AI Ethics — Bias, Fairness & Responsibility

• Explain where AI bias comes from and how to detect it.
• Apply a fairness check to a model's predictions across groups.
• Reason about privacy, consent, and data provenance.
• Adopt the builder's checklist: transparency, human accountability, harm-minimisation.

A hiring model trained on 10 years of mostly-male hires
learns to prefer male candidates — not because it's "evil",
but because it faithfully copied a biased history.

1. Bias — where does it come from?

historical bias   data reflects past unfairness (hiring example)
sampling bias     some groups under-represented in the data
label bias        the labels themselves were assigned unfairly
proxy bias        a "neutral" feature (postcode) stands in for a protected one

2. Fairness — measure it across groups

# check accuracy / approval rate per group, not just overall
for group in df["gender"].unique():
    mask = df["gender"] == group
    rate = model.predict(X[mask]).mean()   # approval rate
    print(f"{group}: approval rate {rate:.1%}")
# big gaps between groups → investigate, don't ship

Overall accuracy can hide that a model works great for one group and badly for another. Always break metrics down by relevant groups.

3. Privacy & consent

• Did the people in your data agree to this use?
• Are you storing more personal data than you need?
• Could outputs re-identify someone who should stay anonymous?

4. Transparency

Can you explain a decision to the person it affects? Prefer interpretable models (trees, logistic regression) for high-stakes decisions; document data sources and known limits.

5. Human accountability

AI should ASSIST consequential decisions, not make them alone.
A human must be able to review, override, and be answerable.
"The algorithm decided" is never an acceptable excuse.

# fairness_audit.py — check a model's behaviour across a group
import pandas as pd
from sklearn.metrics import accuracy_score

# df has: features, true label 'y', a sensitive column 'group', model preds
def audit(df, group_col, y_col, pred_col):
    print(f"overall accuracy: {accuracy_score(df[y_col], df[pred_col]):.2%}\n")
    print(f"{'group':<12}{'n':>6}{'accuracy':>10}{'positive rate':>15}")
    for g, sub in df.groupby(group_col):
        acc = accuracy_score(sub[y_col], sub[pred_col])
        pos = sub[pred_col].mean()
        print(f"{str(g):<12}{len(sub):>6}{acc:>10.1%}{pos:>15.1%}")

# audit(df, "group", "y", "pred")

Sample output

overall accuracy: 88.0%

group           n  accuracy  positive rate
A             520     91.0%          62.0%
B             180     74.0%          31.0%   ← much worse + lower approval

Read the diff

Overall accuracy (88%) looked fine — but group B gets 74% accuracy and half the approval rate of group A. That gap is a red flag: the model may be unfair to group B, possibly due to under-representation in training. The audit makes the invisible visible. A responsible builder investigates and fixes this before shipping — more data for B, reweighting, or not deploying.

AI learns biases from data; accurate ≠ fair. Measure metrics per group, watch for proxy features, respect privacy and consent, prefer transparent models for high stakes, and keep a human accountable. A model card documents all of it. Building powerful AI responsibly is part of the job — not an afterthought. Next: your capstone.