Bruce Mountain, Shruti Kant
This paper examines a retrieval-augmented generation (RAG) customisation of artificial intelligence platform ChatGPT, using papers written by economist Professor Stephen Littlechild from the 1960s to the present, to create “SCL AI Agent” (SCL). SCL seeks to replicate and apply the thinking of Professor Littlechild. Establishing the corpus of Professor Littlechild’s papers, uploading it to ChatGPT and then instructing ChatGPT on how to understand that information and apply it, revealed the need for experimentation and learning-by-doing. Careful configuration sought to reduce hallucination and ensure well-informed responses delivered in Professor Littlechild’s style. Assessment of SCL by regulatory professionals who have had long interaction with Professor Littlechild rated SCL highly, particularly in respect of “insight,”“completeness” and “accuracy.” These assessors were less convinced of SCL’s ability to replicate Professor Littlechild’s written style. However, if users provided SCL with context to their questions and information on the audience for its answers, SCL did deliver responses tailored to those audiences. SCL itself and uncustomised ChatGPT were asked to assess SCL’s answers to the assessors’ questions. They both agreed on SCL’s superiority relative to uncustomised ChatGPT. SCL demonstrated a sophisticated, abstract understanding of Professor Littlechild’s scholarship, although its ability to replicate his imagination is less clear and merits further research. Creating AI agents of other economists and setting them to critique each other’s work could facilitate the more rapid dissemination of insight and understanding. JEL Classifications: A11, Role of Economics; Role of Economists; Market for Economists; C45, Neural Networks and Related Topics; D83, Search; Learning; Information and Knowledge; Communication; Belief; Unawareness; I23, Higher Education; Research Institutions; O33, Technological Change: Choices and Consequences; Diffusion Processes