Key points
- Crawford and Paglen analyse ImageNet and document how people were labelled with demeaning and biased categories.
- The essay established the politics of image datasets as a central AI-ethics problem for visual media.
- It established training-data politics as a central AI-ethics problem for visual media, with vocabulary that travels directly into generative-image critique.
Summary
Crawford and Paglen's "Excavating AI" examines the ImageNet large-scale image dataset, documenting the demeaning, biased, and harmful categories used to label photographs of people. First published as an essay in 2019, with a peer-reviewed version appearing in the journal AI and Society in 2021, it established the politics of image training data as a central problem in AI ethics for anyone working with visual media. The essay is widely assigned in AI-ethics and critical media courses and is the founding text for discussing how training datasets reproduce social bias.
Related items
Source
Source: Excavating AI (essay; AI and Society, 2021) ↗ (Research)
Cite this item
Crawford, K. & Paglen, T. (2019). ‘Excavating AI (the politics of image training data)’, Excavating AI (essay; AI and Society, 2021). Available at: https://excavating.ai/
Your reference manager can also read this page directly: with the Zotero (or Mendeley) browser connector installed, save it straight to your library. Whole-collection exports: RIS, BibTeX, CSL-JSON.