Bias in Computing Systems

You’ll be able to

Explain how bias can be embedded in algorithms and their training data
Describe real consequences of biased automated decisions
Identify ways bias can be detected and reduced

Algorithms are not automatically neutral

It is tempting to assume a computer decision is objective because "the math is neutral." But algorithmic bias is real: computing systems can produce systematically unfair results for certain groups. Bias enters through the data a system learns from, through the choices its designers make about what to measure and optimize, and through the assumptions built into the problem itself. Because the output looks precise and impartial, biased systems can be more dangerous than biased humans — their unfairness is hidden behind an appearance of objectivity and applied at massive scale.

Bias from the data

Many modern systems learn patterns from historical training data. If that data reflects past human bias or an unrepresentative sample, the system learns and repeats the bias — sometimes amplifying it. A hiring tool trained on a company’s past hires, who were mostly one demographic, may learn to favor that demographic and reject equally qualified others. The system is not "malfunctioning"; it is faithfully reproducing the bias baked into its data. "Garbage in, garbage out" applies to fairness, not just accuracy.

Consequences and remedies

Biased systems cause real harm when they gate opportunities: loan approvals, job screening, medical triage, and criminal-justice risk scores have all shown bias against protected groups. Because the harm is systematic and automated, it scales to millions of decisions. Reducing bias requires deliberate effort: auditing systems for unequal outcomes across groups, using more representative data, involving diverse teams who are likely to spot problems, and being transparent about how decisions are made. Bias is rarely removed by accident — it must be tested for.

How bias enters a system

biased or unrepresentative data + designer assumptions → biased algorithm → unfair outcomes at scale

The appearance of objectivity can hide the bias, which is why systems must be explicitly audited for unequal outcomes.

Worked example

A bank uses an algorithm trained on decades of past loan decisions to approve or deny applicants. Explain how this could produce biased results and how the bank might respond.

1.The training data records past decisions, which may reflect historical discrimination against certain neighborhoods or groups.
2.The algorithm learns the patterns in that data, including the discriminatory ones, and reproduces them in new decisions.
3.Because it looks like an objective computer decision, the bias is hard to notice and is applied to every applicant automatically.
4.The bank should audit approval rates across groups, retrain on more representative data, and add human review and transparency to catch and correct unfair outcomes.

Answer: The algorithm can inherit and repeat the discrimination hidden in its historical training data, producing biased loan decisions at scale behind a veneer of objectivity. The remedy is to audit for unequal outcomes, use more representative data, and add transparency and oversight.

Checkpoint

A résumé-screening algorithm is trained on a company’s past hiring decisions, in which one group was rarely hired. The tool now rejects qualified applicants from that group. What is the most accurate explanation?

Watch out

Never assume a computer’s decision is unbiased just because it is automated. A system trained on biased data will reproduce that bias while looking perfectly objective — which makes it harder to detect.

Checkpoint

Which action is most effective for detecting bias in an automated decision system?

On the exam

Two exam essentials on bias: (1) bias often comes from the data a system was trained on, and (2) it must be tested and audited for — it does not disappear because a computer made the decision.

Answer the 2 checkpoints as you read.