Artificial Intelligence Ethics Education in Dance and Creative Computing Spaces

• In Event: Artificial Intelligence in Education: Potentials and Challenges in Ethics, Data Privacy, and Pedagogy

Abstract

Approaches on teaching about the ethical implications of AI systems often focus on the identification and operationalization of biases within AI systems that marginalize and harm individuals and communities. Through dance computing and partnerships with community organizations like STEM From Dance (SFD), we teach learners about AI systems and their impacts, engage in discourse on how to counter inequities perpetuated by AI algorithms and navigate these barriers within equitable, inclusive, and culturally sustaining ways and spaces.

Lee et al. (2021) describe students’ engagement with data from three “layers”—personal experiences with data and data collection contexts, cultural practices and communities that shape how data is constructed and used, and sociopolitical forces that speak to the way data reflects or are shaped by power dynamics and capitalist discourse. We draw on these layers in co-designing curricula with teachers and students that connect dance with AI and creating space for discourse around how AI algorithms that process body position data to detect dance poses are shaped by design decisions that encapsulate biases and inequities.

In summer 2020, we developed XXX (Author, 2021), a web-based, free creative coding environment for easily accessing the results of pose detection algorithms run on dance videos uploaded by learners or realtime webcam feeds. In its initial conception, XXX enables the creation of music videos featuring real dancers interacting with virtual animations. XXX was also used by 6 SFD alumni over a 15-week remote study, marketed as a paid educational internship, as they learned XXX features, worked on individual and group projects, and offered feedback for how to improve the XXX system and curriculum for future learners. All XXX capabilities can be accessed without providing personal information (i.e., no logins) and pose detection is run client-side, meaning no videos are stored.

We gathered numerous creative artifacts generated by learners including brainstorming documents, code, raw dance videos, and final performance videos generated by XXX. In the 15-week study, we gathered transcripts recorded from learning sessions and blogs maintained by participants. Additionally, we account for curricular materials researchers and community partners have developed in teaching with XXX over the last two years.

Because of XXX’s initial conception as a tool for creative expression, curricular materials emphasized system features, language affordances, and examples of prior work to make learners aware of artistic directions they may take their projects (Author, 2022). However, XXX uses existing AI systems for pose detection trained on inaccessible datasets, thus, when XXX performs poorly, users must speculate on possible causes. Furthermore, pose detection systems make assumptions about the human body that are not true of all human bodies (e.g., that there are always 33 key points to identify). We see these limitations as opportunities to develop new curricular materials and interface components that highlight biases implemented in flawed AI systems, and give learners strategies for circumventing these within XXX and similar systems.

Authors

• Kayla DesPortes, New York University
  Presenting Author

• Francisco Castro, New York University
  Non-Presenting Author

• William Payne, New York University
  Non-Presenting Author