“It’s Complicated”: Negotiating Accessibility and (Mis)Representation in Image Descriptions of Race, Gender, and Disability

By Cynthia L. Bennett, Cole Gleason, Morgan Klaus Scheuerman, Jeffrey P. Bigham, Anhong Guo, Alexandra To. 2021

146 citations recorded by the source corpus.

Content creators are instructed to write textual descriptions of visual content to make it accessible; yet existing guidelines lack specifics on how to write about people’s appearance, particularly while remaining mindful of consequences of (mis)representation. In this paper, we report on interviews with screen reader users who were also Black, Indigenous, People of Color, Non-binary, and/or Transgender on their current image description practices and preferences, and experiences negotiating theirs and others’ appearances non-visually. We discuss these perspectives, and the ethics of humans and AI describing appearance characteristics that may convey the race, gender, and disabilities of those photographed. In turn, we share considerations for more carefully describing appearance, and contexts in which such information is perceived salient. Finally, we offer tensions and questions for accessibility research to equitably consider politics and ecosystems in which technologies will embed, such as potential risks of human and AI biases amplifying through image descriptions.

Subfields: Computer Vision and Pattern Recognition, Cultural Studies, Sociology and Political Science, Safety, Robustness & Interpretability, Human-Computer Interaction

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Cultural Studies

In Sociology and Political Science

In Safety, Robustness & Interpretability

In Human-Computer Interaction

Citations in the corpus

Evaluating the Social Impact of Generative AI Systems in Systems and Society (2023, background): “These can include choices about many stages of the optimization process [237, 129]; privacy constraints [24], widely used compression techniques [109, 15, 169] and the choice hardware [273] have all been found to amplify harm on underrepresented protected attributes [28].”
Multimodal datasets: misogyny, pornography, and malignant stereotypes (2021, background): “…data dump of alt-text is besotted with high prevalence of issues such as missing important information, not being descriptive enough, resorting to stereotypical and offensive descriptors, being over descriptive (including ﬁlenames and special characters) or misrepresenting images [32, 30, 33].”
Sociotechnical Harms: Scoping a Taxonomy for Harm Reduction (2022, background): “Prior work identifies representational harms in many algorithmic systems, including through classifiers [40], natural language processing [31], computer vision [104], and image tagging and captioning [27].”
Taxonomizing and Measuring Representational Harms: A Look at Image Tagging (2023, background): “In contrast, the developers of an image tagging system for blind and low-vision users might decide that the consequences of erasing social groups are worse than the consequences of incorrectly applying tags (Bennett et al. 2021; Hanley et al. 2021).”
Directional Bias Amplification (2021, background): “Although this may not always be the case, depending on the downstream application (Bennett et al., 2021), perhaps we could consider simply replacing all instances of gendered words like “man” and “woman” in the captions with “person” to trivially eliminate T !”
Underspecification in Scene Description-to-Depiction Tasks (2022, background): “Bias: As in image-to-text (Bennett et al., 2021), there are risks of text-to-image amplifying societal biases including those concerning gender, race, and disability.”
Sociotechnical Harms of Algorithmic Systems: Scoping a Taxonomy for Harm Reduction (2022, background): “As classification systems areusedacrossmany consequential domains,denying opportunities to self-identify can materially impact marginalized communities, ranging from nonconsensual inclusion in datasets to surveillance and wrongful arrest [28].”