Failing grades climb in Berkeley computer science classes as AI usage rises

Computer science professors at the University of California, Berkeley, are reporting a meaningful shift in the upper end of student grade distribution over the past academic year. The investigation reported by The Daily Californian, highlighted by Hacker News, examines the relationship between the decline in the proportion of students passing the course and the rise in the use of artificial intelligence tools.

According to the data conveyed by Daily Cal, the foundational class CS 61A (Structure and Interpretation of Computer Programs) reached a failing rate of 12 percent in the past academic year. Three years ago the failing rate was around 4 percent. CS 70 (Discrete Mathematics and Probability Theory) recorded an 18 percent failing rate; three years earlier the figure was 7 percent.

This rise has triggered an extensive debate within the academic boards. Professor John DeNero, who has led the course for 30 years, told Daily Cal that 'we can see that students complete their in-class assignments much faster, but their performance on conceptual questions on the final exam drops noticeably'. DeNero added that AI tools can bypass deep comprehension in homework.

The chair of Berkeley's computer science department, Professor Pieter Abbeel, took a careful tone on the causes of the shift. Abbeel said, 'attributing the change to AI usage alone is insufficient; accumulated mathematics-foundation gaps after the pandemic period, curriculum changes at the secondary school stage, and the growth in class sizes also play a role'.

Looked at from the student side, there are dimensions that shed light on lived practice. Maya Chen, a CS 61A student and a representative of the Berkeley Engineering Society, said 'most of my classmates have ChatGPT or Claude write whole assignments for them, and then they have not learned to solve the kind of questions they will face alone in the final exam'.

The way professors are responding varies course by course. Some have eliminated traditional homework in favour of in-class proctored exams and oral assessments. Others have integrated AI use into the curriculum as a tool allowed under limited conditions. A third group has reduced the weight of homework in assessment through a hybrid approach.

The discussion on Hacker News covered a broad spectrum. One group of commenters argued that AI tools require the teaching of mathematics to be redesigned; another said that the classical mathematics curriculum was already regarded as unnecessary for practical computer science programmes. The thread drew more than 1,200 comments and became one of the most followed topics on the site.

The Daily Cal report notes that a similar trend is unfolding across the United States. Parallel declines were recorded in course pass rates in computer science departments at Stanford, MIT and Carnegie Mellon University over the past academic year. Carnegie Mellon's undergraduate computer science dean, Professor Mark Stehlik, told Daily Cal that 'a sector-wide debate on the redesign of teaching and assessment has become necessary'.

Academic studies in the field of educational science have begun to examine the impact of AI use on the learning process on more comprehensive scientific ground. Stanford Education Sciences professor Dr Carrie James recorded in a recent publication that 'AI use reduces the individual cognitive load but does not raise conceptual depth at the same rate'. The study was published in the journal Educational Researcher.

The Daily Californian's reporting underscores that the Berkeley case is only a small slice of a much larger debate. How universities respond to AI use may shape teaching methodology over the next decade. Berkeley's plan for the 2026-2027 academic year includes the reintroduction of proctored exams in at least four computer science courses and the addition of a new mandatory introductory course called 'AI Literacy' to the curriculum.