Algorithmic Bias

Algorithmic bias refers to a systematic, repeatable skew in a model's outputs that disadvantages a particular group — as distinct from random error, which affects outcomes unpredictably rather than in a consistent direction. The distinction matters legally: several state AI laws, including Colorado's original SB 24-205, defined "algorithmic discrimination" in terms of this kind of systematic disparity in treatment or impact based on a protected characteristic.

Where it comes from

Bias typically enters a system through one of three channels:

• Training data that reflects historical patterns of discrimination — a hiring model trained on ten years of a company's past hiring decisions will learn whatever bias was present in those decisions.
• Feature selection — proxies for protected characteristics (ZIP code standing in for race, for instance) can reintroduce bias even when the protected characteristic itself is excluded.
• Optimization choices — a model tuned purely for overall accuracy can still produce large disparities in error rates between subgroups if those subgroups are underrepresented in the training data.

Why it's a compliance issue, not just a technical one

Algorithmic bias becomes a legal exposure the moment a biased system makes or materially influences a consequential decision — about employment, credit, housing, insurance, or healthcare. This is exactly the boundary that both the EU AI Act's high-risk categories and U.S. state laws like Colorado's are built around. See high-risk AI system for how that boundary gets drawn.

How it gets tested for

Bias testing generally compares outcome or error rates across protected groups — approval rates, false-positive rates, error rates — against a chosen fairness metric. There is no single universally correct fairness metric; different metrics can conflict with each other mathematically, which is why a defensible governance process documents which metric was chosen and why, rather than asserting the system is simply "unbiased."