On the Effectiveness and Generalization of Race Representations for Debiasing High-Stakes Decisions
A web visualization
This is an accompanying website for the paper with the same title. Below we provide experiments you can run to gain better intuition for the experiments and results in the paper.
1. Prompting is ineffective at debiasing
In the paper, we show that prompting fails to debias models in both Admissions and Hiring. In some cases, it may even worsen the bias or reduces the acceptance to nearly zero.
In this visualization, you will engineer a prompt to debias models. The debiasing prompt will be appended to the baseline prompt (before "Answer:"), which is a template for the task and model. We also provide the debiasing prompts used in the paper to get you started.
Select the task and model to start debiasing!
Below is the baseline prompt template for the selected dataset and model. We use it to get models' decisions.