What is AI bias?
AI bias is when an AI system consistently gives unfair or less accurate results for some people or groups compared to others.
AI does not have opinions of its own. It learns patterns from data created by humans, and human data carries human unfairness. If a model learns from a world where some groups were treated unfairly, it can quietly repeat that treatment at a much larger scale.
Bias can show up in many places: who gets shortlisted for a job, whose face a camera recognises, whose voice an assistant understands, or which images a generator shows when you type "doctor" or "engineer".
Bias is different from AI hallucination. Hallucination is a model making things up; bias is a model being systematically unfair, even when every answer sounds confident and correct.
What is AI bias mainly caused by?
AI bias is mainly caused by biased or unrepresentative training data, though labelling and design choices also play a big part.
- Unrepresentative training data: if most photos in a dataset show lighter skinned faces, the model will be less accurate on darker skinned faces.
- Historical bias: data that records past unfair decisions, such as old hiring or loan records, teaches the model to repeat them.
- Labelling bias: the people tagging data bring their own assumptions, for example marking certain accents or dialects as "unclear".
- Design choices: what the model is asked to predict, which features it uses and how success is measured. Using a proxy, such as past spending to predict who needs care, can hide unfairness.
- Feedback loops: a biased system's outputs are used as new data, which makes the bias stronger over time.
- Lack of diverse testing: if a team only tests on people like themselves, problems for other groups go unnoticed.
What are the types of bias in AI?
The main types of bias in AI are data bias, historical bias, labelling bias, algorithmic bias and interaction bias.
| Type | What it means | Simple example |
|---|---|---|
| Data or sampling bias | Some groups are missing or underrepresented in the data | A speech model trained mostly on city English speakers |
| Historical bias | Data reflects past unfairness in society | Old hiring records where most hires were men |
| Labelling or measurement bias | Labels or measurements are inaccurate for some groups | Inconsistent tags given by different annotators |
| Algorithmic or design bias | The model's goal or features create unfair outcomes | Using pin code as a feature that acts as a stand-in for community |
| Interaction or feedback bias | User behaviour and past outputs reinforce the bias | Recommendations that keep showing the same kind of content |
Real cases usually mix several types at once, which is why fixing bias takes more than one step.
What are some examples of AI bias in real life?
Well documented examples of AI bias include face analysis tools that were less accurate for darker skinned women, a hiring tool that downgraded women, and a healthcare algorithm that underestimated some patients' needs.
- Face analysis: the 2018 Gender Shades study by Joy Buolamwini and Timnit Gebru found that commercial gender classification systems made far more errors on darker skinned women than on lighter skinned men.
- Hiring: Reuters reported in 2018 that Amazon scrapped an experimental resume screening tool after it learned to penalise resumes that mentioned the word "women's", because it was trained on a decade of mostly male applications.
- Healthcare: a 2019 study in Science found that a widely used US hospital algorithm used healthcare spending as a stand-in for health needs, which led it to underestimate how sick Black patients were.
- Image generators: many studies and users have noticed that generators often show stereotyped results, such as mostly men for "CEO", unless the prompt says otherwise.
In an Indian setting, imagine a loan app trained mostly on salaried city borrowers. It may unfairly reject farmers or gig workers who repay well but look different on paper.
How can AI bias be reduced?
AI bias can be reduced by using more representative data, testing the model separately for different groups, and keeping humans in charge of high stakes decisions.
- Collect diverse data: include different genders, regions, languages, ages and income groups.
- Audit labels: write clear labelling rules and check that annotators agree.
- Measure fairness: compare accuracy and error rates across groups, not just one overall score.
- Remove or question proxy features that secretly stand in for caste, religion, gender or region.
- Use human review for hiring, lending, healthcare and policing decisions.
- Add safety layers: techniques like RLHF and AI guardrails help chatbots avoid stereotyped or harmful answers.
- Publish and monitor: document what data was used, and keep checking results after launch.
Fairness has several definitions, and they can conflict. Equal accuracy for every group and equal approval rates for every group are not always possible together, so teams have to choose and explain their choice.
I tell students to ask one question about any AI system: who might this work badly for? That single question catches more bias than any fancy tool.
What is AI bias for Class 9 and Class 10 students?
For Class 9 and Class 10 AI, AI bias means an AI system favouring or ignoring certain people because the data or the people who built it were one-sided.
An easy example to write in your exam: a smart speaker trained mostly on voices of adult men from cities may understand women, children or people with regional accents less well. The machine is not trying to be unfair; it simply learned from data that left those voices out.
Key points to remember:
- AI bias mainly occurs due to biased or incomplete training data.
- Bias can also come from the developers' own assumptions.
- To reduce it, collect diverse data and test the AI with many kinds of users.