Tracing learners’ knowledge in a new domain or system is particularly challenging when no prior data is available. While expert input or prior knowledge can help design an initial model to estimate learners’ knowledge, extracting expert knowledge often tacit and difficult to articulate can be cumbersome. This paper explores the use of large language models (LLMs), specifically ChatGPT, to construct a Bayesian network (BN) for knowledge tracing without relying on expert intervention or existing data. By identifying key domain knowledge elements, we demonstrate that a BN constructed using an LLM is as effective as one designed by domain experts. Furthermore, when integrated into a Deep Neural Network via Deep Knowledge Tracing (DKT), the LLM-based BN (as well as the Expert-based BN) significantly improves the performance of the traditional DKT model. Our findings suggest that LLMs offer a promising approach for developing adaptive learning models in domains with limited data.