Can large language models (LLMs) be used for digital twin engineering (DTE)? Engineering digital twins (DTs) is a complex process consisting of several phases and involving different disciplines. We argue that an investigation of LLM use in DTE has to define what kinds of DTs are in focus and what DTE phases shall be supported. In our work, we concentrate on the early phases of DTE, with a particular focus on requirements engineering (RE), and we focus on supervisory and operational DTs. This paper investigates the quality of LLM output for defining requirements for DTs. The main contributions of our work are results from an experiment comparing requirements to a DT of an air conditioning facility of a domain expert and ChatGPT and conclusions for prompt engineering resulting from this experiment.