Accueil > Séminaires/Colloques > Soutenances de Thèse et HDR > PhD Defense - Thomas SOUVERAIN. "How AI classifications impact us : (...)
PhD Defense - Thomas SOUVERAIN. "How AI classifications impact us : Explaining AI in financial services and job advertising, from technical solutions to ethical issues"
Date/time : Thusday, December 19th, 1.30 p.m.
Location : Salle des Génovéfains – lycée Henri IV, 23 rue Clovis, 75 005 Paris.
Please arrive 15 mn ahead of time, and note that the defense will happen in French. To enter the building, participants will have to register on the following link (or with the QR code below) : https://forms.office.com/e/yPMbb8jvgU
Committee
- Jean-Gabriel GANASCIA Président, Rapporteur Professeur émérite Sorbonne Université – LIP6
- Marc-Antoine DILHAC Rapporteur Professeur associé Université de Montréal
- Stéphanie RUPHY Examinatrice Professeure des universités École Normale Supérieure – PSL
- Ali OZKES Examinateur Associate professor SKEMA & ; WU Vienna
- Gaëlle JOUIS Référente technique Docteure en informatique & Data Scientist France Travail
- Paul ÉGRÉ Directeur de thèse Directeur de recherche CNRS, École Normale Supérieure – PSL
Abstract
This thesis explores the notion of explainable AI (XIA), involving data-science and philosophy. Thanks to industrial partnerships, we focus on two use cases where AI classifications are deployed, and impact humans on a daily basis : loan granting (partnership with the company DreamQuark) and warnings on illegal job offers (partnership with the French Employment Agency, France Travail). We begin with a critical lens on techniques used to explain AI classifications (the what of explanation). Building our own package, we then examine how classification bias can be mitigated in loan lending, to obtain a fair outcome (the what for of explanation). Finally, we empirically investigate how users of AI interact with algorithmic recommendations, and how these interactions can be optimized (for whom to explain). We argue that explainability solutions are merely the basis of a description which must incorporate causality, fairness, and intelligibility in order to effectively explain AI.
