AI bias usually comes from the data a model learned from, the way a problem was framed, or how results are measured and used. It can appear in hiring, lending, policing, medical tools and image generators. Because models operate at scale, a small bias can affect many people quickly.
How it works
Bias can enter at every stage. In data collection, some groups may be under-represented or recorded differently; face recognition systems trained mostly on lighter-skinned faces have been shown in published research to make more errors on darker-skinned faces. Historical data can encode past discrimination: a model trained on previous hiring decisions may learn to prefer the kinds of candidates who were hired before. Labels added by people carry their assumptions. Proxy features matter too: even if a model never sees ethnicity, postcode or school can stand in for it. Generative models absorb stereotypes from internet text and images, so prompts like 'a CEO' may skew towards one gender or ethnicity. Bias is measured by comparing error rates or outcomes across groups, and fixing it involves better data, testing across groups, adjusting training, adding guardrails and human review. Researchers have shown that different mathematical definitions of fairness can conflict, so teams have to choose which trade-offs suit the situation.
Why it matters
Biased AI can harm people, damage trust and break laws on discrimination and data protection. For businesses using AI to screen applicants, set prices or moderate content, unchecked bias is a legal and reputational risk. For creators, biased image or voice tools can produce stereotyped or exclusionary content that audiences notice. Knowing where bias comes from helps you write better prompts, test outputs across different groups and ask vendors the right questions.
Where you will see it
- Auditing a hiring tool's pass rates across different applicant groups
- Testing an image generator's outputs for stereotypes before publishing campaigns
- Reviewing a lending model for proxy features such as postcode
- Adding diverse examples to training data for a speech recognition system
Limits and common misconceptions
- Removing sensitive fields such as gender does not remove bias, because other data can act as proxies.
- There is no single fairness measure; improving one can worsen another.
- Bias is not only a data problem; design choices and how outputs are used matter too.
How it differs from related terms
- Bias (in AI) vs Training data Training data is a common source of bias, but bias also comes from problem framing, labels and deployment choices.
- Bias (in AI) vs Guardrails Guardrails are controls added to limit harmful outputs; they can reduce visible bias but do not fix its underlying causes.
Frequently asked questions
What causes bias in AI?
The main causes are unbalanced or historical training data, human labelling choices, features that act as proxies for protected traits, and how the problem and success measures are defined. Deployment matters too: using a model on people unlike those in its training data can create unfair results.
Can AI bias be removed completely?
Not completely, but it can be reduced and managed. Teams test performance across groups, improve data coverage, adjust training, add review steps and monitor results after launch. Because fairness definitions can conflict, the goal is usually to make trade-offs deliberate, documented and appropriate to the use.
Related terms
- Artificial intelligence (AI) The field of building computer systems that perform tasks usually requiring human intelligence, such as understanding language, recognising images, predicting and generating content.
- Benchmark A standardised set of tasks with a fixed scoring method, used to measure and compare how well AI models perform a skill.
- Chatbot Software that holds a conversation in natural language, by text or voice, to answer questions, complete tasks or provide information.
- Computer vision The field of AI that enables computers to interpret images and video, identifying objects, people, text, movement and scenes.
- Context window The maximum amount of text, measured in tokens, that a language model can take in and work with in a single request, including its reply.