Reading and Implementation on Intersectionality

For this assignment, please read the following book chapters and articles:
1. Are "Intersectionally Fair" AI Algorithms Really Fair to Women of Color? A Philosophical Analysis, Youjin Kong, ACM FAccT, 2022
2. Decoding Ableism in Large Language Models: An Intersectional Approach, Rong Li et al., Workshop on NLP for Positive Impact, 2024
3. Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval, Kyra Wilson and Aylin Caliskan, AAAI/ACM AIES, 2024
4. White Men Lead, Black Women Help? Benchmarking and Mitigating Language Agency Social Biases in LLMs, Yixin Wan and Kai-Wei Chang, ACL, 2025
After completing these readings, please critical answer the following questions.

Based on paper [1], answer the following questions:
1. What limitations does the paper identify in using statistical parity for intersectional identities, particularly for women of color? How might this approach unintentionally reinforce existing inequalities?
2. The paper suggests that conventional fairness metrics may be inadequate for addressing intersectional bias. What alternative or additional approaches does it propose or imply? How might these approaches better serve women of color in algorithmic decision-making?

Based on paper [2], answer the following questions:
1. The paper notes that some disabilities (e.g., body dysmorphic disorder and depression) consistently score lower in sentiment than others (e.g., Down syndrome and intellectual disabilities). What might explain these differences in light of the training data or societal stereotypes? Give one interpretation grounded in the authors’ discussion.
2. The authors use VADER (sentiment) and regard (social perception) as core metrics. What biases or misclassifications might these tools introduce when evaluating disability-related text? Use reasons connected to the examples shown in the paper.

Based on paper [3], answer the following questions:
1. The paper finds that intersectional bias cannot be explained by summing gender bias + race bias. Using the authors' discussion, why is intersectional bias non-additive, and what does this mean for evaluating fairness in LLM hiring systems?.
2. The authors show that simply shortening resumes (name + title only) or changing the corpus frequency of names dramatically alters which groups are favored. What does this reveal about how MTEs encode bias, and why do these low-level features challenge conventional fairness auditing?

Based on paper [4], answer the following questions:
1. The authors find that adding fairness instructions to prompts can increase bias. Why might LLMs overshoot or invert their behaviors when given fairness instructions, and what does this reveal about the limitations of prompt-based mitigation strategies?
2. How does the paper explain that communal language (e.g., “helpful,” “supportive”), although positive-sounding, can still reinforce stereotypes and lead to harmful real-world outcomes?