Analyzing internal world models of humans, animals and AI

Scientists propose a formal framework to describe internal world models across humans, animals, and artificial intelligence. This approach unifies task, neural, and conceptual spaces, allowing researchers to analyze how organisms abstract general laws from experience to predict outcomes in unfamiliar situations. By linking external experiences with internal brain activity, these models become scientifically tangible and offer a shared language for interdisciplinary study. This structured understanding highlights significant limitations in current AI systems, which lack the capacity for genuine planning and plausibility checking. Unlike biological entities, many AI models operate merely as pattern recognizers without the ability to simulate strategies or correct errors before acting. Bridging this gap requires integrating the dynamic, predictive capabilities found in natural internal world models into artificial systems to enhance their robustness and safety. Furthermore, analyzing these models holds promise for addressing mental health challenges, as deficits in world model formation may underlie conditions like depression or schizophrenia. Improved comprehension of these mechanisms could lead to more targeted therapies and medications. For open data communities, this framework demonstrates the value of standardized, cross-species data integration, enabling the development of more intelligent and reliable AI systems grounded in biological insights.

Source: sciencedaily.com
Published on 2024-07-19