ChatGPT can exhibit biases reflecting patterns in its training data — a critical issue for fair and responsible deployment.

Sources of Bias

• Training data over-represents English and Western perspectives
• Human raters in RLHF bring their own cultural assumptions
• Historical data encodes historical inequalities

Manifestations

• Gender stereotypes in professional descriptions
• Cultural insensitivity in non-Western contexts
• Political bias perceptions (heavily studied and debated)

Reference:

Wikipedia: Algorithmic Bias

https://en.wikipedia.org/wiki/Algorithmic_bias