Surya Tallavarjula graduated from UC Berkeley with a bachelor of arts with honors in computer science and physics. He is now an incoming Ph.D. student at Stanford University’s physics department.
This spring, in my final computer science course at UC Berkeley, I received a B-plus, removing me from summa cum laude standing at the last second.
That grade was calculated correctly; I am not writing to contest it. I am writing because the structure that produced the grade risks undermining educational fairness in the age of generative AI.
The course had no midterm, and homework counted for zero percent of the final grade. A single 35-page final exam accounted for 65% of students’ final grades. These percentages are converted to letter grades after assigned grade bins, which award grades based on predetermined point ranges.
According to Berkeleytime, the course GPA fell to approximately 2.53, down from 3.528 the previous fall, while the failing grade rate rose from under 3% to 17%. The UC Berkeley electrical engineering and computer sciences department’s published guidelines list a typical GPA of 3.0 to 3.5 for upper division courses, with roughly 5% of students receiving D’s and F’s combined. When a course moves this far outside established norms, students deserve more than simply the assurance that the grade calculations were technically correct.
The redesign’s logic is understandable given that generative AI has impacted the traditional homework-heavy model. When problem sets are brutally difficult and heavily weighted, students who don’t outsource their work to AI may be penalized relative to those who do.
Proctored exams seem to be a straightforward alternative: reduce homework weight, eliminate midterms and replace curves with fixed bins. Each choice sounds reasonable in isolation, but these alternatives mean that the students harmed by these changes are often not the ones who were cheating.
Many campus courses have replaced homework grades with in-class quizzes consisting of one or two problems drawn randomly from prior assignments. In theory, this is an elegant compromise; but in practice, it solves one problem while creating others.
When a small number of quizzes accounts for roughly 20% of a final grade, one bad quiz can move a student’s course grade significantly. Grading becomes nearly all-or-nothing. A notation slip or momentary misread can cost more than 2% of a final course grade.
More troubling is what material the format actually rewards. A quiz drawn from previous homework sounds like a test of understanding, but when the stakes are high and grading is unforgiving, the safest strategy is memorization. A student who commits the official solutions to memory can nearly guarantee a perfect score. A student who trusts their ability to rederive the answer may learn more, but they risk a costly error.
A format can be AI-resistant and still pedagogically distorted. Moreover, students determined to game the system often find a way. The students hurt most often are the conscientious ones — those who do the homework even when it carries no grade value, avoid AI shortcuts and treat assignments as practice — when they discover that months of honest work can be outweighed by one unforgiving afternoon.
The irony is that many courses adopting these changes tell students that homework is still the best way to learn while the grading structure sends the opposite message. Assessment design functions not just as a quantitative measure of learning, but it shapes how students study and what they believe learning truly is.
A 17% failure rate in a large EECS upper division course will impact learning trajectories in a way never seen before. Many more students may lose financial aid. Some may delay graduation. Many may be pushed into academic probation by a course design that was never publicly framed as a beta-test.
The answer to AI-assisted cheating is not to make courses so high-variance that honest students fail at abnormal rates. Major structural changes should be reviewed against departmental norms and grade distributions should be monitored when courses undergo substantial redesigns. Assessments should accumulate evidence of learning across multiple assignments rather than concentrate on a single moment.
Surya Tallavarjula graduated from UC Berkeley with a bachelor of arts with honors in computer science and physics. He is now an incoming Ph.D. student at Stanford University’s physics department. Contact him at surya_talla@berkeley.edu or the opinion desk at opinion@dailycal.org.