Working with Large Language Models: impacts on performance, efficiency and perceived responsibility

Large language models (LLMs) and assistive tools that capitalise on LLMs are becoming an integral part of work, yet their impact on experiences during daily work tasks remains underexplored. Across two studies, we examine how LLM use affects task performance and work-related experiences such as responsibility and information processing. In Study 1, professional software developers (N = 40) completed a programming task using ChatGPT-4.0 while receiving prompt engineering advice or no advice. The prompt engineering group performed worse. Regardless of the condition, ChatGPT was associated with lower information processing, control, responsibility, and enjoyment. Study 2 replicated and extended these findings in a longitudinal design. Over 13 weeks, psychology graduate students (N = 18) completed writing and data analysis tasks with and without ChatGPT while receiving training on AI use. ChatGPT reduced time on task and information processing, but ChatGPT use and training only improved writing and not data analysis performance. As in Study 1, perceptions of control, responsibility, and information processing declined with ChatGPT use. While LLMs may boost efficiency, they thus appear to reduce users’ cognitive engagement and sense of task ownership – undermining core aspects of meaningful work experiences.