Author : Dirk H. R. Spennemann
The public release of ChatGPT in late 2022 has resulted in considerable publicity and has led to widespread discussion of the usefulness and capabilities of generative Artificial intelligence (Ai) language models. Its ability to extract and summarise data from textual sources and present them as human-like contextual responses makes it an eminently suitable tool to answer questions users might ask.
Expanding on a previous analysis of the capabilities of ChatGPT3.5, this paper tested what archaeological literature appears to have been included in the training phase of three recent generative Ai language models: ChatGPT4o, ScholarGPT, and DeepSeek R1. While ChatGPT3.5 offered seemingly pertinent references, a large percentage proved to be fictitious. While the more recent model ScholarGPT, which is purportedly tailored towards academic needs, performed much better, it still offered a high rate of fictitious references compared to the general models ChatGPT4o and DeepSeek.
Using ‘cloze’ analysis to make inferences on the sources ‘memorized’ by a generative Ai model, this paper was unable to prove that any of the four genAi models had perused the full texts of the genuine references. It can be shown that all references provided by ChatGPT and other OpenAi models, as well as DeepSeek, that were found to be genuine, have also been cited on Wikipedia pages.
This strongly indicates that the source base for at least some, if not most, of the data is found in those pages and thus represents, at best, third-hand source material. This has significant implications in relation to the quality of the data available to generative Ai models to shape their answers. The implications of this are discussed.