Bias is a systematic tendency for model outputs to reflect or amplify unequal or stereotypical patterns associated with race, gender, culture, or other social contexts. In text-to-image generation, broad training datasets can make prompts such as “a doctor” or “a leader” produce disproportionately stereotypical depictions. section 1.6 section 13.2
Assess bias through its social effects and the difficulty of correcting it:
- Recognize that biased outputs are not merely technical defects; they can reflect and amplify existing social inequalities. section 13.2
- Audit and mitigate stereotypical representations while avoiding overcorrection that produces inaccurate historical depictions. section 1.6 section 13.2
- Treat mitigation as a multidisciplinary challenge involving AI developers, social scientists, and ethicists. section 1.6
Prompt stereotypes and overcorrection
A prompt such as “a doctor” or “a leader” may yield stereotypical images. Conversely, attempts to correct stereotypes can introduce historically inaccurate depictions, such as portraying people of color among Nazi-era German soldiers. section 1.6 section 13.2
