Although multi-level modeling has long been argued to showcase benefits in various application domains, its adoption is still hindered not least because the construction of multi-level models may entail a cumbersome and error-prone re-engineering effort. While many studies have investigated the potential of using LLMs to support the automatic construction of two-level conceptual models, such as UML class diagrams, no research has yet been conducted on using LLMs to support the construction of multi-level conceptual models. In this paper, we report on experiments conducted with ChatGPT to support the re-engineering of flat two-level models into deep multi-level models – a process we refer to as model deepening – using the multi-level modeling language FMML \(^{\textrm{x}}\) . Our findings indicate that while ChatGPT can significantly aid in semantic tasks during model deepening – such as comparing attribute meanings or analyzing type-object patterns – it also presents challenges, sometimes generating erroneous models by removing and duplicating properties. Future research should aim to develop an overarching model-deepening method that integrates probabilistic information sources, such as ChatGPT, with rule-based algorithms, while clearly defining and leveraging the user’s role in guiding and validating the process.