A recent Daily Cal article reports that UC Berkeley computer science professors are seeing a surge in failing grades, coinciding with increased AI usage and dwindling math skills among students. The article cites specific courses where failure rates have doubled or tripled since 2020. As an engineer who ships production code, I want to look past the headlines and ask: what is actually happening, and what can we learn from it?
First, the data. The article mentions CS 61A (intro to CS) and CS 70 (discrete math) as examples. In CS 70, the fail rate went from around 5% pre-pandemic to over 15% in some recent semesters. Professors surveyed students and found that a majority use AI tools like ChatGPT for homework and even exam prep. At the same time, performance on in-person, proctored exams (where AI is not available) has dropped, especially on questions that require symbolic manipulation or multi-step reasoning.
This is not a simple story of "AI makes students lazy." The reality is more nuanced. Many students report that they use AI to fill gaps in understanding when lectures move too fast or when they lack prerequisites. But the data suggests that using AI as a crutch instead of a tutor leads to weaker long-term retention. In CS 70, students who self-reported high AI usage scored 20% lower on the final exam compared to low-AI users, controlling for prior GPA.
From a practitioner's perspective, this mirrors what we see in industry. Junior engineers who rely heavily on Copilot or ChatGPT to write code often struggle to debug or reason about system behavior when the AI gives a wrong answer. They lack the mental models to verify outputs. The Berkeley data is an early warning that the same problem is happening in foundational courses.
What about the math skills specifically? The article notes that students are struggling with basic algebra and proof techniques. In CS 70, which covers combinatorics and probability, students used to be able to derive formulas from first principles. Now many try to prompt an AI for the answer without understanding the derivation. When the AI produces a plausible but incorrect result, they cannot catch the error. This is a direct threat to the kind of rigorous thinking needed for systems programming, cryptography, or machine learning.
Are there any upsides? Some professors argue that AI can be a great tool for exploring ideas or generating test cases, but only if the student already has a solid foundation. The problem is that many students skip the foundation-building phase. The Berkeley CS department is now experimenting with "AI-free" exam sections and requiring handwritten proofs in some courses. They are also updating curricula to teach students how to use AI effectively as a learning aid, not a shortcut.
What does this mean for hiring managers and engineering leads? If you are interviewing Berkeley CS grads in 2026, you may see a bimodal distribution: some with deep fundamentals who used AI as a tool, and others who coasted through and cannot reason about code. Expect to see more emphasis on whiteboard coding and system design interviews that test understanding, not just output. The trend also suggests that bootcamps and online courses that rely heavily on AI tutors may produce graduates with shallow knowledge.
For educators, the takeaway is clear: we need to redesign assessments to measure understanding in an AI-augmented world. Open-book exams with AI allowed but requiring explanation of reasoning might be one path. Another is to focus on projects that require integration of multiple concepts, where AI can help but cannot replace the synthesis.
Finally, a note on cost. The article does not discuss financials, but the trend has implications. Students paying $15k+ per year for a Berkeley CS degree may get less value if they cannot pass core courses. Universities may need to invest in AI literacy programs and tutoring. On the flip side, employers may need to invest more in onboarding and training for new grads.
In summary, the Berkeley data is a concrete signal that AI usage without foundational skills leads to worse outcomes. The solution is not to ban AI but to teach students how to use it responsibly while building their own mental models. For engineers, this reinforces the value of deep understanding over tool proficiency.
Keep Reading:
- Is AI Ruining Our Skills? Early Results Are In and They're Not Good
- My Agent Skill for Test-Driven Development
- Claude Code: Everything You Can Configure That the Docs Don't Tell You
Want to build AI agents that augment your team's skills without eroding them? Try Zlyqor at https://app.zlyqor.com/signup.
