Learning Machine Learning with Personal Data Helps Stakeholders Ground Advocacy Arguments in Model Mechanics

Abstract

Machine learning systems are increasingly a part of everyday life, and often used to make critical and possibly harmful decisions that affect stakeholders of the models. Those affected need enough literacy to advocate for themselves when models make mistakes. To understand how to develop this literacy, this paper investigates three ways to teach ML concepts, using linear regression and gradient descent as an introduction to ML foundations. Those three ways include a basic Facts condition, mirroring a presentation or brochure about ML, an Impersonal condition which teaches ML using some hypothetical individual's data, and a Personal condition which teaches ML on the learner's own data in context. Next, we evaluated the effects on learners' ability to self-advocate against harmful ML models. Learners wrote hypothetical letters against poorly performing ML systems that may affect them in real-world scenarios. This study discovered that having learners learn about ML foundations with their own personal data resulted in learners better grounding their self-advocacy arguments in the mechanisms of machine learning when critiquing models in the world.