What happens when students use AI to complete work they don't yet understand? UC Berkeley's CS department is gathering data on that question - and the results aren't encouraging.
According to a report in the Daily Californian, professors in Berkeley's computer science program are seeing failure rates climb as AI tool usage rises and students show weaker foundational math skills. The pattern is predictable in retrospect: AI handles the hard parts, students pass the intermediate checkpoints, and the gaps compound until they hit an exam or upper-division course that demands real fluency.
The Skill Gap Behind the Grade Data
Berkeley runs one of the most selective CS programs in the country. If this pattern shows up there, it's almost certainly more common at institutions with lower admissions thresholds.
The issue isn't that AI tools exist in classrooms. It's that students are using them to skip foundational work that was never optional. CS fundamentals - discrete math, probability, linear algebra - are deliberately difficult. Working through proofs and problem sets builds the mental models needed to debug complex systems, reason about edge cases, and understand why code fails in unexpected ways. When AI generates the answers, students bypass the process that was the actual point. They learn to produce correct-looking output without building the reasoning that makes output trustworthy.
Failing grades are the cleaner signal. The subtler problem is students who survive by leaning on AI, graduate, and surface the gap in a job - usually when something breaks in production and no one on the team can reason about why.
What Actually Needs to Change
Banning AI tools from CS courses isn't realistic or useful. Tools like ChatGPT and Claudee Code](/tools/claude-code/) are standard in professional software development. Teaching students to avoid them is preparation for a world that doesn't exist.
What would help: oral exams, whiteboard problem sessions, and explicit instruction around when to use AI versus when to work through something yourself. These are harder to administer than timed written tests and harder to grade at scale - which is exactly why most curricula haven't adopted them. Berkeley's grade data is evidence that existing assessments aren't measuring what they used to measure. Redesigning them for an AI-assisted world is unglamorous, resource-intensive work, but it's the actual problem.