ChatGPT and Bard Performance on the POSCOMP Exam
Resumo
Context: Modern chatbots, built upon advanced language models, have achieved remarkable proficiency in answering questions across diverse fields. Problem: Understanding the capabilities and limitations of these chatbots is a significant challenge, particularly as they are integrated into different information systems, including those in education. Solution: In this study, we conducted a quantitative assessment of the ability of two prominent chatbots, ChatGPT and Bard, to solve POSCOMP questions. IS Theory: The IS theory used in this work is Information processing theory. Method: We used a total of 271 questions from the last five POSCOMP exams that did not rely on graphic content as our materials. We presented these questions to the two chatbots in two formats: directly as they appeared in the exam and with additional context. In the latter case, the chatbots were informed that they were answering a multiple-choice question from a computing exam. Summary of Results: On average, chatbots outperformed human exam-takers by more than 20%. Interestingly, both chatbots performed better, in average, without additional context added to the prompt. They exhibited similar performance levels, with a slight advantage observed for ChatGPT. Contributions and Impact in the IS area: The primary contribution to the field involves the exploration of the capabilities and limitations of chatbots in addressing computing-related questions. This information is valuable for individuals developing Information Systems with the assistance of such chatbots or those relying on technologies built upon these capabilities.
