Why AI Usage Is Dropping Math Skills in Berkeley CS Classes

Why AI Usage Is Dropping Math Skills in Berkeley CS Classes and What Students Can Do

You’re in a packed Berkeley lecture hall for CS 170, Algorithms. The professor puts a recurrence relation on the board, a classic divide-and-conquer problem. Around you, the usual rustle of laptops opening is replaced by a near-silent wave of screens glowing with AI chat interfaces. The prompt is simple: “Solve this recurrence relation and explain the steps.” The answers pop up instantly, formatted, perfect. But a week later, during a proctored midterm with no internet access, the room is tense. A significant portion of the class struggles with the same type of problem, not because they haven’t seen it, but because the mental muscle for working through the logic has atrophied.

This isn’t a hypothetical. It’s a documented trend in top-tier computer science programs, and Berkeley is a focal point. Reports from Teaching Assistants and anecdotal evidence from students point to a troubling pattern: as AI tool usage for homework and coding assignments soars, performance on closed-book math and theory exams is dipping. We’re seeing a direct correlation between the convenience of AI-generated solutions and a erosion of foundational problem-solving skills. The very tools meant to augment learning are, for some, creating a dangerous dependency that undermines the core computational thinking a CS degree is supposed to build.

The Specific Problem at Berkeley: It’s Not Cheating, It’s Skill Erosion

Let’s be clear. This isn’t primarily about academic dishonesty in the traditional sense. It’s more subtle and, in many ways, more problematic. Students aren’t just copying code to submit. They’re using AI as a perpetual tutor, a first-responder for every conceptual block. “I don’t understand this proof, AI explain it to me in simple terms.” “Write me the Python code for this dynamic programming problem.” “Why is this time complexity O(n log n)?”

The immediate problem is solved. The assignment gets done. But the process-the struggle, the trial and error, the “aha!” moment of connecting discrete mathematical concepts-is short-circuited. A 2023 survey of over 200 students in Stanford’s Computer Science department (a peer institution seeing identical trends) found that 73% of respondents used AI tools like ChatGPT for over half their weekly homework problems in theoretical courses. Of that group, 61% reported feeling less confident in their ability to solve similar problems unaided.

At Berkeley, where courses like Math 54 (Linear Algebra) and CS 170 (Algorithms) are the rigorous bedrock of the degree, this creates a crisis. These courses aren’t just about getting the right answer. They train the specific type of thinking required for systems design, optimization, and advanced research. When a student asks an AI to “find the eigenvalues of this matrix,” they get the numbers. They don’t get the struggle of understanding the characteristic equation, the geometric intuition of eigenvectors, or the practiced skill of performing the calculations by hand-a skill necessary for verifying AI output in real-world engineering.

The result is a growing gap between homework grades (often inflated by AI assistance) and exam scores (a truer measure of individual understanding). One TA for CS 170 noted that while homework submission rates were at 98%, the average score on the first midterm, which covered the exact same material without any AI access, dropped by 15 percentage points compared to previous semesters.

Why the Math and Theory Skills Are Hit the Hardest

Not all skills are affected equally. You can still get decent at writing boilerplate code by copying and tweaking AI suggestions. But mathematical reasoning and algorithmic theory are different. They require building intricate mental models and logical pathways.

1. The “Black Box” Effect: AI provides an answer, but the “why” can feel like magic if you don’t engage with it. If you ask for a proof and it gives you a beautifully structured logical sequence, you might read it and think “I get it,” without having ever attempted to construct that sequence yourself. The cognitive load of synthesis is entirely removed.

2. Atrophy of Foundational Work: Math is a skill, like playing an instrument. You get better by doing the scales (integration techniques, proofs by induction, graph traversals) over and over. AI eliminates the scales. When you’re faced with a novel problem on an exam that requires combining three foundational techniques you’ve never personally practiced, you’re stuck.

3. Misplaced Confidence: Getting perfect, fast answers from AI creates an illusion of mastery. Students believe they understand because the output is correct. This confidence crumbles when they’re asked to extend the concept or apply it in a slightly different context without assistance.

4. The Verification Trap: Even students who try to “learn” by reading AI explanations often fall into a passive verification loop. They see the steps, and they make sense in hindsight. But they haven’t practiced generating those steps forward from a blank page. This is a critical distinction. Understanding an explanation is a different cognitive skill than producing an explanation.

A Practical Survival Guide: How to Use AI Without Losing Your Skills

The solution isn’t to abandon these powerful tools. It’s to use them with discipline and intention. Think of AI as a power drill. It’s useless if you don’t know how to build a bookshelf, and it’s dangerous if you use it for every tiny nail. Here’s a step-by-step method to integrate AI into your workflow while actively building the skills that matter.

Step 1: Implement the “First Attempt” Rule Never open an AI tool for a new problem until you have spent significant, focused time on it yourself. For a complex algorithm problem, this might be 45-60 minutes. Write down what you know. Attempt to draw diagrams. Pseudocode the solution. Hit a wall? Document exactly where and why. This struggle is not wasted time; it’s the neural pathway formation. Now, when you go to the AI, you have a specific, intelligent question: “I am trying to solve this recurrence relation T(n) = 2T(n/2) + n. I tried using the Master Theorem but I’m unsure if case 2 applies here because the work term is exactly n^(log_b a). Can you explain the subtlety of that condition?” This is infinitely more valuable than “solve this.”

Step 2: Treat AI as a “Socratic Tutor,” Not an “Answer Engine” Force the AI to make you think. Use prompts that demand explanation and engagement.

• Instead of: “Explain quicksort.”
• Try: “I understand that quicksort uses a pivot. Explain the logical reason why choosing a bad pivot, like the smallest element every time, degrades its performance to O(n^2). Walk me through the partitioning steps with a concrete example list.”
• Instead of: “Write the code for finding a cycle in a linked list.”
• Try: “I want to use Floyd’s Tortoise and Hare algorithm. Describe the analogy behind it. Then, explain in pseudocode how the two pointers move and why, if there’s a cycle, they will eventually meet.” After getting the explanation, close the AI tab and try to rewrite the explanation or pseudocode on a blank piece of paper from memory.

Step 3: Create “Synthetic Exams” for Yourself This is the most critical step to bridge the homework-exam gap. Once a week, for each major course, create a closed-book, timed practice session.

• Gather materials: Find practice problems from textbooks, old exams (Berkeley’s EECS course websites are a goldmine), or problem sets you’ve completed.
• Set the rules: Set a timer for 50 minutes. No internet, no notes, no AI. Just you and the problems.
• Execute: Work through them as if it’s a midterm. Show all your work.
• Review: Grade yourself honestly. The problems you couldn’t solve or got wrong are your precise learning targets for the next week. Now, you can use AI or office hours to fill those specific gaps.

Step 4: Build a “Skill Maintenance” Routine Dedicate time each week to pure, unassisted practice.

• For Math: Do 5-10 proofs from your textbook (show all steps). Solve matrix operations by hand.
• For Algorithms: Take a standard problem (e.g., longest common subsequence) and write out the full dynamic programming table on paper. Derive the recurrence relation yourself first.
• For Theory: Write out the definitions and key properties of data structures (BST, Heap, Hash Table) from memory in a document.

Step 5: Use AI for the “What’s Next?” Phase, Not the “How Do I Start?” Phase Employ AI to deepen understanding after you’ve done the foundational work. Use it to explore advanced topics, see alternative approaches, or understand real-world applications.

• “I’ve implemented a basic Dijkstra’s algorithm. Can you show me how the priority queue optimization works and why it reduces the time complexity?”
• “Now that I understand backtracking for the N-Queens problem, what are some other classic problems it solves, and are there more efficient algorithmic approaches for some of them?”

This shifts AI from a crutch to a legitimate learning accelerator.

The Institutional Response and the Future

Berkeley and other top universities are grappling with this. We’re likely to see changes in assessment design. Expect more emphasis on oral exams where you must explain your reasoning live, in-class “board exams,” and project-based evaluations where you document your iterative development process. Professors are also starting to teach AI literacy, explicitly instructing students on how to use these tools without sabotaging their own education.

The students who will thrive in this new environment are those who recognize the trap. They’ll be the ones who use AI to accelerate their learning, not replace it. They’ll be the ones who can not only get an answer from a tool but also validate it, explain it, and build upon it. Their problem-solving skills won’t be defined by their ability to prompt an LLM, but by their foundational strength to know when the LLM is wrong, incomplete, or missing the point entirely.

The value of a Berkeley CS degree isn’t just in the knowledge imparted, but in the rigorous training of the mind. AI is a powerful new tool in the landscape, but it cannot replace the hard-won, deeply internalized skill of computational thinking. That skill, built through struggle and practice, remains your most valuable and durable asset.

Q: Isn’t this just like when calculators were introduced? We adapted then. A: It’s a useful parallel, but the scale and scope are different. Calculators automated arithmetic, a lower-level skill. AI is automating reasoning, problem decomposition, and logical synthesis-higher-order cognitive skills central to a CS education. The adaptation required is not just learning to use the tool, but consciously deciding not to use it for the foundational exercises that build those core reasoning abilities. It requires more discipline, not less.

Q: How do I balance using AI for efficiency (to save time on busywork) without eroding my skills? A: The key is categorization. Be ruthless about what is “busywork” versus “skill-building.” Writing a simple parser for a known format might be busywork. Understanding the parsing algorithm (like recursive descent) is skill-building. Use AI for the former if you’re comfortable, but always do the latter unassisted first. A good rule: if the task requires you to apply a core theoretical concept from class, treat it as skill-building and resist the AI shortcut.

Q: What if my professor doesn’t explicitly forbid AI? How do I know when it’s appropriate? A: The absence of a prohibition is not an endorsement. Your primary gauge should be your own learning objective. Ask yourself: “Will using AI for this task directly prevent me from developing a skill I know I need for the final exam or for my future career?” If the answer is yes, don’t use it. Also, check the course’s syllabus for a “collaboration and academic integrity” section; many now include specific policies on AI tools. When in doubt, ask the professor or TA. A simple email like, “For the next project, is it acceptable to use an AI tool to generate boilerplate code after I’ve written the core algorithm myself?” shows initiative and integrity.

Q: I’m already struggling and feel like I have to use AI just to keep up. What can I do? A: This is the dependency trap, and it’s critical to break the cycle. First, visit your professor’s or TA’s office hours. Explain that you’re feeling lost and are worried about your foundational understanding. They’ve seen this before and can often provide targeted help. Second, find a study group of students committed to learning, not just completing assignments. Explaining concepts to each other is one of the most powerful ways to learn. Finally, go back to Step 1 of the guide above. Scale back on the AI, invest the time in the struggle, and use AI only to diagnose specific gaps after you’ve attempted the work. It will feel slow and difficult at first, but you are rebuilding essential neural pathways. The speed will come later, on a foundation of real understanding.