Get started
Work

‘Person’ == Light-skinned, Western Man, and Sexualization of Women of Color: Stereotypes in Stable Diffusion

Year: 2023
Type: conference-paper
Abstract: We study stereotypes embedded within one of the most popular text-to-image generators: Stable Diffusion.We examine what stereotypes of gender and nationality/continental identity does Stable Diffusion display in the absence of such information i.e. what gender and nationality/continental identity is assigned to 'a person', or to 'a person from Asia'.Using vision-language model CLIP's cosine similarity to compare images generated by CLIP-based Stable Diffusion v2.1 verified by manual examination, we chronicle results from 136 prompts (50 results/prompt) of front-facing images of persons from 6 different continents, 27 nationalities and 3 genders.We observe how Stable Diffusion outputs of 'a person' without any additional gender/nationality information correspond closest to images of men (avg.cosine similarity 0.64) and least with persons of nonbinary gender (avg.cosine similarity 0.41), and to persons from Europe/North America (avg.cosine similarities 0.71 and 0.68, respectively) over Africa/Asia (avg.cosine similarities 0.43 and 0.41, respectively), pointing towards Stable Diffusion having a concerning representation of personhood to be a European/North American man.We also show continental stereotypes and resultant harms e.g. a person from Oceania is deemed to be Australian/New Zealander (avg.cosine similarities 0.77 and 0.74, respectively) over Papua New Guinean (avg.cosine similarity 0.31), pointing to the erasure of Indigenous Oceanic peoples, who form a majority over descendants of colonizers both in Papua New Guinea and in Oceania overall.Finally, we unexpectedly observe a pattern of oversexualization of women, specifically Latin American, Mexican, Indian and Egyptian women relative to other nationalities, measured through an NSFW detector.This demonstrates how Stable Diffusion perpetuates Western fetishization of women of color through objectification in media, which if left unchecked will amplify this stereotypical representation.Image datasets are made publicly available.
DOI: 10.18653/v1/2023.findings-emnlp.465
Language: English
Subfield: Safety Research
Open Access status: hybrid
Award: PURPOSE: ARTIFICIAL INTELLIGENCE (AI) IS WIDESPREAD IN TODAY'S SOCIETY, WITH AI ALGORITHMS BEING INCREASINGLY RELIED ON TO MAKE CONSEQUENTIAL DECISIONS ABOUT INDIVIDUALS. ALTHOUGH THERE WAS INITIAL OPTIMISM THAT AI APPLICATIONS WOULD MAKE A MORE EFFICIENT AND FAIR SOCIETY, ANECDOTAL EXAMPLES OF BIASED OUTCOMES IN AI APPLICATIONS NOW ABOUND FROM DOMAINS AS WIDE AS WEB SEARCH, MACHINE TRANSLATION, UNIVERSITY ADMISSIONS, RESUME SCREENING, AND JOB CANDIDATE ASSESSMENT. YET, BEYOND THESE ANECDOTES, THERE REMAINS NO COMPREHENSIVE STUDY THAT EXAMINES (1) HOW THE EXTENT OF BIASES IN GENERAL-PURPOSE MODALITY REPRESENTATIONS THAT AI RELIES ON PROPAGATE TO DOWNSTREAM DECISIONS, (2) HOW BIASED AI DECISIONS INTERACT WITH THE BIASES OF HUMANS-IN-THE-LOOP, AND (3) HOW AI-MEDIATED OUTCOMES IMPACT THE BIASES OF FUTURE GENERATIONS OF AI MODELS. THE PROPOSED RESEARCH SEEKS TO ADDRESS THESE NEEDS, THEREBY CONTRIBUTING TO NIST'S TRUSTWORTHY AND RESPONSIBLE AI EFFORTS. ACTIVITIES TO BE PERFORMED: IN THIS PROJECT, THE PROPOSERS WILL ANALYZE THE IMPACT OF BIASED GENERAL-PURPOSE MACHINE REPRESENTATIONS GENERATED VIA UNSUPERVISED MACHINE LEARNING ON THE AUTOMATED DECISIONS OF SUPERVISED MACHINE LEARNING ALGORITHMS THAT RELY ON STRUCTURED DATA IN THE EMPLOYMENT DOMAIN. AS A STARTING POINT, THE PROPOSERS WILL FOCUS ON GENDER AND RACIAL BIAS IN RESUME SCREENING AND JOB CANDIDATE ASSESSMENT, AN APPLICATION DOMAIN WITH CLEAR CONSEQUENCES FOR LIFE'S OPPORTUNITIES AND OUTCOMES. THE PROPOSERS WILL DEVELOP A BIAS PROPAGATION AND INTERACTION AUDIT FRAMEWORK THAT GENERALIZES ACROSS MODALITIES, BIAS TYPES, AND APPLICATION DOMAINS. EXPECTED OUTCOMES: THE PROPOSED RESEARCH WILL CONDUCT TWO COMPLEMENTARY STUDIES TO ANSWER THE FOLLOWING IMPORTANT QUESTIONS: (1) GIVEN THAT BIAS REFLECTED FROM SOCIETY TO LARGE-SCALE DATASETS GETS EMBEDDED IN GENERAL-PURPOSE MACHINE REPRESENTATIONS USED FOR PROCESSING DATA, HOW DOES REPRESENTATIONAL BIAS PROPAGATE TO DOWNSTREAM APPLICATIONS? (2) HOW ARE HUMANS-IN-THE-LOOP COLLABORATING WITH AI IN DECISION MAKING AFFECTED BY THE BIASED OUTCOMES OF MACHINES,GIVEN THEIR OWN IMPLICIT BIASES? THE RESEARCH STUDIES SUPPORTED BY THIS GRANT WILL BE PUBLISHED IN THE SCIENTIFIC LITERATURE, SPECIFICALLY BY CONTRIBUTING PAPERS ON AN ANNUAL BASIS AT ARTIFICIAL INTELLIGENCE AND FAIRNESS VENUES FOR COMPUTER AND INFORMATION SCIENCES. INTENDED BENEFICIARIES: SOCIOTECHNICAL ARTIFICIAL INTELLIGENCE (AI) SYSTEMS WITH HUMANS-IN-THE-LOOP MAKE CONSEQUENTIAL DECISIONS ABOUT INDIVIDUALS IN SCENARIOS THAT DETERMINE LIFE'S OPPORTUNITIES AND OUTCOMES. DEVELOPING METHODS AND METRICS TO MEASURE HOW THE BIASES OF SOCIETY, AI, AND HUMANS-IN-THE-LOOP INTERACT IN DECISION MAKING WILL PROVIDE QUANTITATIVE EVIDENCE ON THE IMPACT OF AI ON SOCIETY. THIS PROJECT WILL BENEFIT THE GENERAL PUBLIC BY PROVIDING MUCH-NEEDED RESEARCH TO SUPPORT DEVELOPING STANDARDS FOR IDENTIFYING AND MANAGING BIAS IN AI. FAIR WITH HARMFUL BIAS MANAGED IS A KEY CHARACTERISTIC OF TRUSTWORTHY AI AS DEFINED IN NIST'S AI RISK MANAGEMENT FRAMEWORK, THEREFORE THIS BIAS RESEARCH WILL CLEARLY BENEFIT NIST'S LARGER TRUSTWORTHY AND RESPONSIBLE AI PROGRAM AND ADVANCE THE STATE-OF-THE-ART FOR BIAS EVALUATIONS IN AI SYSTEMS. SUBRECIPIENT ACTIVITIES: THERE ARE NO PLANNED SUBAWARDS.
OpenAlex ID: W4389520170
FWCI: 8.01
Cites: 32
Cited by: 32
Related to: 10