What is bias in AI?
In one line
Ages 8 to 10: If a machine learns mostly from some kinds of examples, it may work well for those and badly for others, and that unfairness is called bias.
Ages 11 to 13: Bias shows up when you test a model on different groups separately and find that it makes more mistakes for some groups than for others.
Ages 14 to 18: Standards guidance groups AI bias into systemic, statistical and human bias, and managing it means testing by subgroup, documenting data and making someone accountable.
Bias in AI means a system's results are unfair or lopsided in a consistent way, often working better for some groups of people than for others. It can come from the examples the system learned from, from choices made by the people who built it, or from how people use its answers.
Last reviewed: 20 October 2026 6 min read
The idea in one picture
Imagine a voice helper that learned mostly from adult voices. Here is an illustrative test with 200 recordings it has never heard:
| Group | Understood correctly |
|---|---|
| Adults | 95 of 100 |
| Children | 70 of 100 |
| Everyone | 165 of 200 = 82.5% |
The overall score, 82.5%, sounds fine. Split by group, the gap appears: the helper makes six times as many mistakes with children as with adults (30 against 5). Nobody had to intend this. The machine simply had fewer children's voices to learn from. A group-by-group test is how bias is found.
How it works, step by step
Bias can enter at every stage, not only in the data.
- Choosing the goal. People decide what the system should do and what counts as success.
- Collecting data. If some groups are missing or rare in the examples, the system can be less accurate for them.
- Labeling. People's judgments and assumptions go into the labels.
- Learning. The model picks up patterns from past data, including unfair ones from history.
- Using the output. People may trust an automated answer too much, or apply it to groups it was never tested on.
A little deeper (ages 11 to 13)
To find bias, test separately for each group that matters and count the mistakes in each. One total can hide a large gap, as the table shows.
Real cases have been found this way. One well-known study tested three commercial systems that guess a person's gender from a photo, and found that darker-skinned women were the most misclassified group. Two common test datasets were mostly lighter-skinned faces. Counting mistakes group by group made the problem visible.
Going further (ages 14 to 18)
A US standards body groups AI bias into three kinds. Systemic bias comes from institutions and history and appears in datasets and practices. Statistical and computational bias comes from samples that do not represent the population, and from modeling choices. Human bias comes from how people think and decide, including how they use automated output. It also notes that zero risk of bias is not achievable, so bias has to be managed, not declared solved.
Practical steps follow from this: document where the data came from and who is missing; report results by group, not only overall; decide which mistakes cost more and for whom; and make someone accountable for checking and for answering people who are affected.
Note that "bias" has a second, unrelated meaning in machine learning: the bias term, an extra number in a model that shifts its output, like the "+ b" in a straight line's equation.
Common mix-ups
- Mix-up: "AI is biased because it has opinions."Fact: A model has no opinions. It reproduces patterns in its data, design and use.
- Mix-up: "Bias only comes from the data."Fact: It also comes from the goal people choose, from labels and from how outputs are used.
- Mix-up: "Bias can be removed completely."Fact: Standards guidance says zero risk is not possible; it has to be measured and managed.
- Mix-up: "The bias in a neural network is the same thing."Fact: That bias is just a number added inside the model.
Try it (no device, 5 minutes)
Ask three people to each draw "a shoe" in ten seconds. Look at the drawings together: which kinds of shoe appear? Now imagine a shoe-spotting machine that learned only from these drawings. List three kinds of shoe it might fail on, such as sandals, boots or baby shoes. Who would be let down by it, and what examples would you add?
Quick check
In the table, how many mistakes did the helper make for adults and for children?
Answer: 5 for adults (100 − 95) and 30 for children (100 − 70).
5 for adults (100 − 95) and 30 for children (100 − 70).Why can a high overall score hide bias?
Answer: The total mixes all groups together, so a large gap between groups can disappear inside one average.
The total mixes all groups together, so a large gap between groups can disappear inside one average.Name the three kinds of AI bias in the standards guidance.
Answer: Systemic, statistical and computational, and human.
Systemic, statistical and computational, and human.Questions
Bias in AI is when a system's results are unfair or lopsided in a consistent way, often working better for some groups of people than for others. It usually comes from the examples, the design choices or the way the system is used.
It can enter through the goal people choose, data that leaves some groups out, labels that carry people's assumptions, patterns from an unfair past, and people trusting or using the output in the wrong place. Standards guidance stresses that bias is not only a data problem.
Researchers testing three commercial systems that guess gender from a face photo found that darker-skinned women were misclassified more often than any other group they tested. Two common test datasets were mostly lighter-skinned faces.
It can be reduced and managed, but standards guidance says zero risk of bias is not achievable. Useful steps are more representative data, testing every group separately, documenting the data and making someone accountable.
Ask who the system was built for and tested on, and who might be missing from its examples. Then look for results split by group, not only one overall score.
Mechna Academy is a city of machines for curious children aged 8 to 18, opening first to ages 8 to 10. In each mission, a child guesses how a machine decides, then opens its Peek Hatch to check, changes one thing and tests it on a case it has never seen. Children learn how AI works this way, and they never chat with an AI.
For grown-ups: join the waitlistPlaces are limited. The waitlist is for parents and guardians 18 and older.