Capturing the unobservable in AI development: proposal to account for AI developer practices with ethnographic audit trails (EATs)

Abstract The prevalence of artificial intelligence (AI) tools has inspired social studies researchers, ethicists, and policymakers to seriously examine AI’s sociopolitical and ethical impacts. AI ethics literature provides guidance on which ethical principles to implement via AI governance; AI auditing literature, especially ethics-based auditing (EBA), suggests methods to verify if such principles are respected in AI model development and deployment. As much as EBA methods are abundant, I argue that most currently take a top-down and post-hoc approach to AI model development: Existing EBA methods mostly assume a preset of high-level, abstract principles that can be applied universally across contexts; meanwhile, current EBA is only conducted after the development or deployment of AI models. Taken together, these methods do not sufficiently capture the very developmental practices surrounding the constitution of AI models on a day-to-day basis. What goes on in an AI development space and the very developers whose hands write codes, assemble datasets, and design model architectures remain unobserved and, therefore, uncontested. I attempt to address this lack of documentation on AI developers’ day-to-day practices by conducting an ethnographic “AI lab study” (termed by Florian Jaton), demonstrating just how much context and empirical data can be excavated to support a whole-picture evaluation of AI models’ sociopolitical and ethical impacts. I then propose a new method to be added to the arsenal of EBA: Ethnographic audit trails (EATs), which take a bottom-up and in-progress approach to AI model development, capturing the previously unobservable developer practices.