AI Dictionary › AI Fundamentals
Mitigazione del Bias
Bias mitigation is the set of techniques used to reduce the distorting effects an AI system inherits from its training data, model design, or the way it is evaluated. While bias describes the problem, a system that systematically favors or disadvantages certain groups or categories, mitigation describes the corrective action, which can be applied at different stages of the model's lifecycle.
Techniques fall into three main families: interventions on data before training, such as rebalancing underrepresented samples or removing spurious correlations; interventions during training, such as loss functions that explicitly penalize disparities between groups; and interventions after training, such as calibrating decision thresholds per group or post-processing the output. No technique fully eliminates bias: it is always a trade-off between overall accuracy and fairness across subgroups.
In modern AI, bias mitigation is central in areas with high impact on people, such as hiring, credit approval, predictive justice, and medical diagnosis, where a skewed system can translate into concrete discrimination and legal consequences. Many companies include periodic fairness audits in their model validation process before release.
The topic is rooted in statistics and social science applied to machine learning since the 2010s, but it became an independent and well-funded research field as automated decision-making systems spread widely in the second half of that decade.
From our network
AGORÀ Intelligence: Enterprise AI Governance Platform
Govern AI at scale: policies, adoption and measurable results on your data. Built for boards and C-suite.
Visit agora-intelligence.com →From the Agora Intelligence blog
📱 Download the Android app (beta) iOS coming soon
Say what you mean. Get what you need.
Grace Certified, the AI coach that trains and certifies your prompt engineering, by Agora Intelligence.