AI bias and fairness: how models go wrong and how to audit them

12 steps · about 8 h 45 min · Intermediate

  1. Fairness | Machine Learning | Google for Developers (interactive, Google for Developers, 1 h)

    This module gives you the overview first: where bias comes from and how to identify, mitigate and evaluate it.

  2. Fairness: Types of bias (article, Google for Developers, 10 min)

    It gives you names for the ways data goes wrong (historical, selection, reporting, automation bias), which you need to describe any real case.

  3. Fairness: Evaluating for bias (article, Google for Developers, 15 min)

    It introduces demographic parity, equality of opportunity and counterfactual fairness with a worked admissions example, and shows how overall accuracy can hide harm to a minority group.

  4. Machine Bias (article, ProPublica, 25 min)

    This is the best-known real audit of a deployed algorithm, COMPAS, and it shows what unequal error rates look like in practice.

  5. Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification (article, Buolamwini & Gebru, PMLR, 30 min)

    It shows a second kind of failure, unrepresentative data, and why you should measure by intersecting subgroups, not one group at a time.

  6. Bias in Criminal Risk Scores Is Mathematically Inevitable, Researchers Say (article, ProPublica, 10 min)

    It explains the follow-up finding that different fairness definitions can conflict, which sets up the metrics in the next steps.

  7. Fairness in Machine Learning — Fairlearn documentation (doc, Fairlearn, 20 min)

    It covers the vocabulary of harms, sensitive features and group metrics that you use to audit a model.

  8. Performing a Fairness Assessment — Fairlearn documentation (doc, Fairlearn, 25 min)

    It walks through the audit steps: name the harms, name the groups, then compute disaggregated metrics with MetricFrame.

  9. Common fairness metrics — Fairlearn documentation (doc, Fairlearn, 20 min)

    It helps you choose between demographic parity and equalized odds, and warns about misusing the four-fifths rule.

  10. COMPAS Recidivism Risk Score Data and Analysis (repo, ProPublica Data Store, 1 h 30 min)

    You can repeat the COMPAS audit yourself using ProPublica's data and its Python and R code, with the group metrics you just learned.

  11. Model Cards for Model Reporting (article, Mitchell et al., arXiv, 40 min)

    After an audit you have to report it, and model cards are the standard format for results broken down by subgroup.

  12. Fairness and Machine Learning: Limitations and Opportunities (reference, Barocas, Hardt & Narayanan, 3 h)

    This free textbook goes deepest on the criteria and trade-offs, so read the chapters that match what you hit in your own audit.