This webinar too place on April 18, 2024, 3:00pm-4:00pm EST.
Click here for the recording
Recent generative AI – particularly models using an LLM architecture – have wowed and produced anxiety in equal parts with their seemingly limitless ability to generate text, images, answer increasingly complex questions, and perform sophisticated problem-solving tasks. Initially, this led many to opine that these systems were reasoning, or ‘thinking;’ what’s more, for some, these successes evidenced that thinking is reducible to those probabilistic analyses and calculations that constitute current AI. Notable critics like Judea Pearl and Angus Fletcher, however, see intrinsic limits to the current paradigm, and caution that current approaches in AI are incapable of genuine causal and narrative thinking/reasoning.
From a humanistic perspective grounded in narrative theory, my talk presents some recent findings which demonstrate that current-generation AI does not think, or reason, causally – a technical limitation that clearly differentiates it from human thinking. As a result of this incapacity, I argue, current AI cannot be said to possess narrative understanding, and therefore has a sharply limited ability to comprehend narratives at even a literal level. The seemingly paradoxical fact that AI can perform sophisticated feats of analysis and interpretation while failing rudimentary reading comprehension tasks (i.e. that it can extrapolate what a given text might ‘mean’ without understanding what happens) supports the idea of a specifically narrative/causal mode of thought that is not reducible to propositional computation and probabilistic prediction. In presenting a small sample of AI responses/analyses, I will argue that this shortcoming isn’t remedied with more data because it is a procedurally distinct way of thinking. In short, my research suggests that narrative (causal) thinking – the type honed and exercised in the humanities – is distinct from calculation, fundamental to our understanding of stories and events, and therefore essential for interacting with our world.
This Webinar forms part of the True Digital Poiesis Project.
______
Michael Benveniste received his BA and MA in English from UC Santa Barbara, and a PhD in English Literature from Stanford University. A first-generation college student, his teaching and research focus on multi-ethnic U.S. literature, narrative theory, politics and causal analysis. He now teaches at San Jose State University and is lead researcher for the Fletcher Lab at Project Narrative (OSU), where he works on narrative cognition and applied narrative theory. In addition, he runs a start-up that fuses expert/symbolic systems and AI for applied narrative analysis.