← Back to feed

UC Berkeley’s EECS department is publishing numbers that the rest of the industry should read carefully. In spring 2026, the CS 10 failure rate reached 35.3%, up from under 10% in prior years. CS 61A hit 10.6%. EECS 127 reached 16.8%. All three cleared the department’s own guideline of 7% for combined D’s and F’s in lower-division courses.

The proximate cause, per Professor Dan Garcia: a “vast increase in academic dishonesty” driven by large language models. Nearly 30 CS 10 students were caught cheating on take-home exams in a single semester.

Dual Failure Mode

What makes the Berkeley data particularly significant is that AI cheating is not the only problem. Associate teaching professor Gireeja Ranade identified a parallel one: incoming students are arriving with foundational gaps in linear algebra, vector calculus, and proofs. In a documented diagnostic administered to Calculus I students between 2021 and 2023, students who passed zero to two of eight mathematical proficiency topics had a 46% failure rate in Calculus I by fall 2023.

The two failure modes compound. AI tools let students pass courses they haven’t learned; the underlying preparation gap makes it harder to catch up when AI is unavailable or when tasks get complex enough that the model hallucination rate matters. The cohort that graduates with a Berkeley CS degree in 2027 or 2028 will include students who passed intro CS via AI assistance but lack the vector calculus that upper-division ML and systems courses assume.

The Faculty Response

More than 1,100 STEM faculty across the UC system signed an open letter to the Board of Regents calling for standardized testing reinstatement by 2027. Signatories include Nobel laureate Jennifer Doudna and Fields medalist Richard Borcherds, both at Berkeley.

The letter cites a UC San Diego Senate-Administration Workgroup report from November 2025: between 2020 and 2025, students entering UCSD whose math skills fell below high school level increased nearly 30-fold since the system suspended standardized testing in 2020. Of those students, 70% tested below middle school level.

The petition’s central argument is counterintuitive in the equity framing: “The SAT mathematics requirement is not an obstacle to equity; rather, it is a prerequisite for it. Failing to measure preparation gaps does not remove barriers; it moves them into the classroom, where they become harder to overcome.” Professors Jitomirskaya and Stankova argue that removing objective measures pushed admissions toward expensive extracurriculars and GPA from elite private schools — effectively making the system less equitable while appearing more so.

What This Means for AI Hiring

The Berkeley data has a direct implication for the AI industry’s talent pipeline. CS graduates from 2027–2029 cohorts will have spent their university years at the intersection of peak AI adoption and degraded foundational preparation. A candidate with a strong Berkeley CS GPA may have passed courses with AI assistance and lack the linear algebra required for non-trivial ML work.

The failure rate numbers are a leading indicator. They’re visible now because Berkeley is publishing them and professors are named on the record. What’s less visible is the pass rate for students who completed courses via AI without being caught — and what those students will do when they arrive in production engineering environments.

The industry has spent three years asking whether AI makes developers more productive. Berkeley’s spring 2026 numbers raise a different question: whether AI is producing credential holders who aren’t actually developers at all.