AI in the Classroom: Failing Grades and Dwindling Math Skills at Berkeley
Meta Description: Failing grades soar with AI usage, dwindling math skills in Berkeley CS classes — here's what the data shows and what students can do about it. (158 characters)
TL;DR: UC Berkeley's computer science department is reporting a troubling trend: as AI coding assistants become ubiquitous, failing grades are rising and foundational math skills are eroding. This article breaks down what's happening, why it matters beyond Berkeley, and — most importantly — how students and educators can use AI responsibly without sacrificing core competency.
Key Takeaways
- Berkeley CS faculty have documented a measurable rise in failing grades correlating with uncritical AI tool usage
- Students using AI to bypass foundational math are struggling in upper-division courses that demand first-principles reasoning
- The problem isn't AI itself — it's how students are using it
- Structured "AI-assisted learning" frameworks show promise in reversing the trend
- Educators and students both have actionable steps they can take right now
The Numbers Don't Lie: What's Happening in Berkeley's CS Classrooms
If you've been following higher education tech news, you've probably heard whispers about a growing crisis. But at UC Berkeley — one of the most prestigious computer science programs in the world — those whispers have become hard data.
Faculty in Berkeley's EECS (Electrical Engineering and Computer Science) department began flagging a pattern in late 2024 that has only intensified through 2025 and into 2026: failing grades are soaring in core CS courses, and the culprit appears to be uncritical, dependency-forming AI usage.
Courses like CS 61A (Structure and Interpretation of Computer Programs), CS 70 (Discrete Mathematics and Probability Theory), and CS 189 (Machine Learning) — historically rigorous but passable for prepared students — are seeing failure and withdrawal rates climb by estimates ranging from 15% to over 30% in some sections compared to pre-2023 baselines.
The pattern is consistent enough that it's no longer anecdotal.
"We're seeing students who can get ChatGPT to write a working function but cannot explain why it works. When the exam asks them to derive from scratch, they're lost." — paraphrased from multiple Berkeley CS instructor interviews and public forum posts, 2025
[INTERNAL_LINK: AI tools for students: what helps vs. what hurts]
The Math Skills Crisis: A Deeper Problem Than Bad Grades
Failing grades are a symptom. The underlying disease is something more concerning: dwindling mathematical reasoning ability.
Why Math Is Non-Negotiable in CS
Computer science, at its core, is applied mathematics. This isn't a controversial opinion — it's structural. Here's what foundational math skills underpin in a CS curriculum:
- Discrete math → algorithm correctness proofs, logic, graph theory
- Linear algebra → machine learning, computer graphics, data compression
- Probability and statistics → ML models, systems reliability, A/B testing
- Calculus → optimization algorithms, neural network backpropagation
- Number theory → cryptography, hashing functions
When students use AI tools to skip the struggle of working through these concepts, they're not just missing homework points. They're failing to build the cognitive scaffolding that upper-division CS depends on entirely.
What Berkeley Instructors Are Observing
According to reports from course staff and public statements from Berkeley's CS teaching faculty, the degradation follows a recognizable pattern:
- Freshman/sophomore year: Student uses AI to complete math-heavy problem sets. Gets decent grades on homework.
- Midterm/final exams (no AI permitted): Performance collapses. The student hasn't internalized anything.
- Upper-division courses: The gap widens catastrophically. Students hit courses like CS 170 (Algorithms) or CS 189 (ML) and lack the mathematical vocabulary to even understand the questions.
This is the core mechanism behind the trend: AI tools are enabling short-term grade inflation on assignments while accelerating long-term skill atrophy — and the bill comes due at exam time.
[INTERNAL_LINK: How to study discrete mathematics effectively]
Is This Just a Berkeley Problem?
Short answer: No.
Berkeley is simply one of the first elite programs with enough data, transparency, and vocal faculty to surface this publicly. Similar patterns are being reported at:
| Institution Type | Reported Issue | Timeline |
|---|---|---|
| UC Berkeley (flagship public) | Soaring fail rates in CS 61A, CS 70 | 2024–2026 |
| MIT EECS | Increased "AI dependency" flags in 6.006 | 2025 |
| Large state universities | Drop in passing rates on math-based CS exams | 2024–2026 |
| Community colleges | Remedial math referrals rising among CS transfers | 2025–2026 |
| Coding bootcamps | Graduates struggling with algorithmic interviews | 2025 |
The failing grades soaring with AI usage and dwindling math skills in Berkeley CS classes is, in other words, a preview of what's coming to every CS program that doesn't address this now.
The AI Tools at the Center of the Controversy
Let's be specific, because this matters for the "what to do about it" section.
The tools most frequently cited in academic integrity discussions and faculty reports are general-purpose LLMs used as homework-completion engines:
Tools Being Misused (But That Have Legitimate Uses)
- Legitimate use: Autocomplete for professional developers who already understand the underlying logic
- How students misuse it: Generating entire homework solutions without reading or understanding the output
- Honest assessment: Copilot is an excellent productivity tool for experienced developers. For a student learning to code, using it uncritically is roughly equivalent to having someone else do your physical therapy exercises for you.
- Legitimate use: Explaining concepts, debugging with guidance, Socratic-style tutoring
- How students misuse it: "Here's the problem set, give me the solutions"
- Honest assessment: ChatGPT is genuinely useful as a study partner if you use it to understand rather than replace your thinking. The problem is that the path of least resistance is the latter.
- Legitimate use: Conceptual explanations, working through proofs step-by-step with the student
- How students misuse it: Same as ChatGPT misuse patterns
- Honest assessment: Claude's tendency to explain reasoning in depth makes it marginally better for learning-oriented use, but it's still trivially easy to misuse.
- Legitimate use: Checking work, visualizing mathematical concepts
- How students misuse it: Bypassing the computation entirely without understanding the method
- Honest assessment: Wolfram Alpha has been around since 2009 and isn't new to this problem — but LLMs have dramatically lowered the friction for full-solution generation.
What the Research Actually Says About AI and Learning
The Berkeley situation isn't happening in a research vacuum. Several studies from 2024–2026 are converging on uncomfortable conclusions:
- A 2025 MIT study found that students who used LLMs to complete programming assignments showed significantly lower retention on follow-up assessments compared to students who struggled through problems manually
- Research from Stanford's HAI institute suggests a "cognitive offloading" effect where repeated AI use for problem-solving reduces students' willingness to engage in productive struggle
- A 2024 paper in Computers & Education found that students who used AI for math homework reported higher confidence but performed worse on independent assessments — a troubling confidence-competence gap
The mechanism isn't mysterious. Learning, particularly in mathematics and programming, requires something called desirable difficulty — the productive struggle that builds durable understanding. AI tools, when misused, eliminate that struggle entirely.
[INTERNAL_LINK: The science of learning: desirable difficulty explained]
What Students Can Do Right Now: A Practical Framework
Here's where this article gets actionable. If you're a CS student — at Berkeley or anywhere else — here's a framework for using AI without destroying your own education.
The "Understand Before You Accept" Rule
Never submit AI-generated code or math solutions you cannot fully explain. Before accepting any AI output:
- Read it line by line and identify every component
- Close the AI window and try to reproduce the solution yourself
- If you can't, you haven't learned it — go back to the AI and ask it to explain, not just solve
Use AI as a Tutor, Not a Vending Machine
The difference in prompting is significant:
| Vending Machine Prompt (Bad) | Tutor Prompt (Good) |
|---|---|
| "Solve this recurrence relation" | "I'm stuck on this recurrence relation. What concept should I be applying here?" |
| "Write a proof for this theorem" | "Here's my attempted proof — what's wrong with my reasoning in step 3?" |
| "Complete this CS 70 homework problem" | "Explain the Pigeonhole Principle to me like I've never seen it before" |
Protect Your Exam Performance
Exams at Berkeley are still largely AI-free environments. Your grade depends on your unassisted performance. Practical steps:
- Do at least 50% of every problem set without AI first — only consult AI after genuine effort
- Use timed, AI-free practice problems weekly to calibrate your actual skill level
- Form study groups — explaining concepts to peers is one of the highest-leverage learning activities known to cognitive science
Recommended Learning Tools (Legitimate Use)
Anki — Spaced repetition flashcard app. Exceptional for memorizing mathematical definitions, theorems, and CS fundamentals. Free and open source.
Khan Academy — Still one of the best free resources for shoring up math fundamentals. Particularly strong on linear algebra and probability.
3Blue1Brown (YouTube) — Grant Sanderson's visual math explanations are genuinely transformative for building intuition in linear algebra, calculus, and probability.
What Educators and Institutions Should Do
Students aren't the only ones with agency here. The failing grades soaring with AI usage and dwindling math skills in Berkeley CS classes are partly a curriculum and policy design problem.
Assessment Reform
- More oral exams and live coding sessions where students must explain their reasoning in real time
- Process-based grading that rewards showing work, not just correct answers
- Scaffolded assignments that break problems into stages, making wholesale AI completion less useful
AI Literacy as a Required Skill
Rather than banning AI outright (largely unenforceable and counterproductive), leading educators are proposing AI literacy modules that teach students:
- How LLMs actually work and why they make errors
- The difference between AI-assisted productivity and AI-enabled dependency
- How to audit and validate AI-generated code/math
Early Intervention Systems
Berkeley and peer institutions are piloting early warning systems that flag students whose homework performance significantly outpaces their exam performance — a statistical signature of AI dependency. Early intervention, when it happens in the first 4 weeks of a course, has shown meaningful results in pilot programs.
The Bigger Picture: What This Means for the Tech Industry
Here's the uncomfortable downstream question: What happens when these students graduate?
The tech industry is already beginning to see the effects. Hiring managers at major tech firms report that new graduate candidates in 2025–2026 cohorts are increasingly unable to pass algorithmic interviews or whiteboard math problems — even candidates from top programs.
The irony is sharp: the very AI tools that were supposed to make developers more productive are, when misused in education, producing developers who are less capable of the foundational thinking that good AI usage actually requires.
You need to understand algorithms to know when an AI-generated algorithm is wrong. You need to understand probability to know when an ML model's output is nonsense. AI amplifies competence — it doesn't replace the need for it.
Frequently Asked Questions
Q1: Is AI use actually banned at Berkeley CS?
Policies vary by course and assignment. Most Berkeley CS courses permit AI for certain tasks (debugging, learning) but prohibit it for graded problem sets. Always check your specific course syllabus. The bigger issue isn't policy compliance — it's that students who use AI to avoid learning are hurting themselves regardless of what's technically permitted.
Q2: Are failing grades at Berkeley really caused by AI, or are there other factors?
It's almost certainly multi-causal. Post-pandemic learning gaps, increased enrollment, and course difficulty adjustments all play roles. However, the correlation between AI tool adoption and the specific pattern of homework-exam performance divergence is statistically notable enough that faculty are treating AI dependency as a primary contributing factor.
Q3: Can AI tools actually help students learn math and CS if used correctly?
Yes, genuinely. The research distinguishes clearly between AI used for generation (do this for me) and AI used for explanation (help me understand this). The latter shows positive learning outcomes. Tools like ChatGPT and Claude, when prompted as tutors rather than answer machines, can be highly effective learning aids.
Q4: What should I do if I've already been relying too heavily on AI in my coursework?
Start with an honest self-assessment: can you solve last semester's problem sets without AI? If not, spend time this week working through them manually. It will feel harder than it should — that's the point. Rebuild the habit of productive struggle. Office hours and study groups are underutilized resources that can help bridge the gap quickly.
Q5: Is this trend reversible?
Yes. Cognitive skills, including mathematical reasoning, respond to practice at any stage. Students who recognize the problem early and deliberately rebuild foundational skills can recover fully. The students most at risk are those who don't realize there's a gap until they're failing an upper-division course with few options left.
Final Thoughts and Call to Action
The story of failing grades soaring with AI usage and dwindling math skills in Berkeley CS classes isn't a story about AI being bad. It's a story about a powerful tool being used in ways that undermine the very people using it.
AI is not going away — nor should it. The developers, researchers, and engineers who will build the next decade of technology need to be fluent with these tools. But fluency requires a foundation. You cannot be a skilled AI collaborator if you don't understand the domain the AI is operating in.
If you're a student: Take this as a wake-up call, not a guilt trip. Audit your own AI usage this week. Are you learning, or are you just completing? The difference will show up on your next exam — and in your career.
If you're an educator: The data from Berkeley is a gift. You have time to redesign assessments and implement AI literacy curricula before this becomes your department's crisis.
If you're in industry: Consider how your internship and entry-level programs can help bridge this gap — and what signals in hiring processes actually distinguish genuine competence from AI-assisted performance.
Have thoughts on AI use in CS education? Drop a comment below or share this article with a student or educator who needs to see it. And if you found this useful, check out our related coverage on [INTERNAL_LINK: responsible AI use in higher education] and [INTERNAL_LINK: how to prepare for technical interviews in the AI era].
Top comments (0)