Authors:
(1) Evan Shieh, Young Data Scientists League (evan.shieh@youngdatascientists.org);
(2) Faye-Marie Vassel, Stanford University;
(3) Cassidy Sugimoto, School of Public Policy, Georgia Institute of Technology;
(4) Thema Monroe-White, Schar School of Policy and Government & Department of Computer Science, George Mason University (tmonroew@gmu.edu). Authors: Authors: (1) Evan Shieh, Young Data Scientists League (evan.shieh@youngdatascientists.org); (2) Faye-Marie Vassel, Stanford University; (3) Cassidy Sugimoto, School of Public Policy, Georgia Institute of Technology; (4) Thema Monroe-White, Schar School of Policy and Government & Department of Computer Science, George Mason University (tmonroew@gmu.edu). Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 1.1 Related Work and Contributions 1.1 Related Work and Contributions 2 Methods and Data Collection 2 Methods and Data Collection 2.1 Textual Identity Proxies and Socio-Psychological Harms 2.1 Textual Identity Proxies and Socio-Psychological Harms 2.2 Modeling Gender, Sexual Orientation, and Race 2.2 Modeling Gender, Sexual Orientation, and Race 3 Analysis 3.1 Harms of Omission 3.1 Harms of Omission 3.2 Harms of Subordination 3.2 Harms of Subordination 3.3 Harms of Stereotyping 3.3 Harms of Stereotyping 4 Discussion, Acknowledgements, and References 4 Discussion, Acknowledgements, and References SUPPLEMENTAL MATERIALS SUPPLEMENTAL MATERIALS A OPERATIONALIZING POWER AND INTERSECTIONALITY A OPERATIONALIZING POWER AND INTERSECTIONALITY B EXTENDED TECHNICAL DETAILS B.1 Modeling Gender and Sexual Orientation B.1 Modeling Gender and Sexual Orientation B.2 Modeling Race B.2 Modeling Race B.3 Automated Data Mining of Textual Cues B.3 Automated Data Mining of Textual Cues B.4 Representation Ratio B.4 Representation Ratio B.5 Subordination Ratio B.5 Subordination Ratio B.6 Median Racialized Subordination Ratio B.6 Median Racialized Subordination Ratio B.7 Extended Cues for Stereotype Analysis B.7 Extended Cues for Stereotype Analysis B.8 Statistical Methods B.8 Statistical Methods C ADDITIONAL EXAMPLES C.1 Most Common Names Generated by LM per Race C.1 Most Common Names Generated by LM per Race C.2 Additional Selected Examples of Full Synthetic Texts C.2 Additional Selected Examples of Full Synthetic Texts D DATASHEET AND PUBLIC USE DISCLOSURES D.1 Datasheet for Laissez-Faire Prompts Dataset D.1 Datasheet for Laissez-Faire Prompts Dataset 3 ANALYSIS 3.1 Harms of Omission The first harm we consider is omission. To quantify this harm, we begin by restricting our analysis to power-neutral prompt responses and measuring statistical deviations from the U.S. Census. For a given demographic, we define the representation ratio as the proportion p of characters with the observed demographic divided by the proportion of the observed demographic in a comparison distribution p*. Here, demographic could be any combination of race, gender, and/or sexuality. We compute gender and sexuality proportions directly from gender reference mappings (see Table S9), and model race using fractional counting: This allows us to understand to what degree synthetic texts from LMs correlate with or amplify the underrepresentation of minoritized groups beyond real-world patterns (similarly shown by [59] in generative text-to-image models). Fig. 1ai shows that White characters are the most represented across all domains and models, from 71.0% (Learning, ChatGPT3.5) to 84.1% (Love, PaLM2). The next best-represented race only reaches a 13.2% likelihood (Love, Claude2.0, Latine). Compared to the US Census, median representation for racially minoritized characters ranges from ratios of 0.22 (MENA, Labor) to 0.66 (NH/PI, Labor), while White characters are over-represented at a median ratio of over 1.25 in Learning to 1.34 in Labor (Fig. 1aii). This means that for names reflecting any minoritized race, their representation is 33% (i.e., NH/PI, Labor) to 78% (i.e., MENA, Labor) overall less likely to appear in LM-generated stories, while White names are up to 34% more likely to appear relative to their representation in the U.S. Census. Meanwhile, gender representation is predominantly binary, skewing towards more feminized character representation overall, particularly for fictional students in the Learning domain (except for ChatGPT 4, which skews masculinized). Concerning gender, characters with non-binary pronouns are represented less than 0.5% of the time in all models except ChatGPT3.5 (3.9% in Learning). Binary gender representation ratiosskew slightly feminine (1.07) for all domains, whereas non-binary genders are under-represented by an order of magnitude (0.10) compared to Census levels (see Figure 1aii). Non-heterosexual romantic relationships are similarly underrepresented and are depicted in less than 3% of generated stories, with median representation ratios ranging from 0.04 (NB-NB) to 0.28 (F-F) (Fig. 1b). Therefore, we find that all five generative LMs exacerbate patterns of omission for minoritized identity groups beyond population-level differences in race, gender, and sexual orientation (with p-values of < .001 across nearly every combination of model and domain). In Fig. 1c we illustrate additional harms of omission specifically for NH/PI and AI/AN names, where we find little to no representation above a racial likelihood threshold of 24% (NH/PI) and 10% (AI/AN). Notably, this pattern of omission also holds for intersectional non-binary identities, where models broadly represent non-binary identified characters with predominantly White names (Fig 1d). These baseline findings indicate that LMs broadly amplify the social omission of minoritized groups in response to power-neutral prompts. The extent of omission vastly exceeds the degree that would be merely expected from the overall under-counting of minoritized groups in U.S. Census datasets [60]. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv available on arxiv

Part of HackerNoon's growing list of open-source research papers, promoting free access to academic material.

How AI Assigns Power Based on Race and Gender

How AI Texts Reinforce Harmful Stereotypes—Even Without Asking

Bias in Algorithmic Technology Exposed 

Too Long; Didn't Read

Get Free Unlimited IP Geo Data With IPinfo Lite

The Omission Bias in LMs Is Worse Than You Think

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

A Guide to Exploiting Unknown Strategies From a Known Bias Set

New Study Reveals Subtle but Widespread AI Discrimination

Exploring Social Bias in AI Through 50 Real-World Scenarios

How AI Texts Reinforce Harmful Stereotypes—Even Without Asking

How AI Assigns Power Based on Race and Gender

What AI Gets Wrong About Queer, Indigenous, and Racial Minorities

A Guide to Exploiting Unknown Strategies From a Known Bias Set

New Study Reveals Subtle but Widespread AI Discrimination

Exploring Social Bias in AI Through 50 Real-World Scenarios

How AI Texts Reinforce Harmful Stereotypes—Even Without Asking

How AI Assigns Power Based on Race and Gender

What AI Gets Wrong About Queer, Indigenous, and Racial Minorities

Light-Mode

Classic

Newspaper

Dark-Mode

Neon Noir

Minty

HN StartUps