Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Systemic discrimination
A form of discrimination embedded in systemic structures and processes, where algorithms reproduce and amplify existing inequalities automatically and at scale. It manifests through persistently unfavorable decisions towards specific groups.
Equalized odds
A fairness criterion requiring that true positive and false positive rates be equal across different demographic groups. This stricter approach than equal opportunity ensures equivalent predictive performance across groups.
Disparate impact
A legal and ethical concept identifying when a seemingly neutral policy or practice has disproportionate and negative effects on protected groups. In AI, it applies when algorithms produce inequitable outcomes even without discriminatory intent.
Representation bias
Systematic under-representation or over-representation of certain groups in datasets, leading to models that are less performant for under-represented groups. This bias directly affects the quality and fairness of predictions.
Counterfactual fairness
A fairness approach based on comparing outcomes for an individual with and without a protected characteristic, all other conditions being equal. It aims to ensure that decisions do not change solely due to sensitive attributes.