Whether rudimentary or sophisticated, the transparent purpose of algorithms is to deliver a result from a well-defined command. Their use, on the other hand, is usually completely opaque, even in the public sector. Case in point with the algorithm used in France to decide which households to audit for benefit fraud. In collaboration with Lighthouse Reports, Le Monde investigated this algorithm, which calculates households' "risk score," designed and used by the Caisse Nationale d'Allocations Familiales (CNAF), the social security agency that distributes benefits including minimum income, family allowances, and housing aids. After a decade of use, little or nothing was publicly known about the criteria used to help determine who, among the 13.8 million recipient households, is deemed most "at risk" of committing fraud or errors in their statements to request benefits.

The workings of such a system, which handles the data of some 33 million citizens, should be public and the object of a democratic debate – that is the principle defended by La Quadrature du Net, a non-profit for the protection of fundamental rights on the Internet. "The purpose of this opacity is to conceal CNAF's practices and its real objective: to recover as much money as possible," said Alex, a member of the non-profit following this issue. In 2022, La Quadrature du Net asked the CNAF to publish the details of its algorithm, as well as a list of all the variables used to calculate it.

The freedom of information request, based on the 1978 law setting transparency requirements for administrations, was initially rejected by the CNAF. The agency felt that revealing its recipe would be tantamount to giving fraudsters all the keys they needed to thwart audits. La Quadrature du Net turned to the Commission of Access to Administrative Documents (CADA), the body responsible for dealing with disputes between citizens and government departments before a possible referral to the administrative court.

On December 15, 2022, the CADA issued a ruling on the subject: While the risk of circumventing audits argued by the CNAF could be understood regarding a system still in use, there is nothing to prevent the CNAF from communicating previous versions of the algorithm.

It was thanks to this breakthrough that La Quadrature du Net and Le Monde – in a parallel request – were recently able to obtain detailed information on the algorithms used to monitor benefit recipients from 2010 to 2020. The non-profit published an analysis on its website (in French), and called for "a debate around the political abuses that have led a social institution to resort to such practices."

You have 65% of this article left to read. The rest is for subscribers only.