Over the past year, as generative AI tools have become common in college classrooms, much of the conversation has centered on academic integrity: how to detect AI use, how to redesign assignments and whether traditional grading still works. These are important questions. But in my experience as a faculty member teaching computer systems and parallel computing, AI has surfaced a deeper issue, one that existed long before ChatGPT entered the classroom.
AI is not primarily changing how students learn. It is revealing how often our courses have allowed students to succeed without fully understanding what they were doing.
When Working No Longer Signifies Understanding
I first noticed this shift during a programming assignment in an upper-level systems course. The task was straightforward: Parallelize a loop and analyze its performance. Many students submitted code that compiled, ran correctly and even produced reasonable speedups.
But during follow-up discussions, something felt off.
When I asked students why one version performed better than another, or what would happen if the data size changed, several could not explain their own code. Some pointed vaguely to “overhead” or “threads” but couldn’t connect those words to actual behavior. A few admitted that they had used AI to generate the initial version and then modified it slightly until it passed the tests.
The code worked. The understanding did not.
This was not a case of dishonesty. It was a case of misplaced confidence on both sides: Students assumed that working output implied comprehension. And, if I am honest, our assessment structure had encouraged that assumption.
Why AI Makes the Gap More Visible
Before AI, producing a working solution usually required enough effort that students were forced to engage with the material along the way. That friction masked weaknesses. Now that friction is gone.
With AI, students can generate code that looks polished and sophisticated in seconds. But the ability to produce a solution has become decoupled from the ability to explain it. When asked to reason about performance, memory behavior or design trade-offs, many students struggle in ways that were less visible before.
This becomes especially apparent in courses like parallel computing, where correctness is only the starting point. A program can produce the right answer while being fundamentally inefficient or poorly designed. In those cases, students cannot rely on output alone; they must reason about how the program interacts with hardware and resources.
In one assignment, for example, two student solutions produced identical results. One scaled cleanly; the other slowed dramatically as input size increased. When asked to predict which would scale better before running the code, several students guessed incorrectly, even though the answer followed directly from concepts we had covered in lecture. The AI-generated solution looked “better,” but the students couldn’t explain why.
That moment made something clear to me: AI wasn’t hiding misunderstanding. It was revealing it.
This Is Not Only a Computer Science Problem
Although I see this most clearly in systems courses, the pattern is not unique to computer science.
Colleagues in engineering describe students who can generate technically correct designs but cannot justify their assumptions. In data science, students can produce models without understanding why certain features matter. In lab-based sciences, students can follow procedures without grasping underlying mechanisms. In writing courses, AI can produce fluent prose that masks weak argumentation.
In each case, AI accelerates production but does not strengthen reasoning. The risk is not that students are using AI: It is that our assessments often reward the appearance of competence rather than the thinking behind it.
AI as a Diagnostic Tool, Not the Root Cause
It is tempting to frame this as an AI problem. I think that would be a mistake.
AI did not create shallow learning. It exposed how often we relied on proxies for understanding: correct answers, clean code, polished writing. Those proxies worked when producing them required sustained effort. They work far less well when generation becomes trivial.
In that sense, AI functions as a diagnostic tool. It reveals a mismatch between what we say we value—critical thinking, reasoning, judgment—and what our assignments actually measure.
Once I began viewing AI this way, the question shifted from “How do I stop students from using AI?” to “What evidence of understanding am I actually asking for?”
What I’ve Changed
I haven’t banned AI in my courses. Instead, I’ve changed what counts as success.
Some of the most effective adjustments have been simple:
- Requiring explanation, not just submission. Students now explain why their solution works, not just what it produces.
- Asking for predictions before execution. Before running code, students must describe what they expect to happen and why.
- Using comparison instead of construction. I often provide multiple solutions, including AI-generated ones, and ask students to evaluate trade-offs.
- Grading reasoning explicitly. A correct answer with weak reasoning no longer earns full credit.
These changes do not eliminate AI use. They make its limits visible. Students quickly realize that AI can help them get started, but it cannot replace understanding.
A Broader Question
The challenge AI presents is not primarily technological. It is pedagogical.
If students can generate acceptable work instantly, then higher education must be clear about what it offers beyond production. The value of a course cannot rest solely on whether a student can produce an answer. It must rest on whether they can explain, critique and adapt that answer in new contexts.
Parallel computing simply makes this tension harder to ignore. But the underlying issue applies across disciplines.
AI is forcing us to confront an uncomfortable question: Have we been assessing learning or just output?
If we take that question seriously, AI becomes less of a threat and more of a catalyst. It pushes us to design courses that make thinking visible, that reward reasoning over reproduction and that treat understanding as something that must be demonstrated, not assumed.
That shift is not easy. But it may be one of the most valuable changes AI brings to higher education.
