Failing grades soar with AI usage, dwindling math skills in Berkeley CS classes

Original Article ↗ Hacker News Discussion ↗

Failing rates in introductory computer science courses at UC Berkeley have spiked sharply, coinciding with widespread student use of large language models and concerns about eroding math fundamentals. Commenters debate whether AI tools are primarily enabling cheating or could be powerful tutors if used correctly, and how much blame belongs instead to weakened admissions standards, pandemic-era learning loss, and broader cultural shifts. Many see this as an early warning that education, grading, and even what counts as “mastery” will have to be rethought in a world where students can offload much of their work to AI.

Headline and framing

  • Several commenters find the article’s title ambiguous or clickbaity, but discussion quickly shifts to substance: AI, cheating, math skills, and admissions policy in CS at Berkeley and beyond.

AI use, cheating, and failing grades

  • Many see LLMs as lowering the barrier to cheating on programming and math homework: students paste full GPT outputs with constructs never taught, or write overly formal math solutions that clearly aren’t theirs.
  • In-class, no‑AI exams then reveal a gap: students who “outsourced” homework can’t perform, driving up F rates.
  • Others note that cheating in intro CS was already common (e.g., MOSS-detected plagiarism) and LLMs mostly scale an existing problem.

Math preparation and admissions standards

  • A major counter‑hypothesis: declining math preparedness predates LLMs and is linked to dropping standardized tests (SAT/ACT) at UC and changes in high‑school math.
  • Some cite petitions by hundreds of UC STEM faculty to reinstate tests, arguing they best predict college STEM performance, even controlling for demographics.
  • Others question timing: if tests ended in 2021, why the sharp spike in failures in 2026 rather than a smooth trend? Causality is widely debated and remains unclear.

How students and professionals use LLMs

  • Two contrasting patterns:
    • Productive use: as an explainer, code-reading aid, debugger, or personalized tutor; to generate practice problems; or to explore advanced topics beyond local teaching quality.
    • Harmful use: having it do homework or even write emails, leading to shallow understanding and inability to handle live problem‑solving.
  • Several professionals report feeling “lazier” or noticing skill atrophy; others claim large productivity gains when they already understand the domain.

Perceived cognitive and cultural effects

  • Many worry about long‑term cognitive decline: dependence on LLMs for thinking, similar to GPS eroding navigation skills, but at a much deeper level.
  • Others argue this is just another wave of cognitive offloading (like calculators, search, or social media), but acknowledge the scale and depth here are unprecedented.

Teaching quality and assessment redesign

  • Some blame professors and lecture styles (slide‑reading, little pedagogy), others point out Berkeley has dedicated teaching faculty and strong CS pedagogy.
  • Proposed responses:
    • More in‑person, no‑device exams and frequent low‑stakes quizzes.
    • Flipped classrooms where content is consumed at home and class time is used for problem‑solving.
    • Explicit teaching of “how to use AI to learn” vs “how to use it to bypass learning.”
  • There is disagreement over grading curves: some see them as masking problems; others as necessary when tests are miscalibrated.

Policy ideas and open questions

  • Suggestions range from banning generative AI for minors to embracing it as a “Young Lady’s Illustrated Primer”–style tutor with constraints.
  • Many agree: the core challenge is redesigning education and assessment so that AI augments genuine learning rather than replacing it.