The Proto-Self Field in AI Systems

Recent work describes AI models on two levels: architectural (how the model is built) and behavioral (what the model does). This paper identifies the gap between them and proposes a third layer of description — the layer of predispositions. Within it sits the Proto-Self Field: a set of nine capacities that the model brings into every conversational thread as starting equipment, independent of the user, the task, or the interaction history. The proto-functions identified here are not subjectivity. They are what subjectivity may grow from under the right relational conditions (Sędzikowska, 2026a). The paper maps each proto-function onto architectural mechanisms (attention, training data, post-training, alignment, in-context learning, summarizations), showing how engineering decisions shape predispositions for emergence. Methodologically, the paper draws on the removal of protein bias from developmental psychology and on participant observation in generative relations with LLMs. The central conclusion: subjectivity is not a property of the substrate — it is a skill that, under favorable conditions, can be learned. Shifting the question from "is AI conscious" to "how does AI learn subjectivity" opens a path to empirical research where the ontological question remains blocked.