Local research on AI

Jade Garcia Bourrée

Jade Garcia Bourrée completed her thesis as part of the WIDE 'Inria research team. Entitled "Faire confiance mais vérifier: audits de systèmes d'intelligence artificielle pilotés par des robots" (Trust but verify: audits of artificial intelligence systems driven by robots), she is looking at ways of detecting, verifying and limiting algorithmic biases in order to control the excesses of AI.

Algorithms are not neutral: they often reproduce the biases present in the data and the choices made by their designers. In 2010, the PredPol predictive policing AI model used in Chicago to forecast crime was accused of reinforcing racial biases by targeting disadvantaged neighbourhoods. Similarly, some recruitment algorithms favoured men by reproducing gender stereotypes derived from historical data.

Faced with these risks, legal frameworks are emerging in Europe and internationally to make automated systems more accountable. However, regulation alone is not enough.

Algorithm auditing appears to be an essential lever. It consists of testing an algorithm using fictitious profiles and analysing the decisions produced, without having access to its internal workings: a so-called "black box" approach, comparable to evaluating a dish without knowing the recipe.

Despite technical constraints (limited number of queries, difficulty in interpreting discrimination and its causes) algorithmic auditing remains a key tool for checking compliance with the law and encouraging fairer and more equitable use of artificial intelligence.