Key points are not available for this paper at this time.
Abstract Launched in late November 2022, ChatGPT, a large language model chatbot, has garnered considerable attention. However, ongoing questions remain regarding its capabilities. In this study, ChatGPT was used to complete national high school exams in the Netherlands on the topic of English reading comprehension. In late December 2022, we submitted the exam questions through the ChatGPT web interface (GPT-3.5). According to official norms, ChatGPT achieved a mean grade of 7.3 on the Dutch scale of 1 to 10—comparable to the mean grade of all students who took the exam in the Netherlands, 6.99. However, ChatGPT occasionally required re-prompting to arrive at an explicit answer; without these nudges, the overall grade was 6.5. In March 2023, API access was made available, and a new version of ChatGPT, GPT-4, was released. We submitted the same exams to the API, and GPT-4 achieved a score of 8.3 without a need for re-prompting. Additionally, employing a bootstrapping method that incorporated randomness through ChatGPT’s ‘temperature’ parameter proved effective in self-identifying potentially incorrect answers. Finally, a re-assessment conducted with the GPT-4 model updated as of June 2023 showed no substantial change in the overall score. The present findings highlight significant opportunities but also raise concerns about the impact of ChatGPT and similar large language models on educational assessment.
Joost de Winter (Wed,) studied this question.