Grade Increases After ChatGPT Adoption―Evidence Points to Task Substitution, Not Learning Improvement

Research from UC Berkeley has revealed that student grades have increased following the launch of ChatGPT. However, this increase does not reflect an improvement in learning ability; rather, it strongly suggests that AI is "completing students' assignments for them."

The study analyzed more than 500,000 grade records. The findings showed that grades increased notably at the time ChatGPT became publicly available, particularly in courses emphasizing writing and programming—fields where AI excels. These subjects directly overlap with tasks that AI performs well.

Most notably, grade improvements concentrated primarily on "homework" evaluations. The fact that grades improved only on home assignments, not on exams or in-class work, indicates that students are using AI as a "ghostwriting tool" for submissions rather than as a learning aid. If actual learning had deepened, we would expect to see similar changes in exam performance, but such evidence has not been found.

This phenomenon reflects the rapid proliferation of generative AI. ChatGPT was released in November 2022 and quickly spread globally among students. Universities have struggled to establish rules governing AI use at the pace of its adoption, and questions about how to maintain academic integrity have become a global concern.

The significance of this research extends beyond simply noting increased cheating. If grades no longer accurately reflect actual learning, the very foundation by which universities measure learning outcomes would be undermined. The implications extend to the broader evaluation systems used by employers and graduate programs to assess talent based on academic performance.

However, there remains room for debate about whether "supplementary use" of AI—such as using it to organize thoughts or learn sentence structure—is inherently harmful. The core issue lies in distinguishing whether AI functions as a tool that supports the learning process or one that entirely substitutes for it. Moving forward, educational institutions will need to fundamentally reconsider how they design assignments and assess student work.

Crucially, Berkeley's study is based on large-scale data exceeding 500,000 records and captures a structural trend rather than isolated incidents. As research that quantitatively demonstrates how generative AI affects education, it is likely to be referenced in future policy discussions and educational design initiatives.