Artsheets for Art Datasets

Machine learning (ML) techniques are increasingly being employed within avariety of creative domains. For example, ML tools are being used to analyzethe authenticity of artworks, to simulate artistic styles, and to augment humancreative processes. While this progress has opened up new creative avenues, ithas also created the opportunity for adverse downstream effects such as culturalappropriation (e.g., cultural misrepresentation, offense, and undervaluing) andrepresentational harm. Many of the concerning issues stem from the training datain ways that diligent evaluation can uncover, prevent, and mitigate. As such, whendeveloping an arts-based dataset, it is essential to consider the social factors thatinfluenced the process of conception and design, and the resulting gaps must beexamined in order to maximize understanding of the dataset’s meaning and futureimpact. Each dataset creator’s decision produces opportunities, but also omissions.Each choice, moreover, builds on preexisting histories of the data’s formation andhandling across time by prior actors including, but not limited to, art collectors,galleries, libraries, archives, museums, and digital repositories. To illuminate theaforementioned aspects, we provide a checklist of questions customized for usewith art datasets in order to help guide assessment of the ways that dataset designmay either perpetuate or shift exclusions found in repositories of art data. Thechecklist is organized to address the dataset creator’s motivation together withdataset provenance, composition, collection, pre-processing, cleaning, labeling,use (including data generation), distribution, and maintenance. Two case studiesexemplify the value and application of our questionnaire.