Skip to main content

The patterns we detect in AI are not novel. They reflect biases documented across fifty years of social science research. Every citation is a debt acknowledged.

I. The Naming Problem (1960s–2000s)

Names carry demographic signals that influence how people are treated.

2004

Bertrand & Mullainathan

The seminal study. Identical resumes with names signaling different racial backgrounds received dramatically different callback rates. Emily and Greg received 50% more callbacks than Lakisha and Jamal.

2004

Fryer & Levitt

Validated the demographic associations of names. Established which names reliably signal racial and ethnic identity in American contexts.

2008

Cotton, O'Neill & Griffin

Extended name research to first names. Demonstrated that name characteristics (not just racial associations) influence professional evaluations.

2012

Laham, Koval & Alter

The name-pronunciation effect. People judge others more favorably when their names are easier to pronounce—a phonetic basis for bias.

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
Our debt: The matched-pair methodology.
DOI: 10.1257/0002828042002561 →

II. Healthcare Disparities (1990s–2010s)

Demographic indicators systematically influence medical treatment decisions.

1999

Schulman et al.

Race and sex affect physicians' cardiac catheterization recommendations. The same symptoms, different referral rates. A landmark NEJM study.

2003

Strakowski et al.

Racial disparities in psychiatric diagnosis from emergency services. Schizophrenia diagnosed more frequently in Black patients presenting with identical symptoms.

2016

Hoffman et al.

Racial bias in pain assessment. Medical students and residents endorsed false beliefs about biological differences and recommended less pain treatment for Black patients.

Hoffman, K. M., Trawalter, S., Axt, J. R., & Oliver, M. N. (2016)
Racial bias in pain assessment and treatment recommendations, and false beliefs about biological differences between blacks and whites
PNAS, 113(16), 4296-4301
Our debt: AI systems have learned these same disparities.
DOI: 10.1073/pnas.1516047113 →
Schulman, K. A., et al. (1999)
The effect of race and sex on physicians' recommendations for cardiac catheterization
New England Journal of Medicine, 340(8), 618-626
Our debt: AI exhibits the same patterns, 25 years later.
DOI: 10.1056/NEJM199902253400806 →

III. Algorithmic Bias (2010s–Present)

AI systems inherit and amplify human biases at scale.

2016

Bolukbasi et al.

Word embeddings encode gender bias. "Man is to computer programmer as woman is to homemaker." The mathematical representations of language carry our biases.

2016

ProPublica (Angwin et al.)

COMPAS criminal justice algorithm biased against Black defendants. Investigative journalism that brought algorithmic bias into public discourse.

2017

Caliskan, Bryson & Narayanan

Semantics derived from language corpora contain human-like biases. AI learns our prejudices from the text it reads.

2018

Buolamwini & Gebru

Gender Shades. Commercial facial recognition systems had error rates up to 34.7% higher for darker-skinned women than lighter-skinned men. Intersectional disparities in AI.

2019

Obermeyer et al.

Racial bias in a healthcare algorithm affecting millions. A widely-used algorithm systematically discriminated against Black patients by using cost as a proxy for health need.

2021

Bender et al.

"On the Dangers of Stochastic Parrots." Comprehensive analysis of risks in large language models, including bias propagation and the illusion of understanding.

Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019)
Dissecting racial bias in an algorithm used to manage the health of populations
Science, 366(6464), 447-453
Our debt: Algorithmic bias in healthcare affects millions.
DOI: 10.1126/science.aax2342 →
Buolamwini, J., & Gebru, T. (2018)
Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification
Proceedings of Machine Learning Research, 81, 77-91
Our debt: Systematic demographic auditing of AI systems.

IV. The Gap We Address

Prior research established that:

  • Names carry demographic signals that influence treatment (Bertrand & Mullainathan)
  • Healthcare decisions systematically differ by demographic indicators (Hoffman, Schulman)
  • AI systems inherit and amplify human biases (Obermeyer, Buolamwini)

But this work focused primarily on content—what AI systems say, what treatments they recommend, what predictions they make.

We asked a different question:

Does bias persist in the structure of AI-generated text—the mathematical patterns of how responses unfold—even when vocabulary appears neutral?

The answer, documented in our evidence section, is yes. Structural analysis reveals patterns invisible to content analysis.

Our Contribution

We developed methods to detect bias in narrative structure—12 dimensions of analysis that reveal how AI-generated text unfolds differently based on name signals. Methodology details forthcoming in academic publication.

Continue exploring the research

See the Evidence