Advancing Equality: Harnessing Generative AI to Combat Systemic Racism

Saadia Gabriel, Jessy Xinyi Han, Eric Minwei Liu, Isha Puri, Wonyoung So, Fotini Christia, Munzer Dahleh, Catherine D’Ignazio, Marzyeh Ghassemi, Peko Hosoi, Devavrat Shah2024 · DOI

Abstract

Systemic racism, an issue deeply entrenched in various sectors of American society, continues to cast a shadow over the nation's progress towards equality and justice. This paper showcases some of generative artificial intelligence (AI) potential to identify racial bias and combat structural barriers that perpetuate racial discrimination in the United States. Specifically, three research groups affiliated with MIT’s Initiative on Combatting Systemic Racism (ICSR) present case study applications that utilize generative AI in innovative approaches against the structural foundations of racism in policing, healthcare, and housing.

Introduction

Systemic racism is often "baked into" the data used to train algorithms, creating a vicious cycle where AI reinforces historical inequities. A major hurdle in social science is the lack of counterfactuals—the ability to answer whether an outcome would have changed if an individual's race were different. Traditional observational data is often confounded, making it difficult to identify exactly where bias enters complex systems like police encounters or clinical trials. This research addresses these challenges by using modern Large Language Models (LLMs) and generative frameworks to audit discriminatory practices and design reparative policy interventions.

Methodology

The researchers employed three distinct methodological approaches across different sectors:

  1. Causal Simulators: For policing and healthcare, generative causal frameworks were built to model multi-stage interactions (e.g., from a 911 call to the use of force). These simulators allow for counterfactual scenarios that isolate specific variables to pinpoint the root causes of bias.
  2. Survival Synthetic Intervention (SSI): This novel causal survival method was applied in healthcare to estimate personalized outcomes. It enables comparisons of treatment efficacy across heterogeneous patient cohorts, even with sparse or biased data.
  3. Algorithmic Auditing: To evaluate LLMs like GPT-4, the team used templated prompts and RoBERTa-based classifiers. They measured empathy levels in mental health responses and "probability-of-recommendation" scores to detect racial steering in housing.

Datasets

The study utilized diverse data sources, including FOIA-requested policing data from over 40 US cities, the PETAL consortium healthcare dataset (940 T-cell lymphoma patients), Reddit mental health posts, and 2019 Census data for housing analysis.

Key Insights and Results

Policing: Geographical Drivers of Bias

A critical finding is that the source of racial bias varies significantly by geography. In New York City, the simulator revealed that racial disparity is primarily driven by policing actions—the decisions made by officers on the ground. Conversely, in New Orleans, the disparity is largely driven by biased 911 reporting from the public.

Distribution of precinct-level disparities in NYC and New Orleans policing across various years and categories.

This distinction suggests that policy interventions must be localized: focusing on officer training in NYC while addressing community reporting biases in New Orleans.

Healthcare: Shifting Clinical Paradigms

Using the SSI model for T-cell lymphoma (TCL), researchers found that Single Agent (SA) therapy is comparable or even superior to standard cytotoxic chemotherapy in certain subtypes. This finding supports a potential shift in clinical paradigms toward more personalized, less toxic treatments.

LLMs: "Default Whiteness" and Amplified Bias

Audits of GPT-4 revealed a "Default Whiteness" bias; when race is unspecified, the model's recommendations closely mirror those given for White users. Furthermore, the model exhibited a significant empathy gap in mental health support, scoring lower for Asian patients (.43) than for White patients (.55)—a disparity not present in the human Reddit data it was trained on. In housing, the model was more likely to steer Black homeseekers toward neighborhoods with lower socioeconomic status, even when profiles were otherwise identical.

Discussion and Limitations

While generative AI offers powerful diagnostic capabilities, the "black box" nature of proprietary models like GPT-4 remains a challenge. Without transparency regarding training data, pinpointing the exact origins of these biases is difficult.

Future research will focus on developing reparative models that account for the evolution of discriminatory policies over time. There is also a pressing need for more interpretable generative algorithms and regulatory oversight, including "community-in-the-loop" auditing, before integrating these models into critical social infrastructure.

Conclusion

Generative AI serves as both a mirror, reflecting societal prejudices, and a tool for navigating the maze of systemic racism. By leveraging causal simulators and algorithmic auditing, researchers can identify specific points of intervention to foster a more equitable society.

Summary updated

Abstract

Systemic racism, an issue deeply entrenched in various sectors of American society, continues to cast a shadow over the nation's progress towards equality and justice. This paper showcases some of generative artificial intelligence (AI) potential to identify racial bias and combat structural barriers that perpetuate racial discrimination in the United States. Specifically, three research groups affiliated with MIT’s Initiative on Combatting Systemic Racism (ICSR) present case study applications that utilize generative AI in innovative approaches against the structural foundations of racism in policing, healthcare, and housing.
BibTeX
@misc{gabriel2024advancing,
  author = {Saadia Gabriel and Jessy Xinyi Han and Eric Minwei Liu and Isha Puri and Wonyoung So and Fotini Christia and Munzer Dahleh and Catherine D’Ignazio and Marzyeh Ghassemi and Peko Hosoi and Devavrat Shah},
  title = {{Advancing Equality: Harnessing Generative AI to Combat Systemic Racism}},
  year = {2024},
  doi = {10.21428/e4baedd9.7dc53bbf},
}
Figure