What is an AI Bias?

Bias enters an AI system primarily through its training data: if historical hiring decisions favored certain demographics, a model trained to mimic that pattern will reproduce the same skew, even without anyone explicitly encoding a discriminatory rule. Bias can also emerge from how a task is framed or how success is measured, a model optimized purely for engagement, for instance, can learn to favor sensational content regardless of accuracy. Mitigating bias requires deliberate, ongoing work rather than a one-time fix: auditing training data and outputs for disparate impact across groups, testing model behavior on edge cases that reflect underrepresented populations, and building human review into decisions that materially affect people, like lending, hiring, or healthcare. Because bias can be subtle and context-dependent, most rigorous approaches treat it as a continuous monitoring problem rather than something eliminated once and then ignored.