Computer Vision and Conflicting Values: Describing People with Automated Alt Text

Accessibility vs. Accuracy Creates a Core Dilemma: Automated alt text tools like Facebook's serve the critical purpose of ensuring blind and low vision people can access image content, yet the paper demonstrates that companies face genuine ethical tensions when deciding what information to include—there are "no easy ways to avoid the normative dilemmas" posed by automation in this context.

Facebook Avoids Describing Identity Categories: The paper reveals that Facebook's automatic alt text policy deliberately excludes identity descriptors such as race, gender, and age, reflecting the company's concern about potential bias and discrimination, though this choice itself raises questions about what information is being withheld from users.

Museums Offer a Different Model: The paper contrasts Facebook's approach with museum practices, which take a more contextual, manual approach to describing cultural artifacts in alt text—suggesting that different communities have developed distinct philosophies about what descriptive information is appropriate and necessary.

No Single "Right Answer" Exists: The authors examine alternative strategies to address these concerns but conclude that "there are no easy ways" to resolve the fundamental tension between providing useful descriptive detail and avoiding potential harms, indicating this is an ongoing challenge requiring ongoing deliberation rather than a problem with a technical solution.

Policy Choices Reflect Unstated Values: The paper demonstrates that decisions about what to automate and what to exclude in alt text descriptions are fundamentally normative choices that reflect companies' assumptions about bias, privacy, and user needs—not purely technical or neutral decisions.