AI bias: when machines repeat our blind spots

An error that isn’t random

Every AI gets things wrong sometimes. We call it bias when the errors are not random but keep hitting the same groups, or when the AI repeats clichés. A face recognition system that fails more often depending on skin tone, a résumé screening tool that rejects certain profiles more often, an image generator that almost always draws a “doctor” as a man: these are biases.

The most cited example is the Gender Shades study by Joy Buolamwini and Timnit Gebru, published in 2018 (Gender Shades paper). It measured that three commercial face analysis systems performed far worse on darker-skinned women than on lighter-skinned men. One cause it pointed to: benchmark datasets made up overwhelmingly of lighter-skinned faces.

Where bias comes from

• Training data that is lopsided or reflects past inequalities.
• Human choices: what gets measured, what counts as a “good” answer, which cases nobody thought to test.
• Use: a tool designed for one context, then applied to a different population.

Why it matters to you

In daily life, bias can be subtle: a translation that always defaults to the masculine, a job ad draft that leans on stereotypes. At work, the stakes are higher if AI helps hire, grade or allocate support. Before handing a decision to a tool, ask who might be disadvantaged and plan for a human check. The Can I trust an agent with this? tool helps you weigh that kind of risk.

In Europe, the EU AI Act and the GDPR regulate these uses, and data protection authorities publish guidance for developers.

Common misconceptions

“A machine is objective by definition.” It reflects its data and the choices of the people who built it.

“Just remove gender or ethnicity from the data.” Other details (first name, zip code, career path) can play the same role indirectly.

“Bias is the same as a hallucination.” A hallucination is made-up information; bias is a repeated imbalance. Both can happen at once.

These questions belong to the wider field of AI ethics.