Computational Modeling of Stereotype Content in Text

Abstract

Stereotypes are encountered every day, in interpersonal communication as well as in entertainment, news stories, and on social media. In this study, we present a computational method to mine large, naturally occurring datasets of text for sentences that express perceptions of a social group of interest, and then map these sentences to the two-dimensional plane of perceived warmth and competence for comparison and interpretation. This framework is grounded in established social psychological theory, and validated against both expert annotation and crowd-sourced stereotype data. Additionally, we present two case studies of how the model might be used to answer questions using data “in-the-wild,” by collecting Twitter data about women and older adults. Using the data about women, we are able to observe how sub-categories of women (e.g., Black women and white women) are described similarly and differently from each other, and from the superordinate group of women in general. Using the data about older adults, we show evidence that the terms people use to label a group (e.g., old people vs. senior citizens) are associated with different stereotype content. We propose that this model can be used by other researchers to explore questions of how stereotypes are expressed in various large text corpora.

Ethical and methodological challenges in building morally informed AI systems

“…Fairness -Bias. Fairness is, by far, the most discussed issue in the literature, remaining a paramount concern especially in case of LLMs and text-to-image models 35,[45][46][47] . This is sparked by training data biases propagating into model outputs 48 , causing negative effects like stereotyping 33,35 , racism 49 , sexism 50 , ideological leanings 47 , or the marginalization of minorities 51 .…”

“…Fairness is, by far, the most discussed issue in the literature, remaining a paramount concern especially in case of LLMs and text-to-image models 35,[45][46][47] . This is sparked by training data biases propagating into model outputs 48 , causing negative effects like stereotyping 33,35 , racism 49 , sexism 50 , ideological leanings 47 , or the marginalization of minorities 51 . Next to attesting generative AI a conservative inclination by perpetuating existing societal patterns 52 , there is a concern about reinforcing existing biases when training new generative models with synthetic data from previous models 53 .…”

Analyzing the color availability of AI‐generated posters based on K‐means clustering: 74% orange, 38% cyan, 32% yellow, and 28% blue‐cyan

“…As can be seen from Figure 8, in terms of the use of value, the AI uses mostly low‐value colors in its main color selection, this may be related to Midjourney's tendency to generate darker tones, 84 while the human‐designed posters and the AI‐generated posters both use mainly high‐value colors in accordance with the second and third colors. When comparing the average value of the poster colors, the AI‐generated posters are in the medium‐value area with a mean value of 0.57 (SD = 0.29) and the human‐designed posters are also in the medium‐value area with a mean value of 0.62 (SD = 0.28).…”

From Development to Dissemination: Social and Ethical Issues with Text-to-Image AI-Generated Art

“…These systems were chosen because they are popular implementations of different state-of-the-art text-to-image generative AI techniques using diffusion models [12] and CLIP image embeddings [13] that were developed primarily by researchers and academics. These models have also been studied in academic research for the content nature, and cultural and social biases of their generated outputs [14,15,16,17,18,19,20]. For this experiment, Stable Diffusion's v1-4 and v2-1 pretrained weights will each be used independently in conjunction with Stable Diffusion Web UI.…”