AI Bias

No jargon, no math

A model learns statistical patterns from its training data, nothing else. If the data under-represents a group or encodes a stereotype, the pattern inherits the skew and repeats it at scale, politely and consistently. The usual hiding places: CV screening trained on a decade of past hires that rarely included certain groups; a fraud filter that learns postcode as a proxy for risk; a demand forecast that ignores your smallest depots because the history did. Proxies are the sneakiest: the model never sees gender or ethnicity, it sees address, school or a career gap, and draws the same line anyway. Then feedback loops lock it in: if the biased system says no, the next dataset holds even fewer examples of yes. You catch bias by testing outputs across groups before deployment (a bias audit), retraining or re-weighting data, routing high-stakes decisions through human review, and re-testing after every model update. The EU AI Act classifies many of these uses as high-risk for exactly this reason.