Where AI bias comes from

AI bias isn’t a glitch bolted on at the end — it’s a reflection of the data a model learned from. Because models are trained on human writing, they absorb human patterns, including the unfair ones. Understanding the source makes the risk easier to spot and manage.

Where it shows up

• Skewed examples. Ask for “a nurse” or “a CEO” and a model may lean on stereotypes that were common in its training text.
• Uneven quality. Models often perform better in widely-represented languages and contexts than in under-represented ones.
• Confident omissions. A model can present a partial, culturally-narrow view as if it were the whole picture.

What you can do

You can’t retrain the model, but you can change how you use it. Ask for a range of perspectives. Be specific to avoid defaults doing the talking. Treat its output as a draft to review — bias is one more reason the human stays the editor. And in any decision affecting people (hiring, lending, grading), never let a model be the unaccountable decider.

A model’s fluency can make a biased answer sound authoritative. Sounding sure is not the same as being fair.

Disclosure matters here too: people deserve to know when AI shaped something that affects them — see when to disclose AI use.