Algorithmic Bias: Definition and Examples

Algorithmic bias occurs when an AI system produces systematically unfair or discriminatory results toward certain groups of people. These biases are not intentional: they emerge from the data used to train the model, the design choices of engineers, or the optimization objectives defined during development.

Biases have multiple sources. Training data can reflect historical prejudices (for example, if a recruitment model is trained on past hiring decisions that favored one gender). Bias can also stem from underrepresentation of certain populations in the datasets, or from proxy variables that indirectly encode sensitive characteristics such as ethnicity or socioeconomic status.

In prompt engineering, understanding algorithmic bias is essential because large language models (LLMs) inherit biases present in their training corpora. A poorly formulated prompt can amplify these biases, while a carefully designed prompt can mitigate them by explicitly asking the model to consider multiple perspectives or to verify the fairness of its responses.

Detecting and correcting algorithmic biases is now a major research field in responsible AI. Techniques such as model auditing, red teaming, and evaluation using fairness metrics help identify and reduce these biases, even though it is practically impossible to eliminate them entirely.