How Does ChatGPT Measure City Similarities?

This paper quantifies GPT’s perceptions of cities by converting the descriptive texts, which are responses from GPT using prompts of different word lengths, into numeric embeddings using the BERT model. Our findings reveal that as the prompted length of the response increases, the stability and reliability of GPT’s textual descriptions of cities increases. Additionally, when comparing our results with previous studies that utilised Street View images to quantify city similarity, we find that the GPT model effectively links cities with similar characteristics, such as geographical proximity and cultural similarity.