Bias

Systematic skew or error in data, modeling, or decisions that can distort results or create unfair outcomes.

Whose preferences did the branch survey measure?

A fictional library posts an online poll about extending digital services. Of 100 replies, 80 favor the change. The count is correct; the interpretation needs care.

Who could answer?

The link appeared on the library website. Regular web users had more opportunities to see it than patrons who rely on the branch telephone.

What can be reported?

Eighty percent of respondents favored the change. Estimating support among all patrons requires a plan for how respondents represent that wider population.

Look for the repeated direction of error

Sampling more people through the same channel could make the estimate more precise for web respondents while leaving the coverage gap in place. Offer another response route and examine how the sampling and wording affect the answer before using the poll to set priorities.

What Bias Looks Like

Bias may appear as worse performance for certain groups, distorted recommendations, unbalanced search results, unfair moderation, or decision systems that reproduce existing inequalities. Sometimes bias is obvious. Other times it hides behind average performance that looks acceptable until results are broken down by subgroup or real-world context.

Bias in AI should not be confused with the mathematical bias term used inside models. The fairness sense of bias is about systematic distortion in outcomes, representation, or treatment.

Follow the idea

Algorithmic bias can carry a skewed signal into repeated decisions.

Bias mitigation includes changing how the evidence is collected.

Examples in the AI guides

Source · NIST guidance on identifying and managing bias

Return to Bias in the A–Z glossary