Scientific research with AI co-participation in chemistry: Prompt engineering and AI ethical challenges in the scientific discovery process

The role AI is playing in scientific research has been more of a co-creator than just a tool in the chemistry domain. In the scientific discovery, AI co-participation can present benefits (e.g., facilitate data-driven insights, accelerate innovation, etc.), but it also introduces ethical challenges (e.g., biases in AI-generated outputs, the opacity of AI decision-making, etc.) that can affect the reliability of research findings and communication. Large Language Model (LLM) based chat interaction, whilst aiding the exploratory tasks, adds another layer of challenges for domain experts. They need to know how to ask their questions in the best way to get a good outcome. This is called prompt engineering, which is the practice of designing and refining inputs to guide AI models toward generating accurate, relevant, and ethical outputs. Prompt engineering significantly shapes AI-generated content, which can have consequences both for accuracy and ethical integrity of scientific discovery process. However, domain experts need to always be aware of prompting guidelines while interacting with LLM-based chats. This offers a cognitive and practical burden for domain experts to use this kind of AI technology. We believe that LLM-based chats are powerful tools for the chemistry domain, but not all human-AI co-creation tasks should impose on domain experts the prompt engineering overload. In this work, we present an investigation into how chemistry experts interact with an LLM-based chat for tasks in two chemistry cases. Our goal was to understand how the experts place and structure their questions, what kind of questions they ask, their output expectations, plus doubts and problems along this process. Our findings guided the features’ design for chemistry experts combining LLM-based chat and other resources to tackle ethical challenges in scientific discovery. This investigation relates to the Discovery Workbench, a framework for human-AI co-creation presented in previous ACS meetings.