Bias in AI: Real Examples and How to Reduce It

01, Oct 2023

Bias in AI happens when an artificial intelligence system produces results that are systematically unfair to certain people or groups. It is not usually caused by anyone deliberately programming prejudice. More often, the system learns patterns from data that already reflect inequality, or it is tested on a narrow group of people and then used on everyone. This guide explains how AI bias arises, gives well-documented types and examples, and sets out practical steps to reduce it.

What AI bias means

An AI model learns from examples. If those examples are skewed, incomplete or shaped by past unfair decisions, the model learns those patterns too, and can apply them at a scale and speed no person could. Google's machine learning course describes several common ways this happens in its module on fairness in machine learning.

Bias matters most when AI influences decisions about people: who gets a job interview, a loan, an insurance price, a medical referral, or extra police attention. But it also shows up in everyday tools, like image generators that default to stereotypes.

Common types of AI bias

Historical bias

The data accurately reflects the past, but the past was unfair. A model trained on years of hiring decisions may learn to favour the kinds of candidates who were hired before, even if those decisions were biased.

Representation or sampling bias

Some groups are under-represented in the training data. Systems trained mostly on images of lighter-skinned faces, or on speech from a few accents, tend to perform worse for everyone else.

Measurement bias

The thing the model predicts is a poor stand-in for what really matters. Using past healthcare spending as a proxy for health needs, for example, can understate the needs of groups who historically had less access to care.

Automation bias

People trust an automated system too much and stop questioning its outputs. Even a fairly accurate model can cause harm if humans rubber-stamp its decisions.

Real-world examples

Facial recognition: independent testing, including by the US National Institute of Standards and Technology, has found that many face recognition algorithms had higher error rates for some demographic groups than others.
Recruitment tools: there have been widely reported cases of experimental hiring systems that learned to downgrade CVs associated with women because the historical data was male-dominated.
Speech recognition: transcription systems have often been less accurate for certain accents and dialects, including many African and regional English accents.
Image generators: prompts like "a CEO" or "a nurse" can produce stereotyped results unless the system has been adjusted.

These examples show that bias is not just a theoretical risk. It affects accuracy, trust and, in some cases, people's rights.

Why bias is hard to remove

Fairness has several definitions, and some of them cannot all be satisfied at once. Our guide to fairness in machine learning explains these trade-offs.
Removing a sensitive field is not enough. Other data, like postcode or school, can act as a proxy for race, religion or income.
Bias can appear after launch when a system is used on a different population than it was tested on.

How to reduce bias in AI

Check the data. Ask who is represented, who is missing and what historical decisions are baked in.
Test across groups. Measure accuracy and error rates separately for different groups, not just overall.
Involve affected people. Users and communities often spot problems developers miss.
Keep a human in the loop. Decisions that affect people should be reviewed by someone who can override the system.
Document and monitor. Record how the system was built and keep testing after launch. The NIST AI Risk Management Framework includes guidance on managing bias as part of overall AI risk.
Give people a way to appeal. If someone is affected by an AI-supported decision, they should be able to question it.

What everyday users can do

You do not need to build AI to be affected by its bias. When you use generative tools:

Be specific in prompts. When creating images with an AI image generator, describe the people you want to see rather than relying on defaults.
Review AI-written content for stereotypes or one-sided assumptions before publishing.
Ask an assistant like TextPixAI AI Chat to suggest alternative perspectives on a topic, then check them yourself.

Bias in generative AI

Generative tools raise their own bias questions. Language models may describe some nationalities, religions or genders with stereotypes, give lower-quality answers in less-represented languages, or assume a Western context by default. Image generators may show a narrow range of skin tones, body types or settings for neutral prompts. Developers reduce these patterns through data curation, testing and safety tuning, but they cannot remove them completely. Google's published AI Principles are one example of a developer committing to avoid creating or reinforcing unfair bias. As a user, specific prompts and careful review are your best defence, especially for content that represents your customers or community.

Reducing bias is ongoing work rather than a one-time fix, so build regular checks into how you use AI.

Key takeaways

Bias usually comes from data and design, not intent.
Test results separately for different groups.
Keep humans able to review and override decisions.
Specific prompts reduce stereotyped generative outputs.

Related guides

Keep reading in this topic cluster:

Frequently asked questions

Can AI ever be completely unbiased?

Probably not entirely, because data and design choices always reflect some perspective. The realistic goal is to measure bias, reduce harmful bias and keep humans accountable.

Who is responsible for AI bias?

The organisations that build and deploy AI systems are responsible for testing and managing bias. Regulators increasingly expect this, especially for high-risk uses.

Is bias in AI illegal?

Discrimination laws apply whether or not a decision is made by AI. In many countries, using a biased system to make unfair decisions about people can break existing law.

How can I tell if an AI tool is biased?

Try varied inputs and compare results. Look for published testing information from the provider, and be cautious if none exists.

Bias in AI: Real Examples and How to Reduce It