France Travail's Job-Search Control Targeting Model — 26 Variables, 7,000-File Pilots and an Unpublished Evaluation

In July 2026, La Quadrature du Net and Radio France's investigation unit revealed that France Travail (the French public employment service) is testing a statistical model, "Ciblage du Contrôle de la Recherche d'Emploi", to select which jobseekers' files are checked for genuine job search. It was presented to France Travail's AI ethics committee on 10 December 2025.
The tool is a decision-tree model (not generative AI) trained on about 60,000 previously checked files and using 26 variables, including education level, length of registration, interviews in the past six months, non-salaried activity, CV visibility and cancelled appointments. Two pilot campaigns of roughly 7,000 files each targeted people registered after a conventional termination (rupture conventionnelle). France Travail says 79% of files the model selected had, retrospectively, received an in-depth review in the original check; reporting notes that this shows the model reproduces past choices, not that it finds non-compliance 79% of the time.
France Travail states that a human adviser keeps the final decision, that the criteria are objective and non-discriminatory, that the tool complies with GDPR and that it produces no individual score. Controls rose from about 200,000 in 2017 to 730,000 in 2025, with a stated goal of 1.5 million in 2027, and the model is reportedly to be extended to other groups with monthly lists.
La Quadrature du Net, which holds only the variable list, calls it opaque profiling: variables such as a missing CV or cancelled appointments may reflect digital exclusion, illness or transport barriers, and training on past control decisions may entrench existing bias. It requests the source code, the data protection impact assessment and detailed pilot results. This is a cautionary, still-unevaluated case: no accuracy, fairness or outcome evaluation has been published.

Read the full analysis: https://www.solutions-numeriques.com/france-travail-experimente-une-ia-pour-selectionner-les-dossiers-a-controler/