AI resistance to quizzes solved by ChatGPT
We gave ChatGPT-4o 10-question multiple choice quizzes to solve — one created manually, the others by generated by the quiz apps QuizGecko, Arlinear, and QuestGen.ai (v1). When we gave ChatGPT a version of the quizzes where we changed each question's correct answer to "None of these answers," resistance to AI improved (v2). Since each question has one correct answer out of four answer options, the chance of getting a wrong answer through a random blind guess is 75%. The human-produced quiz outperformed the AI-generated ones even when the correct answer on each quiz question was changed to "None of these answers" (v2).
Column bar chart showing how changing the format of a quiz improves its resistance to AI. The first quizzes (v1) were all created by a person and generated by quiz apps QuizGecko, Arlinear and QuestGen.ai. The quizzes were modified and tested again (v2) with each question's correct answer changed to "None of these answers." The second version of the quiz, especially the manual quiz edited by a person, improved AI resistance.
