A hiring tool quietly ranks certain resumes lower, and nobody programmed it to discriminate against anyone. A lending model approves fewer applicants from one zip code than another. No explicit rule about location exists anywhere in it. Both cases get called algorithmic bias. Neither involved a person deciding to be unfair on purpose.

What does “algorithmic bias” mean?

Algorithmic bias means a computer system produces results that are systematically unfair to certain groups. No explicit intent to discriminate needs to exist. The system wasn’t told to treat anyone differently. It just learned patterns from data that already contained unequal treatment, then reproduced those patterns at scale. A single wrong prediction isn’t bias. The term applies when the errors form a consistent pattern along lines like race, gender, age, or income.

Where did the term come from?

Concerns about biased automated decisions go back decades, but the phrase gained wide public use in the 2010s. Researchers and journalists began publishing case studies of specific systems producing unequal outcomes. Facial recognition tools that performed worse on darker skin tones. Hiring algorithms that downgraded resumes with certain names. Each case built the same argument. Bias doesn’t require a biased programmer, since it can emerge from the data itself.

How does algorithmic bias actually work?

Most modern systems learn by studying historical examples, not by following hand-written rules. If those historical examples reflect decades of unequal treatment, the system learns those patterns as if they were simply how the world works. Picture a hiring model trained mostly on past successful hires from one demographic. It will often rate similar candidates higher later, even though nobody wrote a rule favoring that group directly.

Bias can also creep in through what data gets collected in the first place. A dataset built mostly from one population produces a system that performs well for that population. It performs poorly for everyone else. This happens constantly with technology that seems neutral on its surface. The system just optimizes for whatever pattern the training data actually contains.

For the term to apply, the unfair outcome needs to be systematic rather than random. Every prediction system makes mistakes sometimes. Algorithmic bias describes mistakes that cluster consistently around a particular group, not scattered errors spread evenly across everyone.

A concrete example

A company deploys an automated tool to screen job applications before a human ever reviews them. Over several months, the tool consistently ranks applicants from certain universities and neighborhoods lower than others. Their qualifications look comparable on paper. An internal audit later traces the pattern back to the training data. It was built almost entirely from a decade of past hiring decisions that already leaned in that direction.

What it’s not

Algorithmic bias isn’t the same as an occasional bad prediction. That happens even in well-designed systems that aren’t biased at all. It’s also not identical to intentional discrimination, where someone deliberately programs unfair treatment into a system. And it isn’t a purely technical bug either, since the system usually works exactly as designed. The unfairness comes from what it learned, not from a coding error somewhere in the system.

Where you’ll encounter it

The term comes up constantly in discussions of hiring software, lending decisions, facial recognition, and criminal justice risk assessments — any place a model quietly stands in for a human judgment call. It also shows up in broader debates about AI regulation. Lawmakers increasingly want systems audited for these patterns before deployment, since the decisions involved can affect people’s lives directly.

Related terms

• Doxxing (L22)
• Cancel Culture (L23)
• Shadow Banning (L24)