A deep dive into how algorithmic bias forms, persists, and how we can systematically address it.
Bias in machine learning is not just a technical problem — it reflects the data, choices, and power structures embedded in our society. In this post, we break down the main types of bias and actionable ways to audit your models.
Types of Bias
- Historical bias — reflects past discrimination in training data
- Representation bias — under-sampled groups produce weaker predictions
- Measurement bias — proxies that correlate with protected attributes
- Aggregation bias — a single model for diverse sub-populations
How to Audit
Start with disaggregated evaluation metrics across demographic slices. Tools like Fairlearn and What-If Tool make this accessible. The goal is not a perfectly fair model — it is a model whose failures are understood and mitigated.