What is algorithmic fairness?

At a glance

• Software learns from your past data, so it copies the biases already in that data, even with no instruction to discriminate^[1].
• There is no single definition of fair: you usually cannot satisfy every fairness measure at once, so you must choose which one fits the use case.
• Regulators treat biased algorithms as illegal discrimination, with real fines: the CFPB hit Apple and Goldman Sachs for a combined 70 million dollars in 2024.
• You are liable even when a vendor built the tool.

Why fair-seeming software discriminates

An algorithm has no opinions. It finds patterns in your data and repeats them at scale. If past hires or loans reflected old inequalities, it learns and applies them, even when the code never mentions race or gender. Bias hides in proxies like zip code or school that quietly track protected traits.

What it means for your business

The COMPAS case shows the trap: a risk tool flagged Black defendants as high-risk far more often than white ones, yet was equally accurate for both^[2] — meeting one fairness standard while failing another. So fairness is a deliberate choice, not a box a vendor checks. AI hiring in NYC requires an audited, published bias check^[3], and lending tools must follow fair-credit laws regardless of automation^[4]. Demand audit results and test outcomes across groups.

Bottom line

The software faithfully reproduces whatever bias your data carries, so assume nothing, audit outcomes across groups, and keep the records that prove you checked.