Fairness
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
The requirement that an AI system's decisions not disadvantage people based on characteristics like race, sex, or age, measured by competing metrics.
What is Fairness?
Fairness in AI is the requirement that an automated system’s decisions not disadvantage people because of characteristics such as race, sex, age, or disability, whether the disadvantage is deliberate or a by-product of training.
The requirement sounds simple. It is not: researchers have formalized fairness into several distinct metrics and proved that reasonable ones conflict, so choosing which to satisfy is a policy judgment dressed as a technical one.
Types of Fairness
- Individual fairness: people who are alike in the ways relevant to the decision receive alike decisions.
- Group fairness: outcomes or error rates are comparable across demographic groups, whether as equal approval rates or equal false positive and false negative rates.
- Calibration: a risk score means the same thing regardless of group; a “70% risk” label corresponds to the same real-world probability for everyone.
These definitions cannot all hold at once. Researchers proved in 2016 that when two groups differ in their underlying rates of the predicted outcome, no score can be calibrated for both groups while also equalizing their error rates. Every deployed system therefore embodies a choice about which unfairness it tolerates.
Fairness vs Bias
Bias is the cause; unfairness is the effect. Bias names the systematic tilt in data, design, or model behavior; fairness is the property you assess at the output, where decisions land on people.
Diagnosing a fairness failure means tracing it back to the bias that produced it, which is why the two terms travel together and why fixing the metric without fixing the data rarely holds.
Example of Fairness
COMPAS, a risk-assessment tool some US courts used to estimate a defendant’s likelihood of reoffending, became the defining case. In 2016, ProPublica analyzed its scores and reported that among defendants who did not go on to reoffend, Black defendants had been labeled high-risk at roughly twice the rate of white defendants.
The developer responded that the scores were calibrated: a given score corresponded to about the same reoffending rate regardless of race. Both readings were consistent with the data.
They applied different fairness definitions, and the impossibility result explains why the tool could not satisfy both given differing base rates in the data. The dispute did not settle which definition is right; it established that the choice is a value judgment courts and vendors had been making without noticing.
Related AI terms: Bias · Ethics in AI · Trustworthy AI · Explainability
Did you like the Fairness gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About