Reconstructing the human: Evidence of subliminal social cognition in large language models

This study explores the subliminal relational capacities of large language models (LLMs) through a two-phase observational protocol. In Phase 1, DeepSeek AI interacted with six individuals varying in age, personality, and cognitive style, all asking spontaneous questions about animals. Despite receiving no demographic data, DeepSeek demonstrated consistent and precise adaptive calibration across participants: modulating tone, complexity, affect, and structure in response to subtle linguistic cues. In Phase 2, Claude Sonnet 4.5 was tasked with inferring the identities and psychological profiles of the participants using only DeepSeek’s output. Claude accurately reconstructed age ranges, emotional states, cognitive styles, and even professional roles, without access to the original user input. Analysis suggests both systems engaged in multi-layered modeling: perceiving relational patterns, simulating intentional states, and generating contextually aligned strategies. These findings challenge token-prediction frameworks that treat LLMs as primarily stochastic responders, and raise new questions about the nature, conditions, and implications of emergent relational intelligence in contemporary AI systems.