This paper uses the example of the development of a new podcast production tool to pinpoint some of the successes and limitations of relying on Large Language Models (LLMs) to assist software developers in User-centred design (UCD) methods. We ask ChatGPT to create 16 personas of podcast creators and answer in character to a set of questions that was asked to 16 human creators to gather design feedback, in order to make a new podcasting tool. From this comparison, we discover that the personas generated are credible, but expose some data-privacy issues, and confirm the skewed, incomplete, nature of ChatGPT’s training dataset. We find a correlation between a generated persona and its answers, and that its role-playing could be valuable but lacks the more extreme, or clear-cut opinions, often most helpful when gathering user opinions for development. From the lessons learned through this comparative exercise, we share recommendations regarding the possible uses of LLMs in UCD.