A systematic error of an AI system that produces results unfairly unfavourable to certain groups, often arising from biased training data.
Algorithmic bias designates the tendency of an AI system to produce systematically distorted results, often to the detriment of certain groups of people. It is not a random failure but a reproducible deviation whose origin most often lies upstream, in unrepresentative training data, in variables correlated with protected characteristics, or in the very design of the model. In insurance, algorithmic bias is a central concern, because it can lead a pricing engine to unduly penalise a category of policyholders and to turn into prohibited algorithmic discrimination. The AI Act addresses it through requirements of data governance and human oversight, while the Product Liability Directive paradoxically excludes infringements of fundamental rights from its scope, sending such litigation back to national laws.
A model trained on historical data reflecting past inequalities may reproduce an algorithmic bias by systematically overpricing a neighbourhood, which the insurer must detect and correct.
biais d'un modèle, algorithmic bias