Bias
Systematic errors in AI outputs resulting from prejudiced training data or flawed algorithms, leading to unfair outcomes.
Definition
Persistent, directional deviations in model predictions that systematically disadvantage (or advantage) certain groups or cases. Bias arises from unbalanced data, labeler prejudice, or mis-specified objectives. Effective governance requires detecting, quantifying (e.g., via fairness metrics), and tracing bias sources to remediate both data and model design.
Real-World Example
A hiring-screening AI trained on historical resumes rejects applicants from a particular university because past hires predominantly came from other schools. HR discovers this bias, augments its dataset with more graduates from the affected university, retrains the model, and monitors acceptance rates to ensure parity across alma maters.
.svg)
