A third of the class just failed. Not because the material got harder. It's the same introductory CS course Berkeley has taught for years. What changed is that students stopped learning how to think without a machine holding their hand.
The Numbers
UC Berkeley's CS 10 ("The Beauty and Joy of Computing," an introductory course designed as a gentle on-ramp to computer science) posted a 35.3% failure rate in Spring 2026. That's more than triple the historical norm. In Spring 2024 and Spring 2025, the rate never exceeded 10%.
CS 61A, the department's flagship intro-to-programming sequence, hit 10.6% F's. EECS 127, an upper-division optimization course, reached 16.8%, more than triple the department's 5% target for upper-division courses.
Both CS 10 and CS 61A averaged a 2.3 GPA (C+). The department's target range is 2.8-3.3.
The EECS department's grading guidelines specify a combined D/F rate of 7% for lower-division courses. CS 10 alone blew past that by a factor of five on F's alone.
Why This Is Happening
Teaching Professor Dan Garcia, who taught both CS 10 and CS 61A, is blunt about the cause. "Some of the numbers that you saw from the number of students who receive failing grades were because we caught them cheating and prosecuted them," Garcia told the Daily Californian. "But in other cases, it's students who are leaning a little too hard on LLMs to do their work for them, and then at exam time just really aren't ready."
Nearly 30 students in CS 10 alone were caught cheating on take-home exams in a single semester. Professor Gireeja Ranade identified a second compounding factor: incoming students increasingly lack preparation in linear algebra, vector calculus, and proof-writing. One student reported that a previous linear algebra course had an "open-internet, open-AI" policy, suggesting the upstream problem feeds the downstream collapse.
The department also cut undergraduate TA positions due to budget-driven wage disputes, forcing the removal of final projects that had previously helped students solidify their understanding before exams.
This Is Not Just Berkeley
The same week these Berkeley numbers dropped, researcher Igor Chirikov at UC Berkeley's Center for Studies in Higher Education published the largest-ever survey of undergraduate AI use across 20 research universities. The numbers:
| Statistic | Value |
|---|---|
| Students who use GenAI | ~66% (two-thirds) |
| Students who use it monthly+ | ~40% |
| Daily users who also cheat with AI | 26% |
| Monthly users who also cheat with AI | 7% |
Chirikov's earlier work found that courses with heavy AI exposure saw A-grade shares rise by 13 percentage points after ChatGPT, about 30% relative to baseline. That's grade inflation on the front end, hiding the skill deficits that only show up when proctored exams reveal the truth.
Over 1,300 faculty members have now signed a petition calling for reinstatement of ACT/SAT scores in STEM admissions, a remarkable reversal of the test-optional trend.
What the Community Is Saying
The Berkeley subreddit is full of CS students processing what happened. Some threads question how AI cheating can even be detected in introductory Python courses. Others ask whether Berkeley's CS degree still carries the signal it once did.
A recurring theme: students who relied on AI for homework throughout the semester found themselves unable to write basic code under exam conditions. The gap between "passed the homework" and "passed the exam" became a chasm.
Dan Garcia put it more evocatively. "I love this phrase my colleague uses: 'Confusion is the sweat of learning.' A lot of students are not putting in the sweat."
So What
The Berkeley crisis is the first well-documented case of AI-assisted learning systematically inflating grades on the front end while cratering exam performance on the back end. The 35% failure rate isn't a teaching problem. It's an assessment problem amplified by a tool that lets students skip the hard parts.
The deeper question is whether this pattern generalizes. If Berkeley's CS program, the best public computer science department in the world, can't keep introductory students who use AI to skip the hard parts of learning, what chance do less selective programs have?
Universities that respond by banning AI outright (Berkeley Law is doing this) will face a different problem: graduating students who have no experience with the tools they'll need in industry. Universities that keep letting AI do the work will face a credential crisis.
Neither path is good. The third path, redesigning assessment around AI rather than fighting it, using proctored exams and oral evaluations, is expensive and faculty-intensive. But the cost of not doing it is a generation of CS graduates who can prompt their way through a take-home but freeze on a whiteboard.
Sources: