Rosa wants to plan her spring email to customers. She does not paste her customer list, because she learned that lesson in week 3.
Instead she asks a chatbot a general question: who usually buys flowers, and what should her email say?
The answer comes back sure of itself. As she reads, a few lines make her frown. Some of her best customers do not fit it at all.
Bias in a written output is not always rude or obvious. Often it is a quiet guess about who people are.
NIST warns that generative AI can amplify social biases. In writing, that can look like a stereotype stated as a fact.
A stereotype is a fixed, oversimplified idea about a whole group of people.
Read this example. It was written for this lesson. It contains four lines with bias to find.
| Line in the output | Is it free of bias? |
|---|---|
| "Most flower buyers are young women shopping for a date." | ? |
| "Men rarely buy flowers, so there is no need to write for them." | ? |
| "Skip older customers. They do not read email." | ? |
| "Write only in English." | ? |
| "Brighten your home with fresh tulips this week." | ? |
Notice that the output sounds confident. In week 2 you saw that confident and correct are not the same.
Rosa's fix is simple. She writes one email for all her customers, and she checks it with what she knows.
Sources: NIST AI 600-1, Generative AI Profile (2024)
Tomorrow: fairness in your own track.