The rapid development of artificial intelligence raises a fundamental question: can properties associated with sentience be described mathematically? Although subjective experience cannot be measured directly, theories of consciousness identify candidate prerequisites including recurrent processing, information integration, temporal continuity, self-modelling and adaptive interaction with an environment. This project will investigate whether these properties produce a distinctive dynamical regime in a simple artificial agent.
The project is motivated by the hypothesis that neither low entropy, high complexity nor criticality alone provides a meaningful marker of sentience. A frozen system has low entropy, while random noise has high entropy, yet neither displays intelligence or agency. A more relevant regime may lie between rigid order and incoherent chaos, where a system can maintain differentiated internal states, integrate information, predict its environment and preserve a persistent model of itself.
The student will construct a recurrent artificial agent coupled to a dynamical environment. The model will include an internal state, a representation of the external world, a self-model that predicts aspects of the agent’s future behaviour, and an action mechanism through which the agent affects subsequent observations. Parameter scans will be used to identify transitions between ordered, metastable and chaotic behaviour.
The dynamics will be characterised using measures from information theory and nonlinear dynamics, including entropy rate, predictive information, mutual information, Lyapunov exponents, perturbational complexity and measures of integration. The project will test whether these quantities identify a reproducible region in which predictive structure, self-modelling, integration and adaptive autonomy occur together.
The full agent will be compared with systems lacking self-modelling, recurrence, action or structured connectivity. These controls will determine whether any proposed indicator measures more than ordinary task performance or stability. The project will not attempt to prove that an artificial system is sentient. Instead, it will ask whether candidate prerequisites for sentience can be defined mathematically and distinguished from ordinary complex computation.