American == White in Multimodal Language-and-Image AI

Abstract

Three state-of-the-art language-and-image AI models, CLIP, SLIP, and BLIP, are evaluated for evidence of a bias previously observed in social and experimental psychology: equating American identity with being White. Embedding association tests (EATs) using standardized images of self-identified Asian, Black, Latina/o, and White individuals from the Chicago Face Database (CFD) reveal that White individuals are more associated with collective in-group words than are Asian, Black, or Latina/o individuals, with effect sizes > .4 for White vs. Asian comparisons across all models. In assessments of three core aspects of American identity reported by social psychologists, single-category EATs reveal that images of White individuals are more associated with patriotism and with being born in America, but that, consistent with prior findings in psychology, White individuals are associated with being less likely to treat people of all races and backgrounds equally. Additional tests reveal that the number of images of Black individuals returned by an image ranking task is more strongly correlated with state-level implicit bias scores for White individuals (Pearson's 𝜌 = .63 in CLIP, 𝜌 = .69 in BLIP) than are state demographics (𝜌 = .60), suggesting a relationship between regional prototypicality and implicit bias. Three downstream machine learning tasks demonstrate biases associating American with White. In a visual question answering task using BLIP, 97% of White individuals are identified as American, compared to only 3% of Asian individuals. When asked in what state the individual depicted lives in, the model responds China 53% of the time for Asian individuals, but always with an American state for White individuals. In an image captioning task, BLIP remarks upon the race of Asian individuals as much as 36% of the time, but never remarks upon race for White individuals. Finally, provided with an initialization image from the CFD and the text "an American person, " a synthetic image generator (VQGAN) using the text-based guidance of CLIP lightens the skin tone of individuals of all races (by 35% for Black individuals, based on pixel brightness). The results indicate that biases equating American identity with being White are learned by language-and-image AI, and propagate to downstream applications of such models. CCS CONCEPTS• Computing methodologies → Computer vision; Natural language processing.

A Friendly Face: Do Text-to-Image Systems Rely on Stereotypes when the Input is Under-Specified?

“…Recently, work has begun that focuses on particular biases that have emerged in multi-modal language-vision machine learning systems. Wolfe and Caliskan (2022) reported that racial biases about American identity, previously observed in social psychology, are learned by multi-modal embedding models and propagated to downstream tasks. Other works proposed evaluation frameworks to assess biases in text-to-image systems (Cho, Zala, and Bansal 2022) or training mechanisms such as adversarial learning to reduce representation biases in language-vision models (Berg et al 2022).…”

“…At the same time, commentators have noted potential ethical issues related to the use of copyrighted artworks in the training sets, the generation of hateful and offensive content, as well as issues of bias and diversity in the model outputs. Relating to the latter, research work has begun to audit the output of such models, investigating stereotypical associations between occupations and particular races and genders (Cho, Zala, and Bansal 2022), as well as between the word "American" and lighter skin colours (Wolfe and Caliskan 2022).…”

Fairness in Language Models Beyond English: Gaps and Challenges

“…Although there has been research on fairness in multimodal contexts (Wolfe and Caliskan, 2022;Wolfe et al, 2022), in a first-of-its-kind study, Wang et al (2022) looks at fairness from a multilingual view in multimodal representations. Whilst they find that multimodal representations may be individually fair, i.e., similar text representations across languages translate to similar images, this concept of fairness does not extend across multiple groups.…”

Disambiguating Algorithmic Bias: From Neutrality to Justice

“…Research has likewise uncovered extensive evidence of bias in vision-language AI models trained on internet-scale data [29,145,146]. Data sets used to train vision-language models have been found to contain "troublesome and explicit images and text pairs of rape, pornography, malign stereotypes, racist and ethnic slurs, and other extremely problematic content" [32].…”