ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›AI MEANINGS›WHAT IS AI BIAS?
AI MEANINGS · MEANINGBEGINNER

What is AI bias? When an AI is skewed against some people

AI bias means an AI system that works better for some people than others or repeats stereotypes. Where it comes from, a real example, and why it matters.

READ6 min
WORDS1,114
SECTIONS7
TYPEMEANING
CHECKED1 OCT 26

AI bias means an AI system that is skewed: it works better for some people than others, or repeats unfair stereotypes. It usually comes from the data it learned from and from the choices of the people who built and use it, not from any intention of the machine.

◈ IN PLAIN TERMS

Imagine a shoe shop that only ever measured feet from one town. Its sizes fit that town well. Everyone else gets shoes that pinch, and nobody planned it that way. AI bias often works like that: the system fits the people it saw most.

In a sentence: “The photo app had an AI bias problem: it found faces instantly in bright light skin but kept missing darker faces, because few were in its training pictures.”

What the word means

The plain point: bias here means unfair favouring, built into a system.

Google’s machine learning glossary defines this sense of the word as “Stereotyping, prejudice or favoritism towards some things, people, or groups over others.”1

NIST, the US government’s standards agency, stresses that data is only part of the story. Human and institutional habits feed in too.2

HOW BIAS GETS INTO AN AI · NO ONE HAS TO INTEND IT
Like the shoe shop that only measured one town’s feet.
Skewed dataThe training data shows some groups far more than others.
Human choicesHow people build and use the system feeds in too.
A lopsided fitThe system works best for the people it saw most.
Unequal mistakesErrors fall harder on some groups, in real decisions.
Reasoning, 1 Oct 2026 — summarises this page’s sections on what the word means (Google and NIST) and the face-photo example.

Where the word comes from

Worry about unfair computer systems is older than today’s AI. In 1996 the researchers Batya Friedman and Helen Nissenbaum set out what “bias” means for a computer system: one that treats some people unfairly, consistently, in favour of others.3 “AI bias” applies the same idea to systems that learn from data.

A simple example

The plain point: a system trained mostly on one group made far more mistakes on another.

A 2018 study tested commercial systems that guess gender from a face photo. The researchers found that common face datasets were mostly of lighter-skinned people, and that the systems made far more mistakes on darker-skinned women.

GroupHighest error rate foundWhat it shows
Darker-skinned womenUp to: 34.7%Shows: the group the systems got wrong most often
Lighter-skinned menUp to: 0.8%Shows: the group the systems got right almost every time

4

Those systems are old now. The pattern is the lesson: whoever is missing from the training data tends to pay for it.

Why it matters to you

The plain point: biased AI can affect real decisions about real people.

  • Big decisions. NIST says AI can make decisions that affect whether a person gets into a school, gets a bank loan or is accepted as a renter.5
  • Tools that judge your writing. One study found AI-writing detectors wrongly flagged more than half of essays by non-native English speakers.6 AI on job applications covers what that means for applicants.
  • Everyday answers. If a chatbot’s examples or pictures all look alike, ask for a wider range and check who is left out.
TAKEAWAY

If an AI tool helps decide something about a person, ask who it was tested on and whether a human checks its calls.

Often confused with

  • Bias, the number: inside a model, a “bias” is also a kind of parameter, a plain number with no fairness meaning. See parameters meaning.
  • Agreeing with you: chatbots also tend to side with the person asking. That is a different slant, called sycophancy. See how AI influences you back.
  • Your own bias: people lean too, and AI can make that easier to indulge. See your own blind spots.
  • Making things up: a wrong fact stated confidently is a hallucination. Bias is a pattern of errors that falls harder on some groups.

Questions people ask

What are examples of AI bias?

A 2018 study found commercial face-analysis systems were far less accurate for darker-skinned women than for lighter-skinned men.4 Another found AI-writing detectors wrongly flagged more than half of the essays by non-native English writers as machine-written.6

Can AI bias be fixed?

It can be reduced, but not by fixing the data alone. The US standards body NIST says biased data is only part of the story; how a system is designed, used and checked matters too.2

Learn more on this site

QUICK CHECK · 3 QUESTIONS

Pick an answer to see if you got it. Nothing is recorded.

1. Where does AI bias usually come from?

  • The machine deciding on purpose to be unfair
  • The data it learned from and the choices of the people who built and use it
  • A fault in the screen you view it on
Show the answer

Answer: The data it learned from and the choices of the people who built and use it. AI bias usually comes from the training data and from people’s choices in building and using the system, not from any intention of the machine.

2. What did the 2018 study of face-photo systems find?

  • They made the most mistakes on lighter-skinned men
  • They worked equally well for every group
  • They made far more mistakes on darker-skinned women
Show the answer

Answer: They made far more mistakes on darker-skinned women. The study found the systems got darker-skinned women wrong far more often, while they got lighter-skinned men right almost every time.

3. If an AI tool helps decide something about a person, what does the page suggest you ask?

  • Who it was tested on and whether a human checks its calls
  • How many people have downloaded it
  • Which company designed its logo
Show the answer

Answer: Who it was tested on and whether a human checks its calls. The page’s takeaway is to ask who the tool was tested on and whether a human checks its decisions.

Related terms: AI winter · Alignment · Attention · Turing test · Vision-language model (VLM) · World model

SOURCES
  1. 1Google for Developers, Machine Learning Glossary: ML Fundamentals, undated glossary, read at source 1 Oct 2026bias (ethics/fairness) is “Stereotyping, prejudice or favoritism towards some things, people, or groups over others.” ↑
  2. 2NIST, There’s More to AI Bias Than Biased Data, NIST Report Highlights, 16 Mar 2022, read at source 1 Oct 2026“While these computational and statistical sources of bias remain highly important, they do not represent the full picture.” ↑a“While these computational and statistical sources of bias remain highly important, they do not represent the full picture.” ↑b
  3. 3Friedman and Nissenbaum, Bias in Computer Systems, ACM Transactions on Information Systems 14(3), July 1996, read at source 12 Oct 2026“we use the term bias to refer to computer systems that systematically and unfairly discriminate against certain individuals or groups of individuals in favor of others.” ↑
  4. 4Buolamwini & Gebru, Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification, Proceedings of Machine Learning Research 81, 2018, read at source 1 Oct 2026datasets were “overwhelmingly composed of lighter-skinned subjects (79.6% for IJB-A and 86.2% for Adience)”; “darker-skinned females are the most misclassified group (with error rates of up to 34.7%). The maximum error rate for lighter-skinned males is 0.8%.” ↑a“darker-skinned females are the most misclassified group (with error rates of up to 34.7%). The maximum error rate for lighter-skinned males is 0.8%.” ↑b
  5. 5NIST, There’s More to AI Bias Than Biased Data, 16 Mar 2022, read at source 1 Oct 2026“AI can make decisions that affect whether a person is admitted into a school, authorized for a bank loan or accepted as a rental applicant.” ↑
  6. 6Liang et al., “GPT detectors are biased against non-native English writers”, Patterns, July 2023, as cited on this site’s page on AI and job applications, read at source 17 Sep 2026 in the full text on Europe PMC: “they incorrectly labeled more than half of the TOEFL essays as ‘AI-generated’ (average false-positive rate: 61.3%).”↑a↑b

Liang, Yuksekgonul, Mao, Wu & Zou, GPT detectors are biased against non-native English writers, Patterns 4(7), July 2023, read at source 17 Sep 2026.

The comparison on this page (a shoe shop that measured only one town) is this site’s own way of explaining, not a claim from these sources.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE