Entre o quadro e o algoritmo: o que a IA pode (e não pode) julgar na análise real?

Abstract

This study investigated the potential and limitations of two Artificial Intelligence tools — the ChatGPT 4.0 language model and the Gradescope assessment platform — in grading examinations from a Real Analysis course. The research, characterized as qualitative, applied, and documentary, was conducted using handwritten exams from 14 undergraduate Mathematics students at UTFPR. The evaluations were analyzed in four stages: autonomous correction by ChatGPT, human grading by the instructor, criterion-based correction by ChatGPT using the professor’s rubric and assisted correction through Gradescope. The findings show that ChatGPT, despite its difficulties interpreting non-standard mathematical notation, produced results closer to human evaluation when guided by explicit criteria. However, in autonomous mode, the model tended to overestimate student performance, especially when superficial mathematical language was present. Gradescope, in turn, proved efficient in organizing and applying rubrics but was unable to automatically grade formal proofs and logical arguments, restricting its effectiveness to structured questions. The study concludes that although both tools offer valuable contributions to the assessment process, neither replaces human judgment in disciplines that require rigorous deductive reasoning. These technologies have potential as complementary instruments to streamline and standardize assessment, but the pedagogical legitimacy of grading remains dependent on the instructor’s expertise.