To Bot or Not to Bot, that's the question

• In Event: Artificial Intelligence and Cognition as Social Praxis

Abstract

While the need to factor in context is frequently flagged in AI research today, the challenge is in operationalising what we mean by “context” and whether/how it can be built into AI agents designed for variegated action. We propose a framework that starts with the premise that human activity and interaction is always situated in a broader social and cultural context, and builds on work that conceptualizes cognition as a situated and embodied activity that “occurs in very particular (and often very complex) environments” (Anderson 2003, 91). Our framework distinguishes between “learning to do” -- learning how to perform specific tasks – and “learning to be” -- how to be the person who undertakes these tasks (Lave 1988; Lave and Wenger 1991; Wenger 1999). For instance, there is a difference between learning how to stitch cloth and learning how to be a well-regarded tailor, but both help a person in their subsequent trajectory (Lave 2011). We question whether current models of learning in AI work well for achieving both these types of learning. If they do not, what kinds of interactions can we expect AI agents to handle successfully and which ones should be left to humans? We use the example of designing a chatbot for the administration of academic admissions to show how employing our framework can help us decide what chatbots can and can’t do. Our paper will contribute to ongoing conversations in the STS community around the nature, potential and limits of machine learning and AI in present times.

Authors

• Sachit Rao, International Institute of Information Technology - Bangalore

• Bidisha Chaudhuri, International Institute of Information Technology Bangalore

• Janaki Srinivasan, International Institute of Information Technology - Bangalore