Study: AI Grading Is Too Unstable to Replace Teachers

A study by Cardiff and Melbourne universities found that generative AI is not as reliable as human teachers in grading written work. Researchers used two versions of ChatGPT to grade 50 biology essays on 7 criteria, with 4 prompt styles, and compared results to human grading.

The AI scores varied greatly and couldn't accurately predict human marks. Except for one case, AI gave higher average scores, with a max gap of 16.1 points, and up to 40 points for a single essay. AI tended to lower high scores and raise low ones, pulling marks toward the middle.

The study says current AI models aren't suitable to replace teachers for grading, and using AI without student consent raises ethics issues. While AI may improve, it's hard to match human grading now.