France Travail Tests AI to Sort Job Seekers

France Travail Tests AI to Sort Job Seekers

France Travail is experimenting with an algorithm designed to distinguish "good" from "bad" unemployed individuals. Already tested on 7,000 cases, this tool aims to industrialize controls and could affect more than six million registered individuals.

France Travail is preparing to automate a crucial step in monitoring job seekers. According to a document revealed by La Quadrature du Net, in partnership with the investigative unit of Radio France, the organization is testing a statistical model capable of identifying individuals whose cases should be prioritized for review. The algorithm does not yet impose sanctions itself, but by selecting who will be monitored, it already determines who will be more exposed to requests for documentation, contradictory procedures, and reductions or suspensions of benefits.

An Algorithm to Predict Who to Monitor First

Presented in December 2025 to the ethics committee dedicated to artificial intelligence at France Travail, the project has a clear name: "Targeting Employment Search Control."

It relies on an algorithm tasked with automatically sorting job seekers based on the data contained in their files, in order to create two categories of profiles: "suspicious" and "non-suspicious," to use the terms from La Quadrature du Net. Cases classified in the first category must then be included on the lists sent to the agents responsible for controls.

To feed its model, France Travail has cross-referenced the conclusions of previous controls with 26 pieces of information drawn from the files. Among these pieces of information are:

• Declared activity
• CV visibility
• Education level
• Desired profession
• Length of registration
• Recent interviews
• Absences or non-salaried activity

This data is used to identify characteristics that have historically been associated with problematic cases.

However, none of this data alone proves insufficient job-seeking efforts. The absence of an online CV or a low number of interviews may also indicate digital difficulties, a disconnect from public services, or inadequate support. The model does not identify a shortcoming; it transforms statistical data into suspicion indicators.

Poverty Transformed into a Risk Signal

Of course, France Travail presents its algorithm as an objective tool capable of neutralizing human biases. However, its own tests tell a different story: in its initial version, the model disproportionately targeted individuals with the lowest incomes.

This is not surprising for a system trained on historical control data. These data also reflect the choices of the administration in a society where unemployment, precariousness, and lack of effort are often associated. The more a population has been monitored, the more its characteristics appear in cases deemed problematic, and the more the algorithm finds reasons to control them further. This is a concern that the Defender of Rights has already raised, pointing to the risk of over-monitoring the most vulnerable groups.

France Travail claims to have corrected this issue by removing the reference daily wage and rebalancing incomes among the selected cases. This intervention is said to have also reduced a bias related to education level. Thus, precariousness did indeed weigh in the classification, either directly or through other variables.

The precedent of the CNAF shows that this bias is not merely a matter of misconfiguration. According to statistics produced by the fund and cited by La Quadrature du Net, households where the head was unemployed represented 10% of beneficiaries in 2024, but 22% of controls derived from data mining. Recipients of the RSA accounted for only 13% of beneficiaries but concentrated 51% of controls. A similar imbalance exists for single-parent families: 16% of beneficiaries versus 49% of controls.

After 7,000 Controls, France Travail Aims to Scale Up

The tool is no longer limited to the datasets used for training. France Travail has tested it during an initial campaign of 7,000 controls, conducted across France with individuals registered after a mutual termination of employment. According to the organization, 50.4% of the selected cases led to a sanction or an increase in support, compared to 33.2% using traditional methods. A second campaign of similar scale was still underway at the time of the project's presentation.

France Travail now plans to extend the model to other groups, to provide control teams with automatically generated lists each month, and to gradually replace existing targeting rules. According to La Quadrature du Net, more than six million registered individuals could then be subjected to this profiling.

At this scale, if the lists grow without additional human resources, agents will have less time to examine each situation. As the European Data Protection Supervisor reminds us, the presence of an agent at the end of the chain is not sufficient to ensure genuine human oversight. The "suspicious" profile produced by the algorithm could then serve as a preliminary conclusion, and human control might be reduced to validating the outcome suggested by the targeting without a thorough re-examination of the case.