(174a) Does "Better" Training Data Improve Chatgpt in Process Safety Applications? Beta-Testing Results of a Customized Gpt

At the 2024 Global Congress on Process Safety, an interesting and informative session focused on the potential for generative artificial intelligence tools such as chatbots to enhance process safety in industry. The presentations focused primarily on using available tools such as Microsoft’s CoPilot, Google’s Bard (now referred to as Gemini), and OpenAI’s ChatGPT. While the performance of these tools has shown promise, significant limitations have also emerged. A central question raised during discussion is whether chatbot performance could be enhanced by training these models using authoritative process safety texts (i.e., textbooks, journal articles) that may not be part of the training data for the technologies due to copyright limitations.

OpenAI’s platform includes an interface that paid subscribers can use to create custom GPT models. Within this interface, a user can upload files that the custom GPT can use in its training. Through this interface, a custom GPT was developed and provided with a collection of process safety publications. This custom GPT was then beta-tested, with its responses compared against those of the free-version of OpenAI’s ChatGPT.

This manuscript and presentation will review the literature and motivations for customized GPT tools, provide a practical guide to building customized GPTs within OpenAI’s interface, and share insight from initial testing of a customized GPT that is trained on safety specific publications. The work will probe the question of whether a dedicated collection of safety texts can improve the outputs from ChatGPT on safety related prompts. Attendees will gain a deeper understanding of the challenges, practical benefits, and future potential of customized AI models in advancing safety practices across various industries.