# Sweden's Försäkringskassan fraud-risk ML model flagged women 1.5x and low earners 2.97x as often, then was quietly withdrawn

> Försäkringskassan's machine-learning risk profile for temporary parental benefit failed its own inspectorate's equal-treatment test in 2018, was found by Lighthouse Reports and Svenska Dagbladet to select women more than 1.5 times and people below median income 2.97 times as often, and was taken out of use in 2025 after Sweden's data protection authority IMY opened supervision.

- Verification status: pending
- Case type: post-mortem
- Provider: Försäkringskassan (in-house machine-learning risk profile)
- Client: Försäkringskassan (Swedish Social Insurance Agency), Government: social insurance / welfare administration (named)
- Sector: government / SE / finance
- Canonical URL: https://theinternetninja.com/stories/sweden-s-f-rs-kringskassan-fraud-risk-ml-model-for-parental-benefit-what-the-ins/
- Source: The Internet Ninja (theinternetninja.com), independent verified-proof platform

## Outcomes

| Metric | Before | After |
| --- | --- | --- |
| On 2017 data (5,082 model-selected, 1,047 random), the model selected women more than 1.5x, foreign background close to 2.5x, no university degree more than 3.31x and below median income 2.97x as often as the opposite group (Lighthouse Reports / SvD) |  |  |
| ISF report 2018:5 found the risk-based selection raised hit rates but in its current design did not pass ISF's equal-treatment test |  |  |
| IMY opened supervision on 19 June 2025; the agency said the ML risk profile was taken out of use a little over a month before its 5 September 2025 reply; IMY closed the case on 18 November 2025 |  |  |

## Verification method

Independent public record only, no contact with the agency. Primary documents fetched live on 2026-10-05 and archived: ISF's summary of report 2018:5, IMY's supervision letter (19 June 2025) and closing decision (18 November 2025) in case IMY-2025-11828, Lighthouse Reports' published methodology with its dataset on GitHub, and Amnesty International's statement of 27 November 2024. Handed to the checker.

## FAQ

**Did Försäkringskassan's fraud algorithm discriminate?**

On 2017 temporary parental benefit data, Lighthouse Reports and Svenska Dagbladet found the model selected women more than 1.5 times, people with a foreign background close to 2.5 times, people without a degree more than 3.31 times and people below median income 2.97 times as often as the opposite group. Sweden's inspectorate ISF found in 2018 that it did not pass ISF's equal-treatment test, but took no position on whether that was discrimination in the legal sense.

**Is the Sweden welfare fraud algorithm still in use?**

Försäkringskassan told the data protection authority IMY on 5 September 2025 that its machine-learning risk profile for temporary parental benefit had been taken out of use a little over a month earlier, with no current plans to reinstate it. IMY closed its case on 18 November 2025 without a finding on lawfulness.

**How was temporary parental benefit risk scoring tested?**

Lighthouse Reports obtained the dataset behind ISF's 2018 review, 6,129 people selected for investigation in 2017 (1,047 at random, 5,082 by the model), and tested it against six statistical fairness definitions with review from eight academics.

## Full case file

## The problem

Försäkringskassan, Sweden's Social Insurance Agency, pays temporary parental benefit (tillfällig föräldrapenning) to parents who stay home with a sick child, and it used a machine-learning model to decide which claims to investigate ([source](https://www.lighthousereports.com/methodology/sweden-ai-methodology/)). Sweden's Social Insurance Inspectorate (ISF) reported in 2018 that the agency "har kommit längst i arbetet med riskbaserade urval inom tillfällig föräldrapenning", that is, temporary parental benefit was where its risk-based selection was most advanced ([source](https://isf.se/download/18.6e75aae16a591304891f1e8/1565330423683/Profilering%20som%20urvalsmetod%20f%C3%B6r%20riktade%20kontroller-ISF-sammanfattning-2018-05.pdf)). Amnesty International states the system "has been used by the Swedish Social Insurance Agency since at least 2013" ([source](https://www.amnesty.org/en/latest/news/2024/11/sweden-authorities-must-discontinue-discriminatory-ai-systems-used-by-welfare-agency/)). The people most flagged had no way to know: Lighthouse Reports writes that recipients "have no idea that they have been flagged by an algorithm" ([source](https://www.lighthousereports.com/investigation/swedens-suspicion-machine/)).

## What was built

The agency built a risk-scoring model that selected the highest-risk applications for control, alongside a random sample it also checked ([source](https://www.lighthousereports.com/methodology/sweden-ai-methodology/)). ISF describes the infrastructure behind it, RaKuR, as "en teknisk infrastruktur för att administrera och köra urvalsprofiler, och distribuera urvalet till handläggningssystemet", a technical platform to run selection profiles and push the selection into casework ([source](https://isf.se/download/18.6e75aae16a591304891f1e8/1565330423683/Profilering%20som%20urvalsmetod%20f%C3%B6r%20riktade%20kontroller-ISF-sammanfattning-2018-05.pdf)). Per Amnesty, "people with the highest risk scores as designated by the algorithm have been automatically subject to investigations by fraud controllers within the welfare agency, under an assumption of 'criminal intent' right from the start" ([source](https://www.amnesty.org/en/latest/news/2024/11/sweden-authorities-must-discontinue-discriminatory-ai-systems-used-by-welfare-agency/)). The agency refused to disclose the model's code or input variables, so it remains a black box to every outside reviewer ([source](https://www.lighthousereports.com/methodology/sweden-ai-methodology/)).

## The outcome

**The inspectorate's verdict (2018).** ISF found that the agency's risk-based selection "ökar träffsäkerheten i uppföljningarna, men att de i sin nuvarande utformning inte klarar ISF:s test för likabehandling": it raised the hit rate, but in its current design did not pass ISF's equal-treatment test ([source](https://isf.se/download/18.6e75aae16a591304891f1e8/1565330423683/Profilering%20som%20urvalsmetod%20f%C3%B6r%20riktade%20kontroller-ISF-sammanfattning-2018-05.pdf)). ISF was explicit that it "tar inte ställning till om den bristande likabehandlingen utgör diskriminering i diskrimineringslagens mening", so it made no legal finding of discrimination ([source](https://isf.se/download/18.6e75aae16a591304891f1e8/1565330423683/Profilering%20som%20urvalsmetod%20f%C3%B6r%20riktade%20kontroller-ISF-sammanfattning-2018-05.pdf)). The two sides conflict and this page shows both: ISF also judged the hit rate "betydligt högre" than the alternatives and said equal treatment could be guaranteed "med små korrigeringar", with small corrections ([source](https://isf.se/download/18.6e75aae16a591304891f1e8/1565330423683/Profilering%20som%20urvalsmetod%20f%C3%B6r%20riktade%20kontroller-ISF-sammanfattning-2018-05.pdf)), while the agency "rejected the ISF's conclusions and questioned the validity of their analysis" ([source](https://www.lighthousereports.com/methodology/sweden-ai-methodology/)).

**The independent measurement (2024).** Lighthouse Reports and Svenska Dagbladet obtained the dataset behind the ISF review: "6,129 people that were selected for investigation in 2017", of which "1,047 were randomly selected and 5,082 were selected by the machine learning model" ([source](https://www.lighthousereports.com/methodology/sweden-ai-methodology/)). On selection rates they found women "more than <span class="kpi">1.5</span> times more likely to be selected by the algorithm than men", people with a foreign background "close to <span class="kpi">2.5</span> times more likely", people without a university degree "more than <span class="kpi">3.31</span> times more likely", and people below the median income "<span class="kpi">2.97</span> times more likely to be selected" ([source](https://www.lighthousereports.com/methodology/sweden-ai-methodology/)). Among people who had made no mistake, a woman was "more than <span class="kpi">1.7</span> times more likely to be wrongly flagged than a man" and a person of foreign background "<span class="kpi">2.4</span> times more likely" than a person with a Swedish background ([source](https://www.lighthousereports.com/methodology/sweden-ai-methodology/)). The weakest load-bearing point is right here: these ratios are one newsroom's analysis of 2017 data, not a regulator's finding, and Lighthouse itself says "it is possible that the discriminatory patterns we outline in this methodology have changed" ([source](https://www.lighthousereports.com/methodology/sweden-ai-methodology/)). Amnesty, which corroborates the affected groups, says it reviewed that same analysis, so it is not a second measurement ([source](https://www.amnesty.org/en/latest/news/2024/11/sweden-authorities-must-discontinue-discriminatory-ai-systems-used-by-welfare-agency/)). The agency's reply, per Lighthouse, "did not refute the findings or design of our analysis" but argued a human investigator always makes the final decision ([source](https://www.lighthousereports.com/methodology/sweden-ai-methodology/)).

**The justification under strain.** The agency estimates it wrongly pays out "SEK <span class="kpi">1.3 billion</span>" a year in temporary parental benefit and attributes "SEK <span class="kpi">800 million</span>" of that to intentional fraud, according to Lighthouse, which relays these agency figures rather than this page fetching them from the agency's annual report ([source](https://www.lighthousereports.com/methodology/sweden-ai-methodology/)). Against that, Lighthouse reports that in 2022 the agency investigated "<span class="kpi">5,520</span> cases of suspected fraud" and referred "<span class="kpi">1,686</span>" to police, and that the following year prosecutors reported outcomes in "1,054 cases: <span class="kpi">166</span> convictions and 83 plea bargains" ([source](https://www.lighthousereports.com/methodology/sweden-ai-methodology/)).

**The regulator and the withdrawal (2025).** Sweden's data protection authority IMY opened supervision on 19 June 2025 in case IMY-2025-11828, noting it had read "en artikel som Lighthouse Reports publicerade den 27 november 2024" about the agency's control work in temporary parental benefit ([source](https://www.imy.se/globalassets/dokument/ovrigt/tillsynsskrivelse-forsakringskassan.pdf)). In its reply received 5 September 2025, the agency stated it "har använt en maskininlärningsbaserad riskprofil" for this work "men att denna togs ur bruk en dryg månad före tidpunkten för Försäkringskassans svar", that is, it had taken the model out of use a little over a month earlier, and had no current plans to bring it back ([source](https://www.imy.se/globalassets/dokument/beslut/2025/beslut-tillsyn-forsakringskassan.pdf)). IMY closed the case on 18 November 2025: "IMY bedömer mot denna bakgrund att det inte finns skäl att vidta ytterligare utredning i ärendet" ([source](https://www.imy.se/globalassets/dokument/beslut/2025/beslut-tillsyn-forsakringskassan.pdf)). That closure is not a ruling: no regulator has found the processing lawful or unlawful, and the withdrawal date rests on the agency's own statement as recorded by IMY ([source](https://www.imy.se/globalassets/dokument/beslut/2025/beslut-tillsyn-forsakringskassan.pdf)).

## Why this matters for buyers

The agency took the model out of use after IMY opened supervision in June 2025, while its own inspectorate had flagged the equal-treatment failure back in 2018 ([source](https://www.imy.se/globalassets/dokument/beslut/2025/beslut-tillsyn-forsakringskassan.pdf)). ISF's 2018 recommendation was that risk-based selection profiles "ska genomgå en certifiering eller någon form av oberoende etisk prövning", independent certification or ethical review ([source](https://isf.se/download/18.6e75aae16a591304891f1e8/1565330423683/Profilering%20som%20urvalsmetod%20f%C3%B6r%20riktade%20kontroller-ISF-sammanfattning-2018-05.pdf)). A buyer of any risk-scoring automation should treat that as the control that matters: a hit-rate gain was never in dispute here, and it did not save the system once the fairness test was run by someone outside it.

## How this was verified

Method: independent public record only, with no contact with Försäkringskassan. On 2026-10-05 the maker fetched and archived to the Wayback Machine: ISF's summary of report 2018:5 "Profilering som urvalsmetod för riktade kontroller" (local copy `sources/isf-rapport-2018-5-sammanfattning.pdf`); IMY's supervision letter of 19 June 2025 and closing decision of 18 November 2025 in case IMY-2025-11828 (local copies in `sources/`; the decision PDF's quoted paragraph was extracted through a reader proxy because one paragraph uses an unmapped font); Lighthouse Reports' methodology and investigation pages of 27 November 2024; and Amnesty International's statement of the same date (live page returns 403 to scripts, retrieved through a reader proxy). The seed named "ISF report 2018:15"; the ISF document located and cited is report 2018:5, which is also the one Amnesty links, and the 2018:15 reference was not confirmed. Date verified: 2026-10-05. Status: `checking`.

## Related case files

- [The Hague District Court ruled the SyRI welfare-fraud algorithm unlawful](/stories/hague-district-court-syri-welfare-fraud-algorithm-unlawful-article-8-echr-2020/) - the step Sweden never reached: a court, not a newsroom, testing a welfare risk model against fundamental rights.
- [The Dutch DPA fined the Tax Administration EUR 2.75M for algorithmic nationality profiling](/stories/dutch-dpa-fines-tax-authority-2-75m-for-unlawful-algorithmic-risk-profiling-in-c/) - a data protection authority that did issue a finding, where IMY closed its file once the model was withdrawn.
- [The UK DWP's machine-learning fraud model under National Audit Office scrutiny](/stories/dwp-ml-fraud-model-universal-credit/) - a benefits-fraud model whose fairness analysis an official auditor examined in public, the transparency Försäkringskassan refused.
- [Michigan's MiDAS unemployment-fraud algorithm and its 93 percent error rate](/stories/michigan-midas-unemployment-fraud-algorithm-40195-cases-93pct-error-20m-settlement-2024/) - what an automated fraud flag costs when nobody checks its error rate before it acts on people.

## Sources

1. **Tier 1 (inspectorate report).** Inspektionen för socialförsäkringen (ISF), "Profilering som urvalsmetod för riktade kontroller", summary of report 2018:5, 2018. https://isf.se/download/18.6e75aae16a591304891f1e8/1565330423683/Profilering%20som%20urvalsmetod%20f%C3%B6r%20riktade%20kontroller-ISF-sammanfattning-2018-05.pdf (archived https://web.archive.org/web/20261005072442/https://isf.se/download/18.6e75aae16a591304891f1e8/1565330423683/Profilering%20som%20urvalsmetod%20f%C3%B6r%20riktade%20kontroller-ISF-sammanfattning-2018-05.pdf)
2. **Tier 1 (regulator record).** Integritetsskyddsmyndigheten (IMY), "Tillsyn enligt dataskyddsförordningen, begäran om information", IMY-2025-11828, 19 June 2025. https://www.imy.se/globalassets/dokument/ovrigt/tillsynsskrivelse-forsakringskassan.pdf (archived https://web.archive.org/web/20261005072413/https://www.imy.se/globalassets/dokument/ovrigt/tillsynsskrivelse-forsakringskassan.pdf)
3. **Tier 1 (regulator decision).** Integritetsskyddsmyndigheten (IMY), "Beslut efter tillsyn enligt dataskyddsförordningen, Försäkringskassan", IMY-2025-11828, 18 November 2025. https://www.imy.se/globalassets/dokument/beslut/2025/beslut-tillsyn-forsakringskassan.pdf (archived https://web.archive.org/web/20261005072450/https://www.imy.se/globalassets/dokument/beslut/2025/beslut-tillsyn-forsakringskassan.pdf)
4. **Tier 2 (independent investigative newsroom, published method and data).** Lighthouse Reports, "How we investigated Sweden's suspicion machine", 27 November 2024. https://www.lighthousereports.com/methodology/sweden-ai-methodology/ (archived https://web.archive.org/web/20261005072416/https://www.lighthousereports.com/methodology/sweden-ai-methodology/)
5. **Tier 2 (independent investigative newsroom).** Lighthouse Reports with Svenska Dagbladet, "Sweden's Suspicion Machine", 27 November 2024. https://www.lighthousereports.com/investigation/swedens-suspicion-machine/ (archived https://web.archive.org/web/20261005072415/https://www.lighthousereports.com/investigation/swedens-suspicion-machine/)
6. **Tier 2 (NGO statement; reviewed source 4's analysis, so not an independent measurement).** Amnesty International, "Sweden: Authorities must discontinue discriminatory AI systems used by welfare agency", 27 November 2024. https://www.amnesty.org/en/latest/news/2024/11/sweden-authorities-must-discontinue-discriminatory-ai-systems-used-by-welfare-agency/ (archived https://web.archive.org/web/20261005072413/https://www.amnesty.org/en/latest/news/2024/11/sweden-authorities-must-discontinue-discriminatory-ai-systems-used-by-welfare-agency/)