Powers of the Soul Beyond AI

Could Large Language Models (LLMs) exhibit rational characteristics traditionally attributed to the human soul? I argue that five features of human rationality will likely remain beyond LLMs and other adaptive physical systems. Insight into truth: using billions of pages of text, a LLM may harvest a sound rule of inference. However the LLM has no insight into why the rule is true. Meta-insight: both humans and machines can follow instructions that constitute an infinite loop. Yet humans can, but machines cannot, recognize that they are in an infinite loop. Free will: once humans realize they are trapped in a loop, they can exercise free will to break out of the loop. By contrast, when a machine is trapped in an infinite loop, an external intervention is required to end the task. Access to necessary conceptual relations: LLMs are inductive learners and cannot justify universal necessary truths. By contrast, a human being can, via insight, see that a conceptual relation is necessarily true. Non-combinatorial creativity: LLMs can recombine the products of human creativity in amazing ways. But unlike humans, they cannot use universal concepts to find a possible item that is not derived from items already instantiated in the world.