Back to library

⚖️Audit AI Models for Bias

Three fairness metrics. One model. They disagree. Walk a synthetic loan classifier through demographic parity, equalized odds, and calibration; see where they conflict; then outline a regulator-defensible audit plan for a resume screener.

Foundations14 drops~2-week path · 5–8 min/daytechnology

Phase 1Three definitions of fair — and why they conflict

Three definitions of fair — and why they conflict

4 drops
  1. 'Fair' is not one thing — it's at least three

    7 min

    'Fair' is not one thing — it's at least three

  2. Bias lives on axes you have to name out loud

    7 min

    Bias lives on axes you have to name out loud

  3. The four-fifths rule is the regulator's first-pass test

    6 min

    The four-fifths rule is the regulator's first-pass test

  4. Why you can't be fair by every metric at once

    8 min

    Why you can't be fair by every metric at once

Phase 2Compute three metrics; watch them disagree

Compute three metrics; watch them disagree

5 drops
  1. Build a synthetic loan dataset with a known bias

    8 min

    Build a synthetic loan dataset with a known bias

  2. Measure selection-rate gap — the simplest fairness metric

    7 min

    Measure selection-rate gap — the simplest fairness metric

  3. Measure error-rate gaps — the fairness metric regulators reach for

    8 min

    Measure error-rate gaps — the fairness metric regulators reach for

  4. Measure calibration — does a predicted 0.8 mean the same thing for both groups?

    8 min

    Measure calibration — does a predicted 0.8 mean the same thing for both groups?

  5. Watch all three metrics disagree on the same model

    9 min

    Watch all three metrics disagree on the same model

Phase 3Pre-, in-, and post-processing mitigation

Pre-, in-, and post-processing mitigation

4 drops
  1. Your PM says 'just rebalance the training data'

    7 min

    Your PM says 'just rebalance the training data'

  2. The data scientist proposes a fairness constraint in the loss function

    8 min

    The data scientist proposes a fairness constraint in the loss function

  3. Post-processing fixes the metric, but the lawyer asks 'is this legal?'

    8 min

    Post-processing fixes the metric, but the lawyer asks 'is this legal?'

  4. AIF360 or Aequitas — pick a toolkit and defend it

    8 min

    AIF360 or Aequitas — pick a toolkit and defend it

Phase 4Outline a regulator-defensible audit plan

Outline a regulator-defensible audit plan

1 drop
  1. Outline a bias-audit plan for a hypothetical resume screener

    10 min

    Outline a bias-audit plan for a hypothetical resume screener

Frequently asked questions

What's the difference between demographic parity, equalized odds, and calibration?
This is covered in the “Audit AI Models for Bias” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
Why can't a single model satisfy all three fairness metrics at once?
This is covered in the “Audit AI Models for Bias” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
When should I use AIF360 vs Aequitas for fairness auditing?
This is covered in the “Audit AI Models for Bias” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
What does a regulator-defensible AI bias audit actually include?
This is covered in the “Audit AI Models for Bias” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
How do pre-, in-, and post-processing mitigations differ in practice?
This is covered in the “Audit AI Models for Bias” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.