A sharp rise in failing grades at UC Berkeley’s computer science department is turning classroom AI use from a policy debate into a performance problem.
UC Berkeley has become an uncomfortable test case for what happens when students can outsource the work of learning faster than universities can redesign the way they teach, grade and police academic honesty.
In several spring 2026 computer science classes, the number of students receiving F grades climbed far above recent norms. The most striking example is CS 10, The Beauty and Joy of Computing, where 35.3% of students received F's, according to Berkeleytime data cited by The Daily Californian. CS 61A, a foundational programming course, had a 10.6% F rate. In spring 2024 and spring 2025, neither class had an F rate above 10%.
That is not a small grading fluctuation. It is a warning light. UC Berkeley’s electrical engineering and computer sciences department guidelines say lower division courses such as CS 10 and CS 61A should have about 7% of students receiving D's and F's. CS 10 moved far beyond that line, and it did so in a course designed partly to introduce students to computing, not to filter out only the most mathematically prepared majors.
The simple explanation is that students used ChatGPT, Claude, Gemini and other tools to complete assignments they did not understand. There is truth in that, but it is not the whole story. Professors also pointed to weaker math preparation and understaffing, both of which matter in large technical classes where small gaps can become failures by the end of the semester.
Professor Dan Garcia, who teaches CS 10, said some failing grades came from academic misconduct cases connected to cheating. Reports around the class also said nearly 30 students were caught cheating on take-home exams in spring 2026. That detail matters because it separates two problems that are often blurred together. One problem is dishonest use of AI. The other is dependency, where students may use AI as a shortcut so often that they never build the mental muscles the exam later tests.
The second problem may be harder for universities to handle. Cheating can be investigated, documented and referred to conduct offices. Dependency is more subtle. A student can ask an AI tool to explain code, generate a solution, fix a bug, summarize a concept and prepare a study guide without ever crossing a clearly stated policy line. The work gets done. The learning may not.
That is why this story reaches beyond Berkeley. Coding classes have always had a tension between collaboration and copying. Students learn by seeing examples, discussing bugs and adapting patterns. Generative AI intensifies that old tension because the tutor, collaborator and answer key now live in the same box.
Universities Are Rewriting The Deal
Berkeley is not treating AI as a minor syllabus issue. UC Berkeley Law has already adopted a stricter default policy for summer 2026 that bars students from using AI for academic work submitted for credit unless an instructor permits it. That is a very different posture from the early ChatGPT period, when many schools hoped broad guidance and disclosure rules would be enough.
At the same time, a large study led by Igor Chirikov at UC Berkeley’s Center for Studies in Higher Education found that students who use generative AI more often are more likely to report cheating. The study covered roughly 95,000 undergraduates and was published in Science in May 2026. It does not mean every frequent AI user is cheating. It does suggest that access to these tools changes student behavior in ways universities cannot ignore.
For computer science, the irony is sharp. These are the students most likely to enter a labor market reshaped by AI coding assistants. They should know how to use the tools well. But they also need to understand algorithms, abstraction, debugging and mathematical reasoning without leaning on a model that can produce plausible work faster than a beginner can judge whether it is right.
This is where universities face a practical choice. They can make more work happen in controlled environments, with in-person exams, oral checks, live coding, project defenses and smaller assessments earlier in the term. Or they can keep grading take-home work as if the old internet still exists. The first path is harder and more expensive. The second path is increasingly fictional.
There is also a market implication here. AI companies sell productivity, but schools are now seeing the cost of productivity without skill formation. For employers, that should be just as important as it is for professors. A graduate who can prompt an assistant but cannot reason through the result is not suddenly more capable. They are more dependent.
The next phase of AI in education will not be decided by whether schools allow or ban the tools. It will be decided by whether they can measure real understanding in a world where answers are cheap. Berkeley’s spring grades show that the adjustment is already here, and it is arriving through transcripts before policy committees have finished their work.
Also read: Databricks is choosing patience over a 2026 IPO • AI data centers are making electricity the next industry bottleneck • AI data centers are turning electricity into the new chip constraint
