AI Fairness: Definition and Examples

AI Fairness is a research and practice domain focused on identifying, measuring, and correcting biases in AI systems. These biases can occur at every stage of a model's lifecycle: in training data, algorithm design, or how results are interpreted and deployed. The goal is to ensure that no group defined by sensitive characteristics (gender, ethnicity, age, disability, etc.) is systematically disadvantaged.

The concept of AI Fairness rests on several mathematical definitions of fairness, which can sometimes conflict. For example, demographic parity requires equal positive decision rates across groups, while equal opportunity focuses on equal true positive rates. Choosing the appropriate fairness metric depends on the application context and the values one wishes to prioritize.

In practice, AI Fairness involves regular model audits, use of representative and diverse datasets, and implementation of debiasing techniques before, during, or after training. Frameworks such as Fairlearn, IBM's AI Fairness 360, and Google's What-If Tool allow developers to evaluate and improve the fairness of their systems.

The issue has become central with the widespread use of large language models (LLMs) in prompt engineering. Responses generated by these models can reflect and amplify stereotypes present in their training data. Prompt formulation plays a crucial role: a well-designed prompt can mitigate bias, while a poorly worded one can exacerbate it.