Cognition and the Machine: Exploring Human-AI Interaction via Predictive Processing

Artificial Intelligence is rapidly becoming ubiquitous in both professional settings and everyday life. The prospect is that modern machine-learning technology of Generative AI could help with decision-making, and thinking and reasoning alike. However, our understanding of human cognition, particularly thinking and reasoning, remains limited. This thesis proposes a conceptual framework for understanding the interaction between humans and AI, drawing on the cognitive theory of Predictive Processing. Predictive Processing posits that the brain constructs and employs models to predict and interpret the external world based on past experiences. It operates through a top-down simulation, where sensory feedback primarily highlights discrepancies between predictions and reality. This simulation extends to imagination, reasoning, and thinking, facilitated by offline mental simulations using domain-specific predictive models. Prediction models are structured hierarchically, with lower levels corresponding to concrete sensory input, while higher levels represent abstract, generalized relationships within specific domains. In human communication, abstract predictions play a crucial role, as individuals mentally simulate predictions and convey them through language to peers who then interpret them based on their own domain-models. Chatbots has the potential of being a similar kind of dialogue partner, evaluating your predictions of the world, helping you with interpretations or mutual examinations of problems. For human users it may seem as if it is a shared dialogue, but what AI based on current technology most of all lack, is the worldly connection and adaptive flexibility of the model it is based upon.