PublicAlgoBeta
All case studies

FR · 4 December 2023

France's Digital Inquisition: the family benefits fund's risk score

Image: Lighthouse Reports, from the original investigation. All rights reserved by the owner.

Reported by
Lighthouse Reports, Le Monde
Story language
EN · FR
Methods
  • FOI / records request
  • Code or model obtained
  • Data analysis
  • Interviews and surveys

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.

lighthousereports.comFrance's Digital Inquisition: the family benefits fund's risk scoreA logistic-regression model used by France's national family benefits fund (CNAF) that gives each household on the benefit rolls a risk score between 0 and 1. The score steers which households get fraud checks.France · Reported by Lighthouse Reports, Le Monde (opens in a new tab)

How they investigated

  1. FOI / records request

    Le Monde filed an access request in June 2023 for the model's code and documentation. The fund sent the current code with most variable names redacted, plus full code for two earlier model generations (2010 to 2014 and 2014 to 2020) and aggregate figures on investigations. Earlier versions had been released after campaigners La Quadrature du Net and Changer de Cap successfully argued their case before the access-to-documents body CADA in late 2022.

  2. Code or model obtained

    Because the current model's variable names were redacted, the reporters analysed the unredacted 2014 to 2020 model as a stand-in. They read the logic, looked at the effect of each of its 33 variables, and built hypothetical household profiles, fixing some variables and varying others, then converted raw scores to a 0 to 1 scale. They used 0.6 as an approximate high-risk cutoff, based on the fund's statistics.

  3. Data analysis

    They combined the fund's aggregate statistics (over 700,000 investigations since 2019, with repayment outcomes) with profile tests of how hard-coded thresholds push households over the cutoff.

  4. Interviews and surveys

    Interviews with insiders and with benefit recipients who had been flagged, and an on-the-record interview with the fund's director. The fund did not answer a detailed list of questions.

Read alongside