Exploring AI-resistant assessments: insights from Large Language Models creating and completing their own research assessments

Higher education institutions have been grappling with the impact of Large Language Models (LLMs) on teaching and learning. LLMs like ChatGPT, Claude, and DeepSeek, can produce high-quality text with a single prompt. This is particularly concerning in research modules, where students are expected to showcase their work in a text-based format. In this study, ChatGPT, Claude, and DeepSeek were prompted to develop three research assessments that an LLM should not be able to complete and were subsequently prompted with the instructions generated. The purpose was to explore the influence of assessment instructions and characteristics on the ability of these models to generate complete responses. The instructions and characteristics included written text, oral discussions, physical attendance, contextual access, realistic content, lecturer approval, and the inclusion of multiple components. The most influential characteristic was the inclusion of an oral component. Other noteworthy characteristics were requiring students to attend live events and including multiple staged components. These findings contribute to the ongoing exploration of assessment design in the age of AI and LLMs.