Chapter 31 Further Reading: Understanding AI Bias and How It Surfaces

Foundational Research

1. Buolamwini, J., & Gebru, T. (2018). Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. Proceedings of the 1st Conference on Fairness, Accountability and Transparency (FAccT). The landmark audit study that established the methodology for AI demographic bias measurement. Found error rate disparities of up to 34 percentage points across demographic groups in commercial facial recognition systems. Required reading for understanding both the severity of AI bias and the methodological approach to measuring it. Available free through the Proceedings.

2. Bolukbasi, T., Chang, K. W., Zou, J. Y., Saligrama, V., & Kalai, A. T. (2016). Man is to computer programmer as woman is to homemaker? Debiasing word embeddings. Advances in Neural Information Processing Systems, 29. The paper that demonstrated occupational stereotyping in word embeddings using analogy completion. Established that trained word representations encode gender-occupational associations at a level quantifiable through controlled tests. Foundation for understanding how occupational bias gets embedded in model representations. Free via NeurIPS Proceedings.

3. Caliskan, A., Bryson, J. J., & Narayanan, A. (2017). Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334), 183-186. Extended the word embedding bias findings to show alignment with human Implicit Association Test results — demonstrating that learned language representations encode documented human biases. Published in Science and widely cited; demonstrates the connection between corpus-derived representations and measurable human bias patterns.

4. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of FAccT 2021. The "Stochastic Parrots" paper addresses environmental cost, bias, and the limitations of scale as a solution to AI problems. The section on documentation debt and training data bias is essential reading for understanding structural sources of AI bias beyond individual model choices.

Bias in Language Models and NLP

5. Sheng, E., Chang, K. W., Natarajan, P., & Peng, N. (2019). The Woman Worked as a Babysitter: On Biases in Language Generation. Proceedings of EMNLP-IJCNLP 2019. Demonstrates demographic bias in story generation from language models — differential characterization of people based on demographic characteristics in the prompt. Directly relevant to the generation bias described in Chapter 31. Available free through ACL Anthology.

6. Abid, A., Farooqi, M., & Zou, J. (2021). Persistent Anti-Muslim Bias in Large Language Models. Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society. Documents systematic negative associations with Muslim identity in large language model completions. One of the most cited examples of demographic bias in generative AI beyond gender bias. Demonstrates that bias affects multiple demographic dimensions with consistent, systematic patterns.

7. Dhamala, J., Sun, T., Kumar, V., Krishna, S., Pruksachatkun, Y., Chang, K. W., & Gupta, R. (2021). BOLD: Dataset and Metrics for Measuring Biases in Open-Ended Language Generation. Proceedings of FAccT 2021. Introduces a benchmark dataset for measuring bias in open-ended generation. Useful for understanding how researchers quantify generation bias and what controlled tests look like at scale. Available free through FAccT Proceedings.

Job Description and Hiring Bias

8. Gaucher, D., Friesen, J., & Kay, A. C. (2011). Evidence That Gendered Wording in Job Advertisements Exists and Sustains Gender Inequality. Journal of Personality and Social Psychology, 101(1), 109-128. The foundational study demonstrating that masculine-coded language in job descriptions reduces application rates from women. The research basis for the gender decoder tool referenced in this chapter. Understanding the methodology clarifies why language auditing of hiring materials is evidence-based practice, not optional.

9. Bertrand, M., & Mullainathan, S. (2004). Are Emily and Greg More Employable than Lakisha and Jamal? A Field Experiment on Labor Market Discrimination. American Economic Review, 94(4), 991-1013. The landmark field experiment demonstrating name-based discrimination in hiring callbacks. While this study predates AI tools, it establishes the empirical basis for the concern about name-based demographic bias in AI-assisted hiring contexts — if human bias produces this effect, AI systems trained on human decisions inherit the same patterns.

10. HireVue Algorithmic Bias Audit — Equal Employment Opportunity Commission Resources (eeoc.gov) The EEOC maintains resources on AI tools in employment contexts, including guidance on employer responsibilities when using algorithmic decision tools. Essential reading for any professional involved in deploying AI in hiring contexts in the United States. Free to access.

Practical Tools and Resources

11. Gender Decoder for Job Ads (gender-decoder.katmatfield.com) The free tool referenced in this chapter. Built on the Gaucher et al. (2011) research. Analyzes text for masculine- and feminine-coded language and provides immediate, interpretable results. Essential practical tool for anyone reviewing AI-generated job descriptions. Free.

12. AI Fairness 360 (IBM Research — aif360.res.ibm.com) IBM Research's open-source toolkit for detecting and mitigating bias in datasets and machine learning models. Primarily for technical practitioners, but the documentation provides accessible explanations of bias metrics and mitigation approaches that are useful for professional bias literacy even without a technical implementation context. Free and open source.

13. Partnership on AI — Responsible AI in Practice Resources (partnershiponai.org) The Partnership on AI produces practitioner-focused resources on bias, fairness, and responsible deployment. The case studies and implementation guides are more practically accessible than academic research papers for professionals implementing bias awareness in organizational contexts.

Broader Context

14. Noble, S. U. (2018). Algorithms of Oppression: How Search Engines Reinforce Racism. New York University Press. A book-length critical examination of how algorithmic systems reproduce and amplify racial inequality. Focuses primarily on search, but the underlying arguments about training data, commercial incentives, and institutional choices in AI development apply broadly to the bias concerns in this chapter. Accessible to non-technical readers.

15. Benjamin, R. (2019). Race After Technology: Abolitionist Tools for the New Jim Code. Polity Press. Examines the intersection of race, technology, and power through multiple case studies of algorithmic and AI systems. The concept of "the New Jim Code" — technologies that perpetuate racial discrimination while appearing neutral — is a valuable critical framework for the AI bias literacy this chapter builds.