Lifting the Curtain on AI Use through Donated Conversations: When Confessional Data Demands Academic Curiosity

Find all information about this seminar here: ABRI Lunch Seminar Ward van Zoonen - VU Amsterdam

Abstract

Generative AI took the world by storm, inviting utopian and dystopian accounts of what these technologies mean for how we work, organize, and collaborate. Interestingly, how people actually use these technologies remains largely hidden as user conversations happen in private, and disclosure of use is highly contested. As a result, we have been left theorizing in the blind. Until now...

The seminar introduces a line of research that aims to lift the curtain and shift us from speculation and self-reported uses to empirically grounded accounts of what GenAI is used for. The seminar will inform you about: a donated ChatGPT archive comprised of the full conversation history of 450 ChatGPT users. In total, 1.1 million turns (prompt-response sequences) are paired with survey data from the same users. In the seminar, we'll show how we collected the data, what such a data export can look like, and what we have done with it so far. We'll share some early quantifications of the data and offer an insight into papers in progress. We show that this data offers unique confessional archives that aren't curated for a specific audience, and offer a unique portal into how work is being mediated by GenAI. The data also goes far beyond work - conversations reflect professional uses, worries about life, social activities, and personal consultations.

The seminar is a mixture of methodology, phenomenology, and discussion. We'll show you the pipeline for how to run a data donation study and how we are approaching analysis. We also encourage discussion from a hopefully interdisciplinary audience for example on: what are people actually doing with ChatGPT when their colleagues aren't watching, and what does this mean for (studies of) work and organising? How does this data stretch the foci of our research - into new and at times uncomfortable territory? How can we make sense of what can be done with this data?