Social sentience in neural language models

This work explore the ability of Neural Language Models (NLMs) to produce and modulate”autobiographical” stories,, thanks to their extensive exposure to social linguistic interactions, with a level of narrative coherence comparable to that of humans. Generative AI based on transformer architecture has demonstrated the ability to perform extraordinary tasks often considered exclusive to human cognitive abilities. The need to clarify the functioning of the algorithmic black box within transformers, combined with the opportunity to use cognitive science tasks and tests in this investigation, has led to a significant field of studies aiming to bridge this explanatory gap. The term”machine psychology” refers to the administration of cognitive tests, typical of human cognition, to NLMs. Contributing to this debate our proposal involves an empirical study on the modulation of autobiographical narrative coherence, an element widely used in cognitive psychology for studying aspects related to self-integrity and fragmentation, emotion modulation, worldview and self-construction. We subjected OpenAI models to tasks requiring story production following a multi-level pre-induction framework, considering three variables: age, mood, and gender. The results demonstrate that NLMs are not only capable of simulating various aspects of the human experience but can also adapt to the designated role and modulate their level of narrative coherence accordingly. This provides evidence of these artificial artifacts’ ability to produce cognitively complex textual elaborations and suggests that the emergence of narrative awareness within transformer architecture, akin to the prelude to consciousness in human, may be possible due to their overexposure to social linguistic interactions.