Domain Atlas / Public benefits & eligibility
Serbia Social Card (Socijalna karta)
Serbia's Social Card (Socijalna karta) registry, given a statutory basis by the Law on the Social Card in force from 1 March 2022 and financed in part by an 82.6 million euro World Bank public-sector loan, cross-links roughly 130 to 135 categories of data from other state registers to verify social-assistance eligibility and flag suspected undeclared income or assets. After the law, named sources report the caseload falling by a range of tens of thousands: government figures cited by Amnesty International show about 35,000 fewer recipients by August 2023, A11 counts at least 44,000 people having lost assistance by early 2024, and the UN Working Group on Business and Human Rights reported over 60,000 without assistance by October 2025. These are largely net caseload declines rather than audited counts of system-caused removals, and the government attributes part of the fall to a stronger economy. Roma are reported among the most affected because informal earnings are misclassified as income, but the registry records no ethnicity, so this is inferred rather than officially disaggregated. As of the latest reporting, Constitutional Court, World Bank Inspection Panel, and UN scrutiny were pending or active, with no court or panel yet ordering changes.[6]
What happened
Serbia's Social Card (Socijalna karta) registry was given a statutory basis by the Law on the Social Card (Zakon o socijalnoj karti), adopted in 2021 and in force from 1 March 2022. Operated by the Ministry of Labour, Employment, Veterans and Social Affairs, the registry cross-links roughly 130 to 135 categories of personal data pulled from other state registers — the Tax Administration, the Ministry of Interior's vehicle records, and property, pension and banking data — covering each social-assistance beneficiary plus relatives and so-called connected persons, to verify eligibility and detect undeclared income or assets. Its establishment was financed as part of an 82.6 million euro World Bank public-sector loan, and the Bank provided technical guidance. The system is semi-automated: it flags discrepancies and sends notifications to social workers at municipal Centers for Social Work, who are meant to investigate before benefits are reduced, suspended, or removed. Amnesty International documented automation bias, with workers telling beneficiaries there was nothing they could do because the new system had decided. The underlying rules are undisclosed, and there is no evidence in the public record that the system uses machine learning; it is best described as rules-based cross-registry data-matching rather than a predictive model.
After the law, named sources report Serbia's financial social-assistance caseload falling by a range of tens of thousands, though the figures are method-dependent and move over time. A11 — Initiative for Economic and Social Rights, reading government data, records 211,266 recipients in February 2022, 182,773 at the end of 2022, and 168,307 by 1 February 2024, with at least 44,000 people having lost assistance since the law took effect. Government figures cited by Amnesty show about 176,000 recipients in August 2023, roughly 35,000 fewer than in March 2022; earlier Thomson Reuters Foundation reporting put removals at about 27,000, some 15 percent of recipients. By October 2025 the UN Working Group on Business and Human Rights, after a country visit from 6 to 15 October 2025, reported that implementation of the law had left over 60,000 citizens without social assistance, with its full report due to the Human Rights Council in June 2026. These are largely net caseload declines rather than audited counts of system-caused removals, and the Serbian government has attributed part of the fall to a stronger economy, so causation is contested.
Documented misclassification errors show how the match collapses informal, cash-based livelihoods into sparse data points: a one-off funeral donation of about 170 euro flagged as disqualifying income; two cars sold for scrap years earlier but still registered as assets; a recyclable-materials collector recorded with income implying an implausible monthly haul; and a single mother whose seasonal jobs and vocational grant were read as regular income. Removed beneficiaries have 15 days to appeal and must wait three months to reapply regardless of changed circumstances, and removal letters frequently reference only unspecified data from the electronic database. Over roughly two years the Ministry processed more than 100,000 notifications of suspected income or asset increases, while beneficiaries filed only 361 appeals against Centers for Social Work rulings; because the two figures cover different populations, the gap illustrates how rarely flags were contested rather than a measured appeal rate. Roma are reported among the most affected, because their informal earnings are misclassified as income; the registry records no ethnicity, so Roma-specific counts are inferred from benefit reliance — a 2019 UNICEF estimate, cited in the reporting, put the share of Roma settlement households receiving some benefit at 84 percent — and from case documentation rather than official disaggregation.
The system is operational and heavily contested, but no court or panel has yet ordered changes. A11 filed a constitutional petition in April 2022, and an ESCR-Net coalition submitted a joint legal opinion to the Constitutional Court of Serbia in late November 2022, arguing the law violates rights to social security, equality and non-discrimination, privacy, due process, and remedy; the case remained pending. A11 filed a Request for Inspection with the World Bank Inspection Panel on 31 March 2024 on behalf of six Requesters including Roma community members, alleging roughly 44,000 people were negatively and disproportionately affected and that, because the system is semi-automated, social workers cannot correct errors recorded in it. The Panel registered the request on 24 August 2024 and submitted its Report and Recommendation to the Board on 12 November 2024; as of 30 June 2025 the Board's decision on whether to authorize an investigation was still pending. A11 has called for the law to be repealed; Amnesty International's stated ask is that the system be suspended and made rights-compliant if its risks cannot be prevented.
The sociotechnical reading
The Social Card is the Atlas's legibility case: a system that does not predict but MATCHES, and whose harm is the collapse of a complex, informal, cash-based reality into the sparse and often stale fields the state happens to hold. A social worker could once understand a funeral gift, a car scrapped years ago, or a season of odd jobs; the registry reads each as durable income or an asset, and a matched discrepancy flows into a benefit cut. The lesson the map draws is that when eligibility is decided by cross-registry data-matching, the quality of the data model — what it captures and how stale it is — matters more than the accuracy of any single flag, and the people whose lives are least legible to the registers, the poorest and Roma especially, are the ones it fails.
What sets this case apart from its neighbours is where the correction is supposed to live, and that it does not. This is not MiDAS, where the operator was removed; social workers are still in the loop. It is not Rotterdam, where an audit pried a live model open and measured its bias; there is no measured error rate here at all. And it is pointedly not SyRI or Robodebt, where a court or a Royal Commission finally bound the system. Serbia's Social Card has a statute behind it and every accountability channel activated at once — a Constitutional Court petition, a World Bank Inspection Panel request, and a UN Working Group — yet as of the latest reporting none has bound it, and the caseload has fallen by a moving range (about 35,000, then at least 44,000, then over 60,000) that is a net decline rather than an audited count of who the system wrongly removed. The map's defining feature is an absence: A11's Inspection Panel request alleges that, because the system is semi-automated, social workers cannot correct errors recorded in it, and around that sit a 15-day appeal window, a mandatory three-month reapplication lockout, and only 361 appeals against more than 100,000 flags. Correction is broken structurally, not by individual neglect, and a wrong flag re-enters the next match because nothing pulls it back out.
Two further threads run through this case and no other in the Atlas. First, the leverage is on the input and record sides, not the model: the Lab's sharpest levers here narrow which registers feed the match, reconcile a flag against its real-world source before it drives a cut, and fund a correction channel a person can actually reach — a better classifier changes nothing, because there is no evidence of a classifier to improve. Second, the World Bank financing opens an international-accountability route absent from purely domestic systems, which is why a development-finance body's own Inspection Panel became one of the venues where this system is being contested. The governance question the Social Card poses is therefore live, not historical: on what published, reconcilable basis may a state cross-link everyone's records and cut off the poorest — and who can correct it when the match is wrong?
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.