CNAF benefit-fraud risk score (France)

The CNAF risk score, known as the “datamining” score, is software that France's family-benefits fund built and runs in house. Each month it gives every benefit-receiving household a fraud-suspicion score between 0 and 1. Controllers at the family-allowance funds open fraud controls on the highest scores first.

What the score weighs

The score comes from a logistic regression, a statistical model that combines weighted facts about a household into one number. The 2014 version weighed about 33 such facts. It covers the data of about 32 million people, close to half of France's population. That makes roughly 13 million scores each month.

It reads what the fund holds on each household: declared income and benefit amounts, household makeup, life events, and contact habits, such as online log-ins.

What raises the score

Le Monde and Lighthouse Reports analysed the 2014 version. On its own weights, markers of economic vulnerability raised the score. They included low income, unemployment, receiving the RSA minimum-income benefit, and a high rent compared with income. Others were rarely logging in online, a recent separation or move, errors in declarations, and being a single parent.

Before 2025, working while receiving the disability allowance, called the AAH, raised it too. Critics called that “the height of cynicism.” In the analysis, a family with a stable income averaged about 0.33. A person working while receiving the AAH averaged about 0.66.

What the score looks for

The score looks for overpayments (indus in French), not proven fraud. It was trained to find overpayments of at least 600 euros a month lasting at least six months. CNAF's own training manual says this target covers 98 percent of fraudulent payments. An overpayment is often an honest administrative or declaration error, not deliberate fraud.

CNAF's annual random-audit survey supplied the examples the score learned from. Each year the fund checks a random sample of files, whatever their score.

Who decides

CNAF calls the score a decision-aid that only sets which files are checked first. Controllers, not the score, decide whether a household was overpaid, and if so whether by fraud or error. The sources read do not say who decides the outcome of the roughly 29 million automated controls.

The most invasive controls happen on site. There, controllers can question who lives in the household, examine bank records, and visit the home.

The household never sees its score and cannot appeal it directly. It can contest only the control decision or the repayment demand that follows.

How many controls there are

Official figures for 2024 record 31.5 million controls involving 6.4 million benefit recipients. About 29 million were automated. The other 2.5 million were document checks or on-site visits. The national anti-fraud service's share of detected fraud rose from 48 million euros in 2021 to 166 million euros in 2024.

What the reporting found

Reporting found that since 2019 the score flagged more than 700,000 investigations. About 35 percent of flagged households ultimately had to repay, about 821 euros on average. About 17 percent were found to be owed money by CNAF. In random-audit controls, about 15 percent of households had to repay.

Of the more than 100,000 people sent each year to in-depth investigations, nearly seven in ten were chosen because of a high score. The analysis could not measure how often people in protected groups were flagged wrongly. Outcome data by protected group was not available.

What each side says

CNAF's own statistics department, the DSER, ran a simulation study. Le Monde and La Quadrature du Net reported it in October 2025. Recipients of the RSA minimum-income benefit were about 13 percent of beneficiaries, but 39 to 41 percent of the highest-scoring 5 percent. Single mothers were about 14 percent of beneficiaries, but 37 to 40 percent of that top group. Households including a foreign national scored higher, even after the nationality variable was removed. The full study is not public.

In October 2025, the French ombudsperson, the Défenseur des droits, filed observations with the Conseil d'État. It found that “a presumption of indirect discrimination appears established,” because the difference rests on beneficiaries' economic vulnerability. It said the tool appears to over-control the most precarious people.

CNAF disputes the discrimination framing. It calls the tool a neutral decision-aid that targets significant and repeated overpayments. One CNAF director argued it was “the opposite of discrimination,” because no one can explain why a given file is targeted. No court had ruled as of the sources read.

How the score was opened up

In 2023, Le Monde and Lighthouse Reports obtained earlier versions of the model under France's freedom-of-information law. They went through the CADA, the Commission for Access to Administrative Documents, after CNAF resisted.

On 15 and 16 October 2024, fifteen organisations led by La Quadrature du Net asked the Conseil d'État to strike the score down. They cited data protection and non-discrimination. Amnesty International France was among them. In January 2026, ten more organisations joined, bringing the coalition to 25. A public hearing was expected in spring 2026.

The 2018 version challenged in the case ran until January 2026, and it had been withheld from the 2023 investigation. CNAF published the redesigned model's source code on 15 January 2026. La Quadrature du Net reported in late February 2026 that CNAF had just quietly published the 2018 version. Its headline said CNAF published its code but left out the essentials.

The redesigned model

A redesigned “2025” model entered production in January 2026. It removed working while receiving the AAH, nationality, housing type, and behaviour as variables.

Critics, including La Quadrature du Net, say it still over-targets the same groups. It keeps triggers such as receiving three or more benefits, receiving over 200 euros a month in benefits, low income, and starting or leaving the RSA. Its effect so far comes from the DSER's simulations, not from measuring it in production.

What this case asks

Watch two things at once. First, the facts that raise the score are the markers of the people a safety net exists to serve. The case file calls this a double penalty: the people the safety net serves are the people it suspects. Counting ordinary errors as overpayments sharpens it.

Second, watch for a loop. CNAF's documented training method uses the random-audit survey. If the results of past controls also shaped retraining, a skew in who gets checked would shape who the score suspects next. The sources do not document that step.

The case also shows that opening a score up is not the same as stopping it. Its weights were forced into view, and the system kept running.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. A pathway counts as closed once it passes on only a few mistakes. It need not stop them all.

This case's budget is 13 units. Each tool costs units from it.

Explore (No Targets) sets no targets. Under Service Targets Only, the targets can be met. The cheapest combinations use two tools and cost 4 of the 13 units. Each pairs Mark AI-written records with Escalate checks, Peer sharing rules, or Assign a challenger.

Under Service and Safety Targets and under All Governance Targets, the targets can also be met. Both levels ask you to close every failure pathway, among other targets. The cheapest combinations cost 11 of the 13 units. They use four tools: Store less data, Mark AI-written records, Peer sharing rules, and Escalate checks at its stronger setting.

Every combination that meets these targets includes Store less data and Escalate checks at its stronger setting. Escalate checks at that setting is the one tool that closes High scores sent to controls. Store less data is the one tool that closes Outcomes written to records and Scores logged to records.

Understand the system costs 4 units at these two levels and 3 at the lower two. While it is in use, it cuts the price of Escalate checks, Peer sharing rules, Mark AI-written records, and Store less data. The cut is 1 unit each. At its stronger setting, which costs 6 units, the cut is 2, though no tool drops below 1 unit. Adding it to the four tools above keeps the total at 11 units at either setting. At the standard setting, its 4 units are offset by 4 units of cuts. At the stronger setting, its 6 units are offset by 6, because Mark AI-written records and Peer sharing rules stop at 1 unit.

Three tools change no pathway on this network: Review on schedule, Keep skills sharp, and Upgrade model. The first limits how far mistakes build on one another. The other two act on the controllers and on the score itself.