Berkeley CS failure rates triple as students lean on LLMs and skip the math

UC Berkeley’s spring 2026 computer science courses posted failure rates far outside department norms: 35.3% of CS 10 students and 10.6% of CS 61A students received F’s, against guidelines that target 7% D’s and F’s combined. Instructor Dan Garcia attributes the spike primarily to academic dishonesty involving Claude, ChatGPT, and Gemini, with nearly 30 CS 10 students caught cheating on take-home exams. Others weren’t formally disciplined but leaned heavily enough on LLMs that they couldn’t perform on in-person assessments.

The problem extends past introductory courses. EECS 127, an upper-division optimization class taught by Gireeja Ranade, saw a 16.8% F rate, with students arriving unprepared in linear algebra, vector calculus, and proofs. Ranade learned that at least one prerequisite linear algebra course at Berkeley permitted open-AI use on homework and exams. Both professors have signed a 1,300-faculty petition to restore SAT/ACT requirements for STEM admissions. Staffing cuts driven by EECS TA wage costs forced Ranade to drop the course’s final project, where students typically scored well.

Office hours that once overflowed are now empty, suggesting students are routing around the human help that used to backstop hard material. Garcia opposes curving and Harvard-style A caps, arguing for fixed thresholds and unlimited A’s tied to clear standards — an approach that exposes, rather than hides, when cohorts are underperforming. Both instructors plan to keep teaching harder material, not less, and want students to rebuild tolerance for difficulty rather than offloading it.