Chapter 9: Further Reading

Annotated bibliography — 19 sources

Foundational Empirical Work

1. Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016). "Machine Bias." ProPublica, May 23, 2016.

The investigation that launched a thousand papers. Angwin and colleagues analyzed COMPAS risk scores for more than 7,000 defendants in Broward County, Florida, and documented substantially higher false positive rates for Black defendants compared to white defendants. The methodology, datasets, and code were published alongside the article, enabling independent scrutiny. The piece is both a landmark work of data journalism and a primary source for understanding why the COMPAS controversy matters. Every student of algorithmic fairness should read the original article rather than relying on secondary descriptions. Available at: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing

2. Larson, J., Mattu, S., Kirchner, L., & Angwin, J. (2016). "How We Analyzed the COMPAS Recidivism Algorithm." ProPublica, May 23, 2016.

The methodological companion to "Machine Bias." Explains in technical detail how ProPublica obtained the data, matched defendants to outcomes, defined their thresholds, and computed the metrics they reported. Essential reading for understanding both the strengths and limitations of the ProPublica analysis. Available at: https://www.propublica.org/article/how-we-analyzed-the-compas-recidivism-algorithm

3. Dieterich, W., Mendoza, C., & Brennan, T. (2016). "COMPAS Risk Scales: Demonstrating Accuracy Equity and Predictive Parity." Northpointe Inc.

Northpointe's formal response to the ProPublica investigation. Presents calibration as the appropriate fairness metric for COMPAS and demonstrates that at each score level, recidivism rates are similar across racial groups. Essential for understanding the defense of calibration as a fairness criterion and for appreciating that the ProPublica/Northpointe debate is not a matter of one side being wrong. Though a vendor document, the statistical arguments are substantive. Available through academic search engines.

The Impossibility Theorems

4. Chouldechova, A. (2017). "Fair Prediction with Disparate Impact: A Study of Bias in Recidivism Prediction Instruments." Big Data, 5(2), 153–163.

The paper that formally proved the impossibility of simultaneously satisfying demographic parity, equalized odds, and calibration when base rates differ. Written in direct response to the COMPAS controversy, Chouldechova derives the mathematical relationship between these metrics and the base rate, proving that any two of the three criteria force the third to fail. The mathematical content is accessible to readers with basic statistics; the policy implications are clearly stated. This is one of the most important papers in the algorithmic fairness literature.

5. Kleinberg, J., Mullainathan, S., & Raghavan, M. (2016, published 2018). "Inherent Trade-Offs in the Fair Determination of Risk Scores." Proceedings of Innovations in Theoretical Computer Science (ITCS 2017). arXiv:1609.05807.

Independently establishes impossibility results closely related to Chouldechova's, approaching the problem from a more theoretical computer science perspective. Kleinberg and colleagues prove that three natural conditions for fair risk scores — calibration, equal false positive rates across groups, and equal false negative rates across groups — are mutually incompatible except in degenerate cases. Together with Chouldechova (2017), this paper forms the mathematical foundation for understanding why fairness metric conflicts are not engineering failures but mathematical necessities.

Foundational Metrics Papers

6. Hardt, M., Price, E., & Srebro, N. (2016). "Equality of Opportunity in Supervised Learning." Advances in Neural Information Processing Systems (NeurIPS) 29, 3315–3323.

The paper that formalized equalized odds and equal opportunity as fairness criteria. Hardt and colleagues provide formal definitions, prove properties of these criteria, and describe postprocessing methods for achieving them. They also discuss the relation between their criteria and the ProPublica-type concerns that motivated the work. The paper is technically demanding but the conceptual sections are accessible and important.

7. Dwork, C., Hardt, M., Pitassi, T., Reingold, O., & Zemel, R. (2012). "Fairness Through Awareness." Proceedings of the 3rd Innovations in Theoretical Computer Science Conference, 214–226.

The foundational paper on individual fairness. Dwork and colleagues formalize the principle that similar individuals should be treated similarly, providing the Lipschitz condition definition described in Section 9.3. They also introduce important distinctions between different fairness goals and discuss the challenge of defining a task-relevant similarity metric. This paper launched much of the subsequent work on individual fairness.

8. Corbett-Davies, S., & Goel, S. (2018). "The Measure and Mismeasure of Fairness: A Critical Review of Fair Machine Learning." arXiv:1808.00023.

A comprehensive and opinionated review of the fairness metrics literature that argues for a decision-theoretic framing: the appropriate fairness criterion depends on the costs and benefits associated with different types of errors for different groups. The authors are critical of approaches that treat fairness as a constraint to be satisfied rather than an objective to be optimized. Useful for understanding alternative framings and for engaging with critiques of the standard metrics.

Intersectional Fairness

9. Hébert-Johnson, U., Kim, M. P., Reingold, O., & Rothblum, G. N. (2018). "Multicalibration: Calibration for the (Computationally-Identifiable) Masses." Proceedings of the 35th International Conference on Machine Learning (ICML).

Introduces multicalibration as a stronger fairness criterion requiring calibration for all efficiently computable subgroups of the population. This addresses the intersectional fairness problem by ensuring that fairness properties hold not just for protected groups defined a priori but for any subgroup that can be computed efficiently. The paper is technically demanding but the concept and motivation are accessible and important for practitioners thinking about intersectional analysis.

10. 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), 77–91.

A landmark empirical paper demonstrating that commercial facial recognition systems from major technology companies perform significantly worse for darker-skinned faces and for women, with the worst performance for darker-skinned women — the intersection of race and gender. "Gender Shades" popularized intersectional analysis in the AI fairness literature and became one of the most widely cited empirical demonstrations of intersectional AI harm. Highly accessible and suitable for assigning to students.

11. State v. Loomis, 881 N.W.2d 749 (Wisconsin Supreme Court, 2016).

The Wisconsin Supreme Court's decision upholding the use of COMPAS scores in sentencing, despite constitutional challenges on due process and equal protection grounds. The decision is extensively discussed in Chapter 9 (Case Study 9.1). Reading the actual opinion is valuable for understanding how courts have reasoned about algorithmic tools in criminal justice. Available through Westlaw or through free legal research databases.

12. Barocas, S., & Selbst, A. D. (2016). "Big Data's Disparate Impact." California Law Review, 104(3), 671–732.

A rigorous legal-technical analysis of how machine learning systems can produce legally cognizable disparate impact under civil rights law, even without discriminatory intent. Barocas and Selbst describe several mechanisms through which algorithmic systems generate disparate impact and analyze the adequacy of existing legal frameworks for addressing them. Essential reading for understanding the legal dimensions of algorithmic fairness.

13. Eubanks, V. (2018). Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin's Press.

A book-length investigation of automated decision systems used in public services — child welfare, social services, criminal justice — that disproportionately affect poor communities and communities of color. Eubanks conducts detailed case studies of three systems: Indiana's automated benefits eligibility system, an algorithm used by Allegheny County child welfare services, and the predictive policing program in Los Angeles. The book is essential for understanding the stakes of algorithmic fairness outside the abstract context and is written for general audiences. Pairs well with Chapter 9's case studies.

Tools and Technical Practice

14. Bellamy, R. K. E., et al. (2019). "AI Fairness 360: An Extensible Toolkit for Detecting and Mitigating Algorithmic Bias." IBM Journal of Research and Development, 63(4/5), 4:1–4:15.

Describes the AI Fairness 360 toolkit, an open-source Python library developed by IBM that implements more than seventy fairness metrics and more than ten bias mitigation algorithms. The paper explains the design philosophy, the fairness metrics implemented, and the bias mitigation approaches (pre-processing, in-processing, and post-processing). Essential technical reference for practitioners implementing fairness analysis. The toolkit is freely available at: https://github.com/Trusted-AI/AIF360

15. Bird, S., et al. (2020). "Fairlearn: A Toolkit for Assessing and Improving Fairness in AI." Microsoft Research Technical Report MSR-TR-2020-32.

Describes Fairlearn, Microsoft's open-source fairness toolkit for Python. Covers the fairness metrics Fairlearn supports, the mitigation algorithms included, and practical guidance for using the dashboard and comparison tools. Fairlearn has a somewhat different design philosophy from AI Fairness 360, emphasizing the selection and comparison of models under fairness constraints. Available at: https://fairlearn.org

Documentation and Process

16. Mitchell, M., et al. (2019). "Model Cards for Model Reporting." Proceedings of the 2019 ACM Conference on Fairness, Accountability, and Transparency (FAccT), 220–229.

Proposes "model cards" as a standard format for documenting machine learning models, including their performance characteristics, intended use cases, and fairness evaluations across demographic groups. Model cards have been adopted by major technology companies and are increasingly expected by regulators. The paper describes the format, rationale, and examples. Essential reading for practitioners building fairness documentation processes.

17. Gebru, T., et al. (2018, 2021). "Datasheets for Datasets." Communications of the ACM, 64(12), 86–92.

Proposes standardized documentation for datasets — "datasheets" — that would describe a dataset's provenance, composition, collection process, recommended uses, and limitations including potential sources of bias. Datasheets for Datasets is the data-documentation companion to Model Cards. Both papers are widely cited and have influenced industry practice. Essential for understanding how documentation can support fairness accountability.

Critical Perspectives

18. Selbst, A. D., Boyd, D., Friedler, S. A., Venkatasubramanian, S., & Barocas, S. (2019). "Fairness and Abstraction in Sociotechnical Systems." Proceedings of the ACM Conference on Fairness, Accountability, and Transparency (FAccT), 59–68.

Argues that the dominant technical approach to algorithmic fairness fails by abstracting away the social context in which fairness problems arise. The paper identifies several "traps" that algorithmic fairness work falls into when it treats fairness as a purely technical problem: the framing trap (applying fairness in ways that preserve rather than challenge existing social structures), the portability trap (assuming solutions developed in one context apply in another), and others. Essential for understanding the limits of purely technical approaches.

19. Benjamin, R. (2019). Race After Technology: Abolitionist Tools for the New Jim Code. Polity Press.

Benjamin introduces the concept of the "New Jim Code" — the embedding of racial discrimination into technological systems that are presented as neutral and objective. The book examines facial recognition, predictive policing, healthcare algorithms, and housing technologies. Benjamin argues that racial bias in technology is not an accidental byproduct of engineering decisions but reflects and reinforces the racial order of society more broadly. Critical but accessible, and provides essential context for understanding why fairness metrics alone are insufficient to address algorithmic discrimination.

All journal articles can be found through institutional library databases. Papers cited from arXiv are freely available at arxiv.org. The ProPublica investigation is freely available online. Legal opinions are available through Westlaw, LexisNexis, or free legal research resources including Google Scholar's case law search.