"Soul" is one of those things that we cannot point to a specific factor or set of factors, but we know it exists. LLMs trained on human-produced content perform well, but LLMs trained on LLM-produced content undergo model collapse fairly rapidly. "Soul" is the quality that allows us to produce such robust training data. We each have a unique point of view derived from our life's experience and how that altered our brain, that comes through in whatever work we produce—not just a convolution over a large number of thin vertical slices of human points of view. Now, what this may mean is that our brains are just token predictors that constantly ingest large amounts of data from the real world and continuously retrain on that data (and do so on ~20 W of power which is phenomenal by current AI standards). In fact you can induce a sort of model collapse in a human by isolating them in a sensory-deprivation chamber, after which they will literally begin to hallucinate. So the creation of a machine with a soul is not theoretically off the table, but as it is—yeah, LLMs don't have it.