wunder beta🤖 Bias in AI
An AI once guessed a person's gender from a photo. It was almost always right for lighter-skinned men — and wrong nearly a third of the time for darker-skinned women. How can a computer be so unfair?
What you’ll learn
- How Does AI Learn?Understand that many AIs learn from examples (training data) rather than typed rules, and that the AI cannot know what it was never shown.Instead of following rules, many AIs learn patterns from huge numbers of examples called training data, the way a person learns 'cat' from seeing many cats. An AI can't peek outside its training data, so those examples become its whole world.
- Lopsided Examples, Lopsided AIDefine bias as an AI working better for some groups than others due to imbalanced training data, not computer emotions.When training data is lopsided, the AI comes out lopsided — 'garbage in, garbage out.' This unfairness is called bias. It isn't the computer being mean; it's an echo of whatever examples the AI was shown, as a cause-and-effect chain makes clear.
- A Real Story: FacesExamine the 2018 Gender Shades study and read the measured accuracy gap between demographic groups.Scientists Joy Buolamwini and Timnit Gebru tested face software and measured its accuracy across groups. It erred just 0.8% for lighter-skinned men but up to 34.7% for darker-skinned women — the same program giving very different results, which is bias made visible.
- Where the Bias Came FromTrace the measured bias back to lopsided training data and understand that bias can arise without any bad intent.A common face data set (IJB-A) was about 79.6% lighter-skinned and only 20.4% darker-skinned, so AIs got little practice on darker-skinned faces. The bias came from imbalanced examples gathered through careless, not evil, choices.
- Is the Computer Being Mean?Correct the misconception that biased AI reflects computer emotions or that some faces are inherently unreadable.Learners predict why a face app fails on some groups, then check: computers have no feelings and no group's faces are magically harder. The real cause is lopsided training data — change the examples and the results change.
- How People Fix ItExplain that because bias is measurable, people can reduce it through testing, balanced data, and ongoing checks — and adopt the habit of asking about an AI's training data.Because bias comes from data, it can be measured and fixed: measure across groups, balance the examples, test again, and keep watching. After the study, companies improved. The lasting habit is to ask what examples any AI learned from.
Questions this course answers
How do many kinds of AI learn to recognize things?
Instead of following typed rules, many AIs learn from examples called training data — like learning 'cat' by seeing thousands of cat photos. Whatever is in the training data shapes what the AI becomes.
Why does it matter which examples an AI is shown?
An AI can't peek outside its training data. If it only sees orange cats, it may be confused by a black cat. The examples it studies become its entire world, so leaving groups out causes problems.
What does it mean to say an AI is 'biased'?
Bias means the AI works better for some groups than others — not because the computer has feelings, but because it echoes the lopsided pattern of its training data. Garbage in, garbage out.
In the 2018 study, the face software had a 0.8% error for lighter-skinned men but up to 34.7% error for darker-skinned women. What does this gap show?
It was the same program, yet it was almost perfect for one group and wrong more than a third of the time for another. That large gap in accuracy between groups is exactly what bias looks like when you measure it.
Joy Buolamwini found that a common set of practice faces (IJB-A) was about 79.6% lighter-skinned and only 20.4% darker-skinned. How does this explain the errors?
If darker-skinned faces are only a small slice of the examples, the AI barely practices on them and makes more mistakes on them later. The imbalance in the training data caused the imbalance in the results.
What is the real reason a face app can work worse for one group of people?
Computers have no feelings, and no group's faces are magically unreadable. The app failed because its training data was lopsided and it barely practiced on that group. Change the examples, and the results change.
Grounded in trusted sources
- Buolamwini, J. & Gebru, T. (2018), 'Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification', Proceedings of Machine Learning Research (PMLR) v81
- MIT News (2018), 'Study finds gender and skin-type bias in commercial artificial-intelligence systems' (news.mit.edu/2018/study-finds-gender-skin-type-bias-artificial-intelligence-systems-0212)
- MIT Media Lab, Gender Shades project (media.mit.edu/projects/gender-shades)
- Buolamwini, J. (2017), 'Gender Shades' thesis, MIT (dataset composition of the IJB-A and Adience benchmarks)
Every Wunder lesson is built from real, reputable sources — never invented.
Related Science courses
Wunder is a personalized learn-anything platform — tell it any topic and it builds a beautiful, fact-checked course in minutes, with narration, a knowledge check, and a college-style University track.
Browse more Science courses · All topics · Home
© 2026 Wunder Learning LLC · Terms & Privacy