Algorithmic Bias in AI & Artificial Intelligence

Historical Bias is the most difficult to address because it is present in 'accurate' data. If women were historically excluded from leadership roles, a perfect model will learn that 'being female' is not a feature of a leader. Representation Bias occurs when the sample data doesn't reflect the target population. If you build a medical AI using only data from urban hospitals, it will likely fail for rural populations, creating a 'Representational Gap' in healthcare quality.

Measurement Bias happens when the features we choose to measure are flawed proxies for reality. Using 'standardized test scores' to measure 'intelligence' is a classic example—the score is influenced by wealth and access to tutoring, not just raw ability. Aggregation Bias occurs when a single model is used for a diverse population where different groups should have different models. These errors are 'baked in' during the data preparation phase.

Evaluation Bias is the final trap. It occurs when the 'Benchmark' or 'Test Set' used to approve a model is itself biased. If a vision model is tested using the same biased dataset it was trained on, it will appear highly accurate. This creates False Confidence, where a developer believes their model is ready for the real world, only for it to fail spectacularly when it encounters a diverse set of users for the first time.