · 2024
Denmark's benefits agency runs about 60 fraud models, and it refused Amnesty's requests for a joint audit and full access to code and data. With redacted documents on four models and dozens of interviews, Amnesty still raised serious discrimination concerns: one model scores 'foreign affiliation', ties to countries outside the EEA, to prioritise investigations. It is a good example of building a credible case when code and data are refused. As far as we could find, nothing has changed since.
UK · 2024
A records release forced the UK's Department for Work and Pensions to admit statistically significant disparities in a fraud tool that scores Universal Credit advance requests. Its own later figures show non-UK nationals were referred about 2.27 times as often as UK nationals in 2024/25, and about 3.35 times in 2025/26. DWP says a caseworker decides every flagged case. It published its fairness assessment and committed to retrain the model, which stays in use until a retrained version is tested.
UK
From 2015 the UK Home Office used a 'streaming tool' that rated visa applications red, amber or green, which decided how much scrutiny they got. After a long effort to find out how it worked, the Joint Council for the Welfare of Immigrants (JCWI) and Foxglove brought a judicial review arguing it discriminated on nationality. In August 2020, before a hearing, the Home Office confirmed nationality was one of its factors and dropped the tool pending a redesign, without accepting the claims.
ES · 2024
RisCanvi scores people in Catalan prisons, and its results shape furlough, parole and prison-grade decisions. An outside team audited it without access to its weights, using a small public dataset and interviews. They found it affected 7,713 people in 2022, and that professionals changed the final rating in under 5% of cases. The Justice Department began an update after the audit, which was reported as still underway in March 2026.
ES · 2022
Spain's VioGen scores a woman's risk of being attacked again, and that score determines the level of police protection. An outside audit tested it against a public homicide dataset and interviews with survivors. It found that about 73% of 475 homicide victims had no prior police report, so the system never scored them. In January 2025 the Interior Ministry presented VioGen 2 with recalibrated algorithms and removed the lowest risk level, saying it was a planned update, not a reaction.
ES
BOSCO is the software that checks who qualifies for Spain's bono social, a discount on electricity bills for vulnerable households. Rejected applicants contacted Civio, and their cases showed it was turning away people who were entitled, including widowed pensioners. Civio asked for the source code in 2018 and was refused at every stage until September 2025, when the Supreme Court ruled it had a right to the code. We could not find reporting on whether the code has since been handed over.
PL
From 2014, Polish job centres placed unemployed people into one of three profiles using a scored, computer-based interview, and the profile decided what help they could get. The scoring logic was confidential, and the questions only became public after the Panoptykon Foundation used freedom of information law. Its research and the Human Rights Commissioner's referral took the system to the Constitutional Tribunal, which ruled it unconstitutional in 2018. Profiling ended in 2019.
AT
Austria's public employment service (AMS) built an algorithm, AMAS, to sort jobseekers into groups with high, medium and low chances on the job market. Its development documentation showed that gender and citizenship lowered some people's predicted chances, and that care duties counted only for women. After concerns from researchers, the Ombudsman Board and digital-rights groups, the data protection authority stopped it in August 2020. The courts later sided with the AMS, which had deleted all AMAS data and said it would not revive the system.
RS · 2023
Serbia's Social Card registry, funded with World Bank support, merges data from many government databases to profile people applying for social assistance. Amnesty found that recipients fell by about 35,000 after the law took effect, and says Roma and people with disabilities were hit hardest, partly because of wrong source data. The ministry rejected the findings. The World Bank's Inspection Panel first recommended an investigation, then dropped it after the Bank reported corrective steps such as training and workshops. In December 2025 the Board approved not investigating, which closed the case.
The research behind it: what algorithm registers show and leave out, and how governments are held to account.
Ada Lovelace Institute, AI Now Institute, & Open Government Partnership · Joint report, published 24 August 2021 · 2021
A comparative review of registers, impact assessments, audits and bans, drawing on the experience of the people who implemented them; context and public participation decide whether these tools work.
Cath, C. & Jansen, F. · Techné: Research in Philosophy and Technology 26(3):395-412 · 2022
The Amsterdam and Helsinki registers mostly list uncontentious uses and leave out sensitive systems such as welfare fraud detection, so treating them as accountability risks ethics theatre.
Ananny, M. & Crawford, K. · New Media & Society 20(3):973-989 · 2018
Being able to see a system is not the same as understanding or governing it; the authors set out limits of the transparency ideal and argue for accountability across whole sociotechnical systems.
Das, D. & Guha, S. · arXiv preprint 2609.24883 (submitted 21 September 2026); ACM DOI listed on the arXiv page · 2026
Comparing 8,368 records from 17 instruments covering 72 countries, registers share a descriptive core but rarely ask about appeals, risks, legal bases or external evaluation.
Das, D., Tessono, C., Ahmed, S. I., & Guha, S. · arXiv preprint 2604.15514 (submitted 16 April 2026) · 2026
Across all 409 systems in Canada's federal register, 86% are internal efficiency tools, and the register consistently leaves out human discretion, training and how uncertainty is handled.
