To identify common themes in the corpus, we undertook two independent coding processes.

 i.      Topic coding from abstracts

One of us undertook a simple coding process by reading paper abstracts or introductions and recording regularly appearing topics, informed by knowledge of key topics in the literature we reviewed in 2004, and alert to novel topics. This generated a set of 48 topical codes, which have been implemented as a filter in our online collection.

ii.      Thematic analysis with AI support

In parallel, one of us adapted the method described by Kriukow (2024) and made use of ChatGPT 4.5[1] to assist in a process of supervised thematic coding. Following this method, a round of initial coding required ChatGPT to identify up to 20 initial codes for each work, and for each suggested code to give a direct illustrative quotation from the work in question. These initial(and their illustrative quotations) were compiled in a spreadsheet, and a fraction of outputs were manually spot-checked to check on possible hallucinations (none were identified; had any been identified, human judgment would always have taken precedence). In a second round, ChatGPT assisted with development of focussed codes by sorting and combining clusters of related initial codes. Some focussed codes were also decided on manually. Finally, six major themes were identified entirely manually using sense-making processes based on knowledge of the field, with each theme grouping together sets of focussed codes. Side-by-side comparison of  topic codes (described above) with the focussed codes and themes generated in this AI-assisted process gave us confidence.

[1] OpenAI. (2025). ChatGPT 4.5 (April 2 version) [Large language model]. https://chat.openai.com/