Formalising Social Signaling in AI World Models: Structural Limits and Architectural Extensions

Abstract Contemporary AI systems increasingly emulate social behaviour, yet their underlying architectures remain primarily epistemic: they optimise for accurate modelling of environmental regularities without incorporating persistent internal goal dynamics that shape agent identity over time. This paper develops a formal account of social signalling within world-model based architectures and identifies structural conditions under which socially grounded meaning can emerge. We show that purely predictive objective functions are insufficient to preserve distinctions that are irrelevant to environmental prediction but essential for social coordination, a limitation called predictive collapse. We then define an architectural extension in which system-relative goal dynamics are coupled to latent state evolution, enabling the stabilisation of socially meaningful representations. A proof-of-concept implementation in a social gridworld illustrates how coupling goals to internal dynamics induces representational differentiation in otherwise prediction-equivalent states. The proposed framework clarifies the computational requirements for integrating social signalling into world models and provides architectural design principles for cognitively grounded artificial social agents.