Stop 13 · Business Communication & StorytellingTopics · 5/7

Ethics & Bias in Machine Learning

In short

A model learns from history, and history isn't always fair. If past decisions were biased, your model can copy that bias at scale, quietly, with a confident-looking score.

Models trained on historical data can encode and amplify existing societal biases — recognizing this risk and testing for it is a core professional responsibility, not an optional add-on.

The mascot holds a clipboard and studies a line of young cricketers in pads and helmets.

Think of it like this

The selector with a blind spot

A selector picks only players from a few big-city academies because past champions came from there. Talented kids from small towns never get a trial, so his 'data' keeps proving him right. Models trained on past hires or loans can fall into exactly this loop.

Watch it happen

Where bias sneaks in

Data

Labels

Features

Model

Deployment

Feedback

1/6Some groups are missing or rare in the data, so the model sees too few examples of them.

The big ideas

What you need to know

Bias at every stage

Data collection, labelling, features, evaluation and deployment can each add bias. Check all of them.

Dropping a column isn't enough

Removing 'gender' does little if other features act as proxies for it.

Measure per group

Compare approval rate, recall and false positive rate across groups, not just overall accuracy.

Keep humans in the loop

For high-stakes calls like loans, hiring or health, let people review and appeal decisions.

See it in code

Walk through the code

python

1import pandas as pd2 3df = pd.DataFrame({4    "group":  ["A", "A", "A", "A", "B", "B", "B", "B"],5    "actual": [1, 0, 1, 1, 1, 0, 1, 1],6    "pred":   [1, 0, 1, 1, 0, 0, 0, 1],7})8 9df["approved"] = df["pred"] == 110pos = df[df["actual"] == 1]11 12print(df.groupby("group")["approved"].mean())   # approval rate per group13print(pos.groupby("group")["pred"].mean())      # recall per group

1/3Toy results: each row has a group, the true outcome and the model's prediction.

A tiny fairness check. Group A's approval rate is 0.75 and B's is 0.25; deserving B applicants are missed far more often.

Save yourself the pain

Common mistakes

Oops

Deleting the sensitive column and calling the model fair.

Do this instead

Keep it aside for testing. Measure outcomes per group and look for proxy features.

Oops

Only reporting overall accuracy.

Do this instead

Break metrics down by group. A 90% accurate model can still fail one group badly.

Your mission

Run a group fairness check

~35 min

Quick check

You removed 'gender' from a hiring model, but it still favours men. What's the most likely reason?

Go deeper

1Read a real case study of a biased model that caused a public issue.
2Learn to check model performance across different demographic subgroups, not just in aggregate.
3Practice identifying a potential bias risk in a hypothetical model idea of your own.