Expanding the role of large language models in biology assessment
Abstract
T he rapid development of large language models, or LLMs, has created new opportunities to reconsider how instructors approach time-intensive tasks, such as grading and feedback. Naidu et al. provide a timely examination of this application in undergraduate biology education through their evaluation of LLM-based grading using authentic student work (1). The study moves discussion of artificial intelligence in assessment beyond speculation and toward practical evaluation. Rather than suggest ing that LLMs should replace instructor judgment, the authors provide a foundation for examining how these technologies may complement human assessment. Their approach is also notable for its low financial and technological barrier to entry. Naidu et al. evaluated ChatGPT GPT 4o, Claude 3.7 Sonnet, and Gemini 2.0 Flash using 200 essays from an introductory biology course (1). The models assessed student work with a standardized rubric under zero shot and few shot prompting conditions, and their scores were compared with those of a trained human grader. Agreement varied by model, prompting method, and rubric component. Performance was strongest for writing-related criteria and more variable for scientific content, reflection, ethi cal reasoning, and policy development. The models also differed from one another, demonstrating that conclusions about AI grading cannot necessarily be generalized from a single LLM (1). Several features strengthen the study. The use of authentic student writing gives the investigation greater educational relevance than studies based on artificial or highly constrained responses. Evaluating several contemporary LLMs is also important because models may interpret the same assessment task differently. Saez et al. similarly found meaningful variation when comparing six LLMs used for grading in higher education (2). Together, these findings suggest that AI-based assessment should consider not only whether LLMs can grade student work, but also which models, prompts, rubrics, and assessment types produce reliable results.