algorithmic bias

basicwidespread

Algorithmic bias is when a computer system consistently and unjustifiably treats certain groups of people more disadvantageously than others.

In plain terms

Imagine a chef who learned to cook from only one grandmother's recipes. They become skilled, but salt everything exactly as she did, without noticing. Artificial intelligence learns similarly from given examples: if the examples are one-sided or preserve old injustices, the system faithfully repeats them.

Bias is not a single random error but a systematic shift: the machine repeatedly makes mistakes in the same direction. For example, it consistently assigns lower scores to women, older people, residents of a specific area, or members of a minority.

It is important to know this because such systems now screen resumes, score loan applications, and rank medical reports. Human bias appears in individual cases handled by an officer, while machine bias occurs simultaneously in thousands or millions of decisions, often invisibly.

How it works

The system learns from historical data, so it also learns past inequalities: if a company previously hired mostly men, the model sees this pattern as a 'success' indicator.
It is also a problem if certain groups are barely represented in the data: the system knows less about them and thus makes more errors.
Designers' decisions also matter: what they measure, what they consider a good outcome, and what they use as a proxy for something (e.g., healthcare spending instead of disease severity).
Even if sensitive data (gender, age, origin) is omitted, other data can reveal it: the postal code, the school name, or shopping habits.
Bias can become self-reinforcing: the machine's decisions generate new data, and the next version learns from these.

An everyday example

You submit an application for a job, and your resume is first scored by a screening program. If the program learned from the company's previous hires, where there were few people over fifty or women with young children, it may rank yours lower. Meanwhile, no one looked at your skills.

Who is affected

jobseeker, new graduateResume screening and video interview analysis programs can filter you out before a human sees your application. → Work & HR

bank customer applying for a loanCredit scoring may unfairly penalize you based on your residence or other traits that reveal nothing about your ability to repay.

patient, doctorA diagnostic system trained mainly on data from one type of patient group may be less accurate for others. → Health

developerYou have to check the training data and the results group by group. For high-risk systems, the EU AI Act explicitly requires that possible biases in the training data be examined and addressed. → Code

A common misconception

MythMachines only calculate, so they cannot be biased and are therefore more objective than humans.

In realityThe machine works from data collected by people and goals set by people, so it adopts human biases. Moreover, numbers create an appearance of objectivity, making them harder to dispute.

Limits and risks

Bias cannot be eradicated once and for all; it can only be measured, reduced, and regularly re-evaluated.
There are multiple mathematical definitions of a 'fair' decision. In most real-world situations, these cannot be met simultaneously, so someone must make a value-based choice.
From the outside, you mostly cannot see whether a decision was made by a biased system, so it is important to be able to request explanations and human review.
Large language models, and the agents built on them, also carry biases, so do not automatically accept their answers and decisions as neutral either.

Similar, but not the same

hallucination – A hallucination is a fabricated, false statement, while bias is a systematic, one-sided unequal treatment that can even be based on accurate data.
statistical bias – In statistics, bias is a neutral technical term for systematic deviation in estimation, whereas algorithmic bias refers to unfair consequences for people.
Discrimination – Discrimination is a legal concept and human or institutional behavior, while algorithmic bias is a system property that can lead to discrimination.

Related concepts

Broader

MI-etika, trustworthy artificial intelligence, Responsible AI

Narrower

data distortion, Sampling bias, gender bias

On the edge

algorithmic fairness, Automation bias, Explainable AI, training data, transparency, human oversight

A brief history

The phenomenon was systematically described by Batya Friedman and Helen Nissenbaum in 1996 in the study "Bias in Computer Systems." It became widely known in 2016 when ProPublica's article "Machine Bias" showed that the COMPAS risk assessment program used in US courts incorrectly classified Black defendants as high risk more frequently.

Where things stand

By the end of 2026, bias testing is increasingly a standard part of serious AI development. The EU Artificial Intelligence Act (Regulation 2024/1689) requires high-risk systems to examine and address possible biases in training data. These obligations are becoming applicable in stages. Measurement methods are not yet uniform in practice, and the autonomous decisions of agents create a new oversight task.

Status: 2026 November

In the news

Appears in 1 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Two federal courts dismiss AI and keyword hiring bias claims; Sirius XM also wins dismissalOct 27, 2026 · Work & HR

Also known as

algorithmic bias · algorithmic bias · algorithmic bias · AI bias · machine bias · algorithmic bias · AI bias · machine bias

Sources

NIST SP 1270: Towards a Standard for Identifying and Managing Bias in Artificial Intelligence

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.