OrientIA: Reclaiming AI-Driven Guidance - Towards a Sovereign, Neutral and Equitable System for All

OrientIA

Ella, FAOUZI, Paris-Dauphine-PSL, Matteo, LEPIETRE, Paris-Dauphine-PSL

Training / education / pedagogy

In France, post-secondary and career guidance relies on a public system that is structurally insufficient: approximately one school psychologist for every 1,500 students, and 63% of high schools unable to offer individual consultations to all their final-year students (CNESCO). These constraints directly shape the choices students make: 25% of students in urban areas apply to selective programs, compared to just 10% in rural areas (education.gouv). In the absence of an available and competent advisor, young people have spontaneously adopted generative AI as a substitute: 86% of students use it in their studies (L'Etudiant) and 42% of 18–25-year-olds use it daily (Franceinfo). However, these tools present serious limitations for this purpose, both in terms of the reliability of the information they produce and the protection of the personal data they receive. It is in this context that we propose OrientIA, a free, unlimited, and sovereign public digital guidance service powered by AI, designed to regulate and specialize a usage that already exists. Built on official French data sources, the tool delivers reliable, cited, and contextually relevant responses grounded in the French educational and professional system, where general-purpose tools draw indiscriminately from web data. Users can freely explore programs, discover lesser-known careers, compare pathways, and simulate career outcomes, with no quotas or imposed workflows. Optional access to a human guidance counselor is available at any time, in a separate space fully isolated from the AI conversations, ensuring complete confidentiality and allowing each user to remain in control of their own journey.

The consequences of inadequate guidance are measurable. 56% of 18–24-year-olds report having made an ill-suited choice due to insufficient information (Opinionway/Edumapper, 2023). One-third of high school graduates change programs within their first year of higher education (IPP, 2022), up from 21% eight years earlier (IGÉSR), and each reorientation costs public finances an average of €13,200 (Bechichi, IPP, 2026). These figures reflect a well-documented reality: the lack of access to reliable information about programs and career outcomes leads a significant proportion of young people into pathways that do not match their profiles. The use of general-purpose LLMs as a substitute has not solved this problem, it has introduced new ones. These tools guide students based on institutions' web visibility rather than on the fit between a program and the user's profile, lending a commercial bias the appearance of objective advice. They also require users to share particularly sensitive information : academic records, aspirations, family circumstances, with proprietary foreign models, with no guarantee of confidentiality. Beyond the individual, this is a matter of sovereignty: entrusting the guidance of an entire generation to tools operated by foreign companies represents a loss of control over how French youth is directed toward the country's economic and professional sectors. It is the least supported young people who bear the heaviest consequences. Those with an informed network and structured support use these tools critically and as a complement. Those with access to none of these resources rely on them exclusively: for them, a misguided recommendation will be corrected by no one. It will shape their choice. Finally, public intervention is also required on regulatory grounds: the EU AI Act, applicable from August 2026, requires AI systems used in education to ensure algorithmic transparency, bias detection, and decision documentation, requirements that none of the tools currently used by young people for guidance currently meet.

Our team brings together a computer science student and an international economics student, a complementarity that naturally shaped our approach: technical feasibility on one side, analysis of inequalities and public policy on the other. Our motivation is also rooted in an asymmetric personal experience. One of us benefited from structured support and an informed network in Paris. The other, from a outside major urban centers, grew up in an environment where the teaching staff was less equipped on these matters and where access to a guidance counselor was never a reality. It is both our academic training and our lived experience that drive this project, in the face of an information asymmetry documented at the national level by the CNESCO, INSEE, and IPP. This dual perspective led us to a shared conviction: the guidance problem in France is not a problem of individual willpower, it is a problem of access to information. AI, if properly regulated and specialized, is the first tool capable of addressing it at scale. Through this challenge, we aim to rigorously document the problem through an individual-level empirical study and to demonstrate the technical feasibility of a sovereign and neutral alternative to the general-purpose LLMs currently used by young people for guidance. We submit this work for review by researchers and practitioners in order to enrich our thinking and explore the conditions under which it could be taken further. Beyond the competition, we hope to contribute, at our level, to laying the first foundations of a public AI-powered guidance service that promotes equity rather than undermining it.

Our initial contribution rests on two complementary empirical deliverables, developed in parallel with the final report: rigorously documenting the problem on one hand, demonstrating the feasibility of the solution on the other. A quantitative pilot study : The existing literature documents the scale of reorientations in higher education and identifies the information deficit as a likely explanatory factor, without however establishing this link at the individual level. The work of the IPP (Bechichi, 2024; Terrier et al., 2023), IGÉSR (2021), CNESCO, and CEREQ establishes the causal chain at an aggregate level, but no study directly tests, at the individual scale, whether the quality of guidance received directly affects the likelihood of reorientation. It is this missing link that our pilot seeks to address. The protocol is structured around a central question: are the quality and accessibility of guidance received correlated with the probability of reorientation, independently of the student's sociodemographic characteristics? It relies on two complementary levels of analysis: the effect of the information received on reorientation on the one hand, and the effect of social and geographic origin on the quality of accessible guidance on the other. A third analysis tests whether the use of AI as a source of guidance, whether for choosing a program or for exploring careers and professional outcomes, actually improves the quality of information available to students at the time of their decision, a question that has not, to our knowledge, been studied in the existing literature and whose results should inform our argument on the limitations of general-purpose LLMs for this use. The 20-question survey, designed to be completed in under 5 minutes with conditional logic for students who have reoriented, is distributed to students at the L2 level and beyond, across all disciplines, with a target of 150 to 200 respondents. The analysis is entirely coded in R and relies on a main logistic regression with odds ratio computation, Mann-Whitney and Kruskal-Wallis tests for Likert-scale variables, Chi-squared tests with Cramér's V for categorical variables, and a VIF test to check for multicollinearity. Survey available here: https://docs.google.com/forms/d/e/1FAIpQLSey_7AznA5zYms_ww3nbCupgXpdPJC_Aqmq7-liuPtSByBEBw/viewform A benchmarked RAG prototype : We are developing OrientIA, a RAG system specialized on French public orientation data: Parcoursup Open Data, ONISEP, ROME 4.0 from France Travail, and SecNumEdu labels from ANSSI. At this stage, the collection and merging pipeline is operational and has produced a dataset of 443 enriched records (acceptance rates, admitted student profiles, official labels, career outcomes) covering the cybersecurity and data/AI domains. The core technical innovation is a label-based re-ranking mechanism: after vector search, programs holding official labels (SecNumEdu, CTI, CGE) are prioritized over institutions that are simply better indexed online, directly correcting the marketing bias documented in our analysis. The boost coefficients are optimized via grid search to empirically quantify the effect of the mechanism. The benchmark, already implemented, evaluates OrientIA against ChatGPT and raw Mistral on 32 questions in a double-blind setup, scored by Claude (Anthropic) on 6 criteria. Seven configurations have already been tested, revealing phenomena that we will document in the final deliverable: retrieval constraint effects, coefficient calibration sensitivity, and differentiated impact of data enrichment depending on question type. The entire codebase is open source, for a total development cost under €20. Source code available on GitHub: https://github.com/matjussu/OrientIA What these two contributions aim to establish : The pilot study aims to empirically test whether the guidance information deficit is correlated with the probability of reorientation and whether the use of general-purpose LLMs actually improves or fails to improve students' knowledge of real career outcomes. The prototype aims to demonstrate that a tool specialized on sovereign data can produce more neutral and better-sourced recommendations than general-purpose tools, using the same base model. Together, these two contributions form the empirical and technical foundation of a broader proposal: that of a public AI-powered guidance service whose governance, architecture, and deployment conditions we outline here, hoping that this work can foster a collective discussion on how AI can become a genuine lever for equal access to information for young people in France.