UC Berkeley Computer Science Failure Rates Spike Amid AI Concerns
Failure rates in UC Berkeley's core computer science courses jumped sharply in spring 2026, with 35.3% of students in CS 10 failing the course and 10.6% failing CS 61A. Both figures represent a dramatic shift from spring 2024 and 2025, when failure rates stayed below 10%. The surge has prompted faculty to examine how generative AI and LLM tools are reshaping student learning and academic integrity.
Average grades in both courses landed at a C+ (2.3 GPA), below the EECS department's expected range of 2.8 to 3.3. The department typically anticipates around 7% D and F grades in lower-division courses.
Cheating Cases and Hidden Learning Gaps
Teaching professor Dan Garcia, who taught both courses, said a "primary driver" of the failures was a "vast increase in academic dishonesty" tied to tools like ChatGPT, Claude, and Google Gemini. Nearly 30 students in CS 10 were caught cheating on take-home exams, with additional cases referred to the Center for Student Conduct.
Beyond outright misconduct, Garcia identified a subtler problem: students complete assignments using AI tools without understanding the underlying concepts. "Students are leaning a little too hard on LLMs to do their work for them, and then at exam time just really aren't ready," he said.
This gap between assignment completion and actual competency becomes visible during in-person exams, when students cannot rely on AI assistance.
Mathematical Foundations Remain Weak
Faculty also point to uneven preparation in mathematics. Associate teaching professor Gireeja Ranade said students in EECS 127 (Optimization Models in Engineering) struggled with foundational linear algebra, vector calculus, and mathematical proofs. The course recorded a 16.8% failure rate, well above the department's typical 5% benchmark for upper-division classes.
Ranade learned during office hours that at least one student had taken a linear algebra course that permitted open-internet and open-AI use for assignments and exams. Those students arrived in advanced courses without the fluency they needed.
Staffing Cuts Reduce Hands-On Learning
Structural constraints have made instruction harder. Ranade said staffing shortages forced removal of a major project component from EECS 127, a segment that provided guided, hands-on learning with teaching assistant support.
EECS Department Chair Jelani Nelson said the university reduced both undergraduate computer science enrollment and the number of undergraduate teaching assistants due to rising TA wage costs.
Students Avoid Office Hours
Faculty report a striking change in how students seek help. Ranade said office hours, once "overflowing" with students, now see "very low engagement" despite repeated encouragement to attend.
Garcia described office hours that were sometimes entirely empty. "I used to have full office hours, and for the first time, I was having nobody come to my office hours," he said. "It was just so surprising to sit in my office alone."
The shift raises a question: Are AI tools quietly replacing traditional forms of learning interaction?
Rethinking How to Teach
Garcia plans to explicitly address the spring 2026 outcomes with future students and explore ways to identify those needing foundational support.
Ranade argues the solution is not to simplify instruction but to deepen it. Students must develop analytical and critical thinking skills even as AI tools become ubiquitous. "We really need to make sure that we are preparing our students to be solid, contributing citizens and leaders," she said.
Garcia reflected on what he called the essential difficulty of learning itself. "Confusion is the sweat of learning," he said. "A lot of students, I think, are not putting in the sweat."
A Broader Signal for Higher Education
What unfolds at Berkeley may signal a larger shift across higher education. AI for Education is no longer a peripheral tool-it is embedded in how students complete assignments, approach problem-solving, and think about learning itself.
The challenge facing universities is not simply how to regulate AI, but how to preserve intellectual struggle in an environment where answers are instantly available.
The data at Berkeley suggests that when struggle is reduced or outsourced, performance gaps become sharply visible. Whether that represents a temporary adjustment or a longer-term structural shift remains unclear. What is certain is that the traditional link between assignment completion and learning outcomes is under pressure, and institutions are only beginning to understand the consequences.
Your membership also unlocks:
