Collaborative Auditing of Machine Learning Applications: High School Youth Identifying Issues of Algorithmic Justice (Poster 4)

• In Event: Perspectives on Promoting Algorithmic Justice in Artificial Intelligence and Machine Learning Education in K–12

Abstract

Objectives: Current efforts to promote artificial intelligence/machine learning (AIML) literacy emphasize the importance of ensuring that learners have opportunities to engage with AIML ethically and critically by considering its potential benefits and harms (e.g., Touretzky & Gardner-McCune, 2022). Our poster presents results from an exploratory study in which we investigate how to incorporate algorithm audits into K-12 AIML education, with a focus on critical inquiry in ML contexts by having youth design projects and then collaboratively audit each other's projects.
Theoretical framework: We adapted algorithm auditing, a method that involves “repeatedly querying an algorithm and observing its output in order to draw conclusions about the algorithm’s opaque inner workings and possible external impact” (Metaxa et al., 2021), for youth to collaboratively and systematically investigate how their peers' ML projects work. We build on current research on the design process of ML applications (Tedre et al., 2021) in K-12 and recent studies on how non-expert adult users can audit algorithmic systems to find harmful algorithmic behaviors (Shen et al., 2021; DeVos, 2022; Lam et al, 2022), by having youth design and audit AIML projects.
Methods & Data Sources: We analyzed pre and post interviews from a two-week long workshop conducted in 2023 with 15 Black and Latinx youths (ages 14-16) during which they designed ML-powered applications (that included image, pose, and motion classifiers) and collaboratively conducted algorithmic audits of each others’ projects. During the interviews we presented participants with (1) faulty ML classifiers and their training datasets (2) an everyday algorithm auditing interview task that required them to analyze the results of an AIML image generator. These tasks were adapted from recent work in computing education research and everyday auditing research (DeVos et al., 2022; Solyst et al., in press). We analyzed the transcribed interviews in two rounds using thematic analysis (Braun and Clarke, 2012).
Results: When presented with faulty ML classifiers and their training data sets youths identified potential reasons for inappropriate behaviors. In pre-interviews, youths expressed concerns about the diversity of the data used to train the classifiers. In post-interviews they showed more nuanced understandings of data issues that may cause harmful bias identifying class imbalance issues, spurious relationships, and concrete next steps to create representative datasets. In the auditing task of an image generator, prior to the workshop youths voiced basic understandings of issues of algorithmic justice identifying potential racist and sexist behaviors in ML generated images. After the workshop, youths voiced how potential harmful biases and behaviors may emerge from the qualities of the data used to train the model and the biases of its creators.
Significance: We provide evidence of how audits may support youth in understanding that the harmful consequences and implications of ML applications are closely intertwined with functionality failures and dataset design (Raji et al., 2022). At the same time, after the workshop youth improved their computational communication skills (Lui et al., 2020) being able to discuss issues of algorithmic justice with greater detail and with concrete terms.

Authors

• Luis Morales-Navarro, University of Pennsylvania

• Yasmin B. Kafai, University of Pennsylvania