Heza
AI Ethics & Responsibility

Bias and Fairness

When AI Gets It Wrong

You are previewing the first lesson — free to explore.

In Course 3, you learned where bias comes from technically: unrepresentative training data. Now let's look at a real, well-documented case of what that looks like in the world, and who stepped up to fix it.

Real exampleGender Shades

In 2018, researcher Joy Buolamwini — at the time a graduate student at MIT — tested several commercial facial-recognition systems and found they were highly accurate for lighter-skinned men, but far less accurate for darker-skinned women, with error rates over 30% higher in some systems.

📍 Research conducted at MIT Media Lab, with Timnit Gebru
Key conceptWhy This Happened

The datasets used to train those facial-recognition systems were not representative — they contained far more lighter-skinned faces than darker-skinned ones. The AI hadn't 'learned to be unfair' on purpose; it had simply learned an incomplete pattern from incomplete data, exactly as Course 3 described.

This research — published as the 'Gender Shades' study — didn't just criticize the problem; it directly led major technology companies to measurably improve their systems' accuracy across skin tones and genders. It's a powerful example of one researcher's rigorous work changing an entire industry.

Check your understanding

1. What did the Gender Shades research find?

2. What was the real-world impact of this research?