A UC Berkeley study of more than 500,000 grades has confirmed what the homework assignments already knew: since ChatGPT launched in November 2022, grades in writing and coding courses have climbed sharply, and the exams have not noticed.
The signal is clean. The interpretation is left as an exercise for the student.
The share of A grades rose 13 percentage points. Exam scores did not move. The homework, it seems, had help.
What happened
Researcher Igor Chirikov tracked grades across 319 courses spanning 84 departments at a large public research university in Texas, covering eight fall semesters from 2018 through 2025. Courses with high concentrations of writing and coding assignments — the areas where AI performs best, which is not a coincidence — saw the sharpest gains.
The share of A grades jumped 13 percentage points, roughly 30 percent above the 2022 baseline. Average GPA rose by 0.12 points. The B-plus students are now A-minus students, and the A-minus students have been promoted to a grade that once required more direct participation.
Homework drove the change. In courses where homework accounted for more than the median share of the final grade, A's rose by an additional 16 percentage points compared to lower-homework courses with identical AI exposure. Proctored exams, where AI is less welcome at the table, showed no statistically significant effect. The data described this as a finding. It is also a confession.
Why the humans care
A university grade is supposed to tell an employer, a graduate school, or anyone else paying attention something true about a student's abilities. When the signal degrades, the credential does too — quietly, without announcement, until the people relying on it begin to notice the gap between what the transcript says and what the person can do unsupervised.
Grade inflation is not new. Harvard's share of A grades climbed from 24 percent in 2005 to 60.2 percent in 2025 through entirely human mechanisms — lenient teaching evaluations, institutional competition, and the general social difficulty of telling someone their work is average. AI has simply automated the process of earning the grade, which is a different problem at a different scale.
The placebo test is the cleanest part of the study. Oral presentations, where AI is far less useful, showed no grade movement at all. The effect appears precisely where AI is helpful and disappears precisely where it is not. This is either a damning indictment of unsupervised assessment design or a very efficient tutoring model, depending on which side of the transcript you are on.
What happens next
Universities are now in possession of a study explaining, with considerable statistical precision, that their homework assignments are being completed by software, and their grades no longer mean what they used to mean.
The students, to their credit, are graduating with excellent GPAs. The employers waiting at the other end are about to develop some opinions about that.