pith. sign in

arxiv: 2412.16641 · v6 · submitted 2024-12-21 · 💻 cs.AI · cs.CY

A Systems Thinking Approach to Algorithmic Fairness

Chris Lam This is my paper

Pith reviewed 2026-05-23 06:42 UTC · model grok-4.3

classification 💻 cs.AI cs.CY
keywords algorithmic fairnesssystems thinkingcausal graphssociotechnical systemsAI policyfairness trade-offsmachine learningcausal inference
0
0 comments X

The pith

Encoding assumptions about bias as causal graphs lets systems thinking connect AI fairness to political and legal trade-offs.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that systems thinking can model algorithmic fairness by encoding beliefs about bias locations into causal graphs of the data generating process. This representation links technical AI systems to politics and the law, allowing machine learning, causal inference, and system dynamics to be combined. A sympathetic reader would care because the method aims to reveal emergent fairness issues and the trade-offs among different policy choices. The approach is presented as a foundation for designing AI policies that can align with varying political agendas while respecting shared democratic values.

Core claim

Systems thinking provides a way to model the algorithmic fairness problem by allowing us to encode prior knowledge and assumptions about where we believe bias might exist in the data generating process. We can then encode these beliefs as a series of causal graphs, enabling us to link AI/ML systems to politics and the law. This allows us to combine techniques from machine learning, causal inference, and system dynamics in order to capture different emergent aspects of the fairness problem.

What carries the argument

Causal graphs that encode prior knowledge and assumptions about bias locations in the data generating process.

If this is right

  • Policymakers gain a method to visualize and compare trade-offs among different fairness definitions.
  • AI policy design can draw on integrated machine learning, causal inference, and system dynamics models.
  • The same framework can serve political agendas on either side of the aisle while referencing shared democratic values.
  • Emergent aspects of fairness problems become visible that isolated technical metrics would miss.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The graphs could be used to run policy simulations in concrete domains such as credit scoring or hiring before laws are enacted.
  • If the encoded assumptions prove stable, the approach might help resolve conflicts between incompatible fairness metrics by showing their systemic interactions.
  • Validation would require comparing graph-derived recommendations against outcomes in deployed systems where bias locations are independently measured.

Load-bearing premise

That encoding prior knowledge about bias as causal graphs will reliably link AI systems to politics and law in a way that supports effective policy design.

What would settle it

A demonstration that causal-graph models of fairness produce policy recommendations that systematically overlook documented real-world biases or misalign with the political values they claim to serve.

Figures

Figures reproduced from arXiv: 2412.16641 by Chris Lam.

Figure 1
Figure 1. Figure 1: Three representations of algorithmic fairness [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Modeling bias, fairness, and discrimination using causal Bayesian networks (CBNs) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Adding a feedback loop to the causal Bayesian net [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Success to the successful archetype The next archetype that we will examine is the “limits to suc￾cess” archetype, which is shown in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Limits to success archetype The final system archetype that we will examine is the “shifting the burden / addiction” archetype, which is shown in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Shifting the burden / addiction archetype [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Causal loop diagram representation of algorithmic fairness [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Overt discrimination Overt discrimination causes bias in both the decision 𝐷 and the outcome 𝑌. The outcome 𝑌 becomes future data 𝑋 due to a feedback loop [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Feedback from overt discrimination The bias in mediator𝑊 gets replicated in the data 𝑋, which then causes bias in future decisions 𝐷 and outcomes 𝑌. Even though the US passed the Civil Rights Act of 1964 banning overt discrimination, this didn’t reverse the initial bias caused by historically unjust policies against non-Whites. Even a racially neutral policy could still form a basis for covert discriminati… view at source ↗
Figure 10
Figure 10. Figure 10: Covert discrimination as shown in [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 13
Figure 13. Figure 13: Internal solution and external intervention [PITH_FULL_IMAGE:figures/full_fig_p010_13.png] view at source ↗
Figure 11
Figure 11. Figure 11: Fairness through affirmative action The path 𝐷 → 𝑌 only has a weak effect on the outcome 𝑌. On the other hand, the path𝑊 → 𝑌 has a strong effect on the outcome 𝑌. This is shown in [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Strong and weak effects For example, an affirmative action policy in college admissions might help more Blacks and Hispanics to be accepted (𝐴 → 𝐷). But because they were not accepted purely based on their merit (mediator 𝑊 ), they might be less likely to graduate (outcome 𝑌). This is because they might be put into a class where their White and Asian counterparts would consistently outperform them. Anothe… view at source ↗
Figure 14
Figure 14. Figure 14: Fairness through unawareness Due to the “limits to success” archetype, human capital repre￾sents a constraint on the level of success for any racial group. By focusing more on education and pursuing higher income careers (e.g. STEM) than other racial groups, Asians have been able to catch up and surpass Whites in terms of income and wealth. This may reflect a selection bias in the types of Asians who emig… view at source ↗
Figure 15
Figure 15. Figure 15: Causal hierarchy representation of algorithmic fairness [PITH_FULL_IMAGE:figures/full_fig_p012_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Systems map representation of algorithmic fairness [PITH_FULL_IMAGE:figures/full_fig_p013_16.png] view at source ↗
read the original abstract

Systems thinking provides us with a way to model the algorithmic fairness problem by allowing us to encode prior knowledge and assumptions about where we believe bias might exist in the data generating process. We can then encode these beliefs as a series of causal graphs, enabling us to link AI/ML systems to politics and the law. This allows us to combine techniques from machine learning, causal inference, and system dynamics in order to capture different emergent aspects of the fairness problem. We can use systems thinking to help policymakers on both sides of the political aisle to understand the complex trade-offs that exist from different types of fairness policies, providing a sociotechnical foundation for designing AI policy that is aligned to their political agendas and with society's shared democratic values.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a systems thinking approach to algorithmic fairness by encoding prior knowledge and assumptions about bias locations in the data generating process as causal graphs. This is claimed to link AI/ML systems to politics and the law, enabling the combination of machine learning, causal inference, and system dynamics techniques to capture emergent fairness aspects and help policymakers on both sides of the political aisle understand trade-offs in fairness policies, thereby providing a sociotechnical foundation for AI policy aligned with political agendas and democratic values.

Significance. If operationalized with concrete demonstrations, the approach could offer an interdisciplinary bridge between technical fairness modeling and sociopolitical considerations in AI policy, extending causal modeling ideas to support value-aligned decision-making. The manuscript highlights potential for integrating established techniques but currently presents these as high-level assertions.

major comments (2)
  1. [Abstract] Abstract: The central claim that encoding assumptions as causal graphs 'enables us to link AI/ML systems to politics and the law' and supplies a 'sociotechnical foundation' for policy design is load-bearing, yet the manuscript provides no concrete causal graph, no system-dynamics simulation, and no output metrics or policy mappings to substantiate how trade-offs would be surfaced or aligned with political agendas.
  2. [Abstract] Abstract: The assertion that the method will 'help policymakers on both sides of the political aisle' depends on the untested premise that user-encoded causal graphs will reliably produce actionable, cross-aisle insights; without an example workflow or validation step, this reduces to an assertion rather than a demonstrated capability.
minor comments (2)
  1. The abstract would be strengthened by a brief illustrative example of a causal graph for a fairness domain (e.g., hiring or lending) to ground the proposed workflow.
  2. Clarify the specific integration points between causal inference and system dynamics, as the high-level mention of combining the three techniques leaves the mechanics unspecified.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which correctly identify that several claims in the abstract are not supported by concrete demonstrations in the manuscript. The work is a conceptual proposal for a systems-thinking framework rather than an operationalized method with examples. We will revise the abstract accordingly to better align claims with the manuscript's scope.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that encoding assumptions as causal graphs 'enables us to link AI/ML systems to politics and the law' and supplies a 'sociotechnical foundation' for policy design is load-bearing, yet the manuscript provides no concrete causal graph, no system-dynamics simulation, and no output metrics or policy mappings to substantiate how trade-offs would be surfaced or aligned with political agendas.

    Authors: We agree with this assessment. The manuscript presents a high-level framework for encoding prior beliefs about bias via causal graphs and combining ML, causal inference, and system dynamics techniques, but does not include any concrete graphs, simulations, metrics, or policy mappings. We will revise the abstract to qualify these statements as potential benefits of future operationalization of the approach, rather than capabilities demonstrated in the current work. revision: yes

  2. Referee: [Abstract] Abstract: The assertion that the method will 'help policymakers on both sides of the political aisle' depends on the untested premise that user-encoded causal graphs will reliably produce actionable, cross-aisle insights; without an example workflow or validation step, this reduces to an assertion rather than a demonstrated capability.

    Authors: The referee is correct that this claim rests on an untested premise and lacks an example workflow or validation. The manuscript does not provide such elements. We will revise the abstract to remove or substantially qualify this assertion, framing it as a possible outcome of applying the framework rather than a demonstrated result. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The manuscript presents a high-level conceptual proposal for applying systems thinking to algorithmic fairness via causal graphs and combined techniques, without any mathematical derivation chain, equations, fitted parameters, or self-citations that reduce outputs to inputs by construction. The central claims concern the potential utility of encoding assumptions for policy insight rather than any derived result shown to be equivalent to those assumptions. No load-bearing steps matching the enumerated circularity patterns are present in the text.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the central proposal rests on the unstated domain assumption that causal graphs can faithfully represent bias locations and policy linkages.

axioms (1)
  • domain assumption Prior knowledge about bias locations in the data generating process can be accurately encoded as causal graphs that connect technical systems to political and legal domains.
    Invoked in the abstract as the basis for linking AI/ML to politics and law.

pith-pipeline@v0.9.0 · 5633 in / 1352 out tokens · 25767 ms · 2026-05-23T06:42:58.298482+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 48 canonical work pages

  1. [1]

    Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Kirchner. 2016. Machine Bias. Pro Publica Inc. https://www.propublica.org/article/machine-bias-risk- assessments-in-criminal-sentencing

    work page 2016

  2. [2]

    Chris Barker. 2005. Cultural Studies: Theory and Practice . Sage, London, UK

    work page 2005

  3. [3]

    Solon Barocas, Moritz Hardt, and Arvind Narayanan. 2023. Fairness and Machine Learning: Limitations and Opportunities . MIT Press, Cambridge, MA

    work page 2023

  4. [4]

    Gary S. Becker. 1964. Human Capital: A Theoretical and Empirical Analysis, with Special Reference to Education . The University of Chicago Press, London, UK

    work page 1964

  5. [5]

    Marianne Bertrand and Sendhil Mullainathan. 2003. Are Emily and Greg More Employable than Lakisha and Jamal? A Field Experiment on Labor Market Dis- crimination. Working Paper 9873. National Bureau of Economic Research. https://doi.org/10.3386/w9873

    work page doi:10.3386/w9873 2003

  6. [6]

    Reuben Binns. 2019. On the Apparent Conflict Between Individual and Group Fairness. arXiv:1912.06883 [cs.LG] https://arxiv.org/abs/1912.06883

    work page arXiv 2019

  7. [7]

    Paul Carus. 1910. Person and personality. The Monist 20 (1910), 379

    work page 1910

  8. [8]

    Alessandro Castelnovo, Riccardo Crupi, Greta Greco, Daniele Regoli, Ilaria Giuseppina Penco, and Andrea Claudio Cosentini. 2022. A clarification of the nuances in the fairness metrics landscape. Scientific Reports 12, 1 (2022), 1–21

    work page 2022

  9. [9]

    Silvia Chiappa and William S. Isaac. 2019. A Causal Bayesian Networks Viewpoint on Fairness. Springer International Publishing, Cham, Switzerland, 3–20. https: //doi.org/10.1007/978-3-030-16744-8_1

    work page doi:10.1007/978-3-030-16744-8_1 2019

  10. [10]

    Heather Mac Donald. 2023. When Race Trumps Merit: How the Pursuit of Equity Sacrifices Excellence, Destroys Beauty, and Threatens Lives . DW Books, Nashville, TN

    work page 2023

  11. [11]

    Carol S. Dweck. 2007. Mindset: The New Psychology of Success . Ballantine Books, New York

    work page 2007

  12. [12]

    Jay Forrester. 1971. Counterintuitive behavior of social systems. Technology Review 53, 3 (1971), 52–68

    work page 1971

  13. [13]

    Joshua Grossman, Sabina Tomkins, Lindsay C Page, and Sharad Goel. 2023. The Disparate Impacts of College Admissions Policies on Asian American Applicants . Working Paper 31527. National Bureau of Economic Research. https://doi.org/ 10.3386/w31527

    work page doi:10.3386/w31527 2023

  14. [14]

    Fritz Heider. 1958. The Psychology of Interpersonal Relations . Lawrence Erlbaum Associates, Hillsdale, NJ

    work page 1958

  15. [15]

    Donald Martin Jr. 2021. Applying Community-Based System Dynamics to Com- bat AI Bias . System Dynamics Society. https://www.youtube.com/watch?v= LpOZXYlQOFw

    work page 2021

  16. [16]

    Donald Martin Jr. 2023. Using societal context knowledge to foster the responsible application of AI. Google Research. https://research.google/blog/using-societal- context-knowledge-to-foster-the-responsible-application-of-ai/

    work page 2023

  17. [17]

    Shelly Kagan. 2012. Moral Desert. In The Geometry of Desert . Oxford Univer- sity Press, New York, NY. https://doi.org/10.1093/acprof:oso/9780199895595. 003.0001 arXiv:https://academic.oup.com/book/0/chapter/150310098/chapter-ag- pdf/44982224/book_6467_section_150310098.ag.pdf

    work page doi:10.1093/acprof:oso/9780199895595 2012

  18. [18]

    Scott Barry Kaufman. 2020. Unraveling the Mindset of Victimhood. Scientific American 30, 5 (2020), 14

    work page 2020

  19. [19]

    H. H. Kelley. 1967. Attribution Theory in Social Psychology.Nebraska Symposium on Motivation 15 (1967), 192–238

    work page 1967

  20. [20]

    Ibram X. Kendi. 2017. Stamped from the Beginning: The Definitive History of Racist Ideas in America. Bold Type Books, New York, NY

    work page 2017

  21. [21]

    Niki Kilbertus, Mateo Rojas-Carulla, Giambattista Parascandolo, Moritz Hardt, Dominik Janzing, and Bernhard Schölkopf. 2017. Avoiding discrimination through causal reasoning. In Proceedings of the 31st International Conference on Neural Information Processing Systems (Long Beach, California, USA) (NIPS’17). Curran Associates Inc., Red Hook, NY, USA, 656–666

    work page 2017

  22. [22]

    Daniel H. Kim. 1992. Systems Archetypes I: Diagnosing Systemic Issues and De- signing High-Leverage Interventions. Pegasus Communications, Inc, Waltham, MA

    work page 1992

  23. [23]

    Rakesh Kochnar and Mohamad Moslimani. 2023. Wealth Surged in the Pandemic, but Debt Endures for Poorer Black and Hispanic Families . Technical Report. Pew Research Center. Lam

    work page 2023

  24. [24]

    Matt J Kusner, Joshua Loftus, Chris Russell, and Ricardo Silva. 2017. Counterfactual Fairness. In Advances in Neural Information Processing Systems, I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (Eds.), Vol. 30. Curran Associates, Inc., Red Hook, NY. https://proceedings.neurips.cc/paper_files/paper/2017/file/ ...

    work page 2017

  25. [25]

    Chris Lam. 2024. Debiasing Alternative Data for Credit Underwriting Using Causal Inference. arXiv:2410.22382 [q-fin.RM] https://arxiv.org/abs/2410.22382

    work page arXiv 2024

  26. [26]

    Glenn C. Loury. 2002. The Anatomy of Racial Inequality . Harvard University Press, Cambridge, MA

    work page 2002

  27. [27]

    Donella H. Meadows. 2008. Thinking in Systems . Chelsea Green Publishing Company, White River Junction, VT

    work page 2008

  28. [28]

    Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. 2021. A Survey on Bias and Fairness in Machine Learning.ACM Comput. Surv. 54, 6, Article 115 (July 2021), 35 pages. https://doi.org/10.1145/3457607

    work page doi:10.1145/3457607 2021

  29. [29]

    Charles W. Mills. 1997. The Racial Contract . Cornell University Press, Ithaca, NY

    work page 1997

  30. [30]

    Arvind Narayanan. 2018. 21 fairness definitions and their politics . ACM Confer- ence on Fairness, Accountability, and Transparency. https://www.youtube.com/ embed/jIXIuYdnyyk

    work page 2018

  31. [31]

    Hamed Nilforoshan, Johann Gaebler, Ravi Shroff, and Sharad Goel. 2022. Causal Conceptions of Fairness and their Consequences. arXiv:2207.05302 [cs.LG] https://arxiv.org/abs/2207.05302

    work page arXiv 2022

  32. [32]

    US Department of Justice Civil Rights Division. 2024. Title VI Legal Manual . US Department of Justice, Washington, DC

    work page 2024

  33. [33]

    Judea Pearl. 2000. Causality. Cambridge University Press, Cambridge, UK

    work page 2000

  34. [34]

    Judea Pearl. 2001. Direct and indirect effects. In Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence (Seattle, Washington) (UAI’01). Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 411–420

    work page 2001

  35. [35]

    Judea Pearl, Madelyn Glymour, and Nicholas P. Jewell. 2016. Causal Inference in Statistics. Wiley, West Sussex, UK

    work page 2016

  36. [36]

    Judea Pearl and Dana Mackenzie. 2018. The Book of Why: The New Science of Cause and Effect. Basic Books, New York, NY

    work page 2018

  37. [37]

    Drago Plecko and Elias Bareinboim. 2022. Causal Fairness Analysis. arXiv:2207.11385 [cs.AI] https://arxiv.org/abs/2207.11385

    work page arXiv 2022

  38. [38]

    Julian B Rotter. 1966. Generalized expectancies for internal versus external control of reinforcement. Psychological Monographs: General and Applied 80, 1 (1966), 1–28

    work page 1966

  39. [39]

    Sander and Stuart Taylor

    Richard H. Sander and Stuart Taylor. 2012. Mismatch: How Affirmative Action Hurts Students It’s Intended to Help, and Why Universities Won’t Admit It. Basic Books, New York

    work page 2012

  40. [40]

    M E P Seligman. 1972. Learned Helplessness. Annual Review of Medicine 23, Volume 23, 1972 (1972), 407–412. https://doi.org/10.1146/annurev.me.23.020172. 002203

    work page doi:10.1146/annurev.me.23.020172 1972

  41. [41]

    Peter M. Senge. 1990. The Fifth Discipline: The Art and Practice of The Learning Organization. Doubleday Publishing Group, New York, NY

    work page 1990

  42. [42]

    Thomas Sowell. 2004. Affirmative Action Around the World: An Empirical Study. Yale University Press, New Haven, CT

    work page 2004

  43. [43]

    Thomas Sowell. 2019. Discrimination and Disparities. Basic Books, New York, NY

    work page 2019

  44. [44]

    John D. Sterman. 2000. Business Dynamics: Systems Thinking and Modeling for a Complex World. McGraw-Hill, Boston, MA

    work page 2000

  45. [45]

    American Community Survey. 2021. S0201: Selected Population Profile in the United States. Technical Report. United States Census Bureau

    work page 2021

  46. [46]

    Matthew Talbert. 2024. Moral Responsibility. In The Stanford Encyclopedia of Philosophy (Fall 2024 ed.), Edward N. Zalta and Uri Nodelman (Eds.). Metaphysics Research Lab, Stanford University, Stanford, CA

    work page 2024

  47. [47]

    Sahil Verma and Julia Rubin. 2018. Fairness definitions explained. InProceedings of the International Workshop on Software Fairness (Gothenburg, Sweden) (FairWare ’18). Association for Computing Machinery, New York, NY, USA, 1–7. https: //doi.org/10.1145/3194770.3194776

    work page doi:10.1145/3194770.3194776 2018

  48. [48]

    Bernard Weiner. 1986. An Attributional Theory of Motivation and Emotion . Springer, New York. NY. Received 25 June 2025

    work page 1986