What Is AI Bias? Real Examples
AI bias is systematically unfair performance for some groups. Learn where it comes from, common real-world patterns, how to check for it and the ethics principles behind fair AI.
Short answer: AI bias means an AI system gives systematically worse or unfairer results for some people or groups than for others. It usually comes from the data a system learned from, the way the problem was framed, or the way the system is used, not from the machine "wanting" to be unfair. Understanding bias is one of the core AI ethics principles, because fairness has to be checked, not assumed.
Where Does AI Bias Come From?
An AI model learns patterns from examples. If the examples are unbalanced or reflect past unfairness, the model copies the pattern. The common sources are:
Unbalanced data. A model trained mostly on one kind of person or language works worse for others.
Historical bias. Data records past human decisions. If those were unfair, the model learns to repeat them.
Proxies. Even if you remove a sensitive field, other fields, such as a neighbourhood or a school, can stand in for it.
Labels and design choices. What counts as a "good" result is decided by people, and their assumptions shape the model.
Feedback loops. A system's decisions change the data it sees next, which can strengthen the first error.
The chapter on where bias comes from goes deeper into each of these.
What Are Real Examples of AI Bias?
Here are common patterns, described in general terms. The numbers in the code below are made up for illustration.
Area | What can go wrong | Likely cause |
|---|---|---|
Hiring tools | Candidates from one group are ranked lower | Trained on past hiring decisions that favoured another group |
Face and voice recognition | Works better for some faces, accents or languages than others | Training data under-represented those people |
Credit and risk scoring | Similar people get different scores by group | Proxy features and historical patterns |
Image and text generators | Stereotypes appear, such as who is shown in which job | Patterns in the web text and images used for training |
Language tools | Weaker answers or translations for languages with less data | Less training text in those languages |
These issues are well documented by researchers and journalists, but how large they are depends on the system and the date, so judge each system by its own evidence.
How Can You Check a System for Bias?
A first step is simple: compare how often each group gets a positive outcome. This is called the selection rate. Here is a small runnable example on made-up data:
# (group, was_selected) for 10 applicants in each group; made-up toy data
decisions = (
[("A", True)] * 6 + [("A", False)] * 4 +
[("B", True)] * 3 + [("B", False)] * 7
)
def selection_rate(group):
rows = [selected for g, selected in decisions if g == group]
return sum(rows) / len(rows)
rate_a = selection_rate("A")
rate_b = selection_rate("B")
print(f"Group A: {rate_a:.0%}")
print(f"Group B: {rate_b:.0%}")
print(f"Ratio (lower / higher): {min(rate_a, rate_b) / max(rate_a, rate_b):.2f}")
A big gap, here 60% against 30%, is a warning sign that needs investigation. It does not prove unfairness by itself, because groups can differ in legitimate ways, and there are several competing definitions of fairness that cannot always be satisfied together. See fairness and its trade-offs.
What Are the AI Ethics Principles Behind Fair AI?
Most frameworks share a few ideas: fairness (do not treat groups unjustly), transparency (explain how and when AI is used), privacy, safety and accountability (a person or organisation is responsible for outcomes). The chapter on core ethical principles explains them with examples.
How Can You Reduce Bias?
Use data that represents the people the system will serve.
Test results separately for different groups, not only the overall average.
Keep a human in the loop for decisions that affect people's lives, with a way to appeal.
Document what the system is for, what it was tested on and where it fails.
Monitor after launch, because data and behaviour change over time.
Common Mistakes
Assuming "the data is objective". Data records human choices.
Deleting the sensitive column and calling it fair. Proxies can bring the bias back.
Trusting the average. A model can be accurate overall and poor for one group.
Over-trusting the output. People tend to accept an AI answer without question; read when to rely on AI output.
Key Takeaway
AI bias is unequal performance or treatment caused by data, design and use. You cannot see it in an overall average, so compare groups, keep humans accountable and document limits. Fairness is a habit you practise, not a switch you turn on.
Ready to start? You will learn to spot bias, ask the right fairness questions and make responsible decisions when you use or build AI. Begin with the free course AI Ethics and Responsible Use: Beginner.
Related articles
More in AI Ethics
What Not to Share With ChatGPT: 7 Privacy Rules
Seven simple privacy rules for using ChatGPT and other AI chatbots: what never to type, why it matters, and safer ways to get help.
3 min readNov 11, 2026How Reliable Are AI Detectors?
AI detectors guess from writing patterns and can flag honest human work. Learn why false positives happen and what students and teachers should do instead of trusting a score.
3 min readNov 5, 2026What Is a Deepfake and How Can You Spot One?
A deepfake is AI-generated or altered video, image or audio made to deceive. Learn the harms, the warning signs, the verification habits that work, and what to do if you are targeted.
4 min readNov 2, 2026Precision, Recall and F1 Score, Explained Simply
What precision, recall and F1 score mean, how to calculate them from a confusion matrix, and how to choose which one matters for your problem.
4 min readNov 16, 2026Training vs Validation vs Test Data
Why machine learning data is split into training, validation and test sets, what each is for, how to split it and how to avoid data leakage.
3 min readNov 13, 2026Classification vs Regression in Machine Learning
Classification predicts a category and regression predicts a number. See the differences, examples, metrics and a tiny Python demo of both.
4 min readNov 10, 2026
Comments
Sign in to join the discussion.

