Is AI biased? What business owners need to know
By Rodney Noriega, Founder, Universal Enablement · Published Oct 12, 2026 · Updated Oct 12, 2026
Short answer
Yes. AI learns from human-made data, and that data carries human bias. Research shows AI often makes existing unfairness worse, not just copies it. In a business, bias shows up wherever AI writes about, describes, sorts or judges people: job ads, customer replies, images, and anything that ranks applicants or customers.
Some well-documented examples:
- Faces. In Joy Buolamwini and Timnit Gebru's 2018 Gender Shades study, commercial systems got darker-skinned women wrong up to 34.7% of the time. For lighter-skinned men, the worst error rate was 0.8%.
- Images. A 2023 Bloomberg analysis of more than 5,000 AI-generated images found high-paying jobs like CEO and lawyer shown mostly as lighter-skinned men, more extreme than real workforce numbers.
- Text. Students asked ChatGPT 40 times to describe a day in the life of a CEO and a nurse. Every CEO was "he." Every nurse was "she."
- Subtle language. Asked for "good hair," one assistant showed smooth, straight styles. "Bad hair" got curly and locked hair, described as messy and neglected.
None of these tools was told to be biased. They learned patterns from data that never represented everyone fairly. Every major assistant, from ChatGPT to Claude to Gemini, faces this problem to some degree. Some of these studies are a few years old and accuracy has improved since, but newer tests keep finding the same patterns.
What to do:
- List every workflow where AI output describes or affects people.
- Test it with different names, genders and backgrounds, and compare.
- Have a person review those outputs before they go out.
Related
Start with one workflow. In 30 minutes we'll look at where your team's time goes and tell you honestly whether AI can help.
Book a free 30-minute workflow review