Fairness, bias, transparency, accountability and the wider social effects of AI systems.
56 terms
Proxy discrimination occurs when an AI model disadvantages a protected group by relying on correlated features instead of the protected attribute itself.
Representation bias arises when training data under-samples or misrepresents some groups, so an AI model performs worse for the people it rarely saw.
Representational harm occurs when an AI system portrays social groups in demeaning, stereotyped or invisible ways, shaping how people are seen and valued.
A saliency map is a heatmap that shows which input features, usually image pixels, most influenced a model's prediction, supporting AI transparency.
SHAP explains machine learning predictions by assigning each input feature a Shapley value, its fair share of the difference from the average prediction.
Trustworthy AI describes AI systems that are lawful, ethical and technically robust, with evidence that they are fair, safe, transparent and accountable.