Academic integrity in an age of LLMs (#1410)

Topics/tags: Teaching

I spend way too much time thinking about issues of academic honesty. Do all faculty? I’m not sure. These days, such thoughts are inevitable for most CS faculty. After all, the world is telling our students that Large Language Models (LLMs) are the future of programming, so many of them are turning to tools like Claude Code, GitHub Copilot, or ChatGPT to help them with assignments. Unfortunately, while using such tools may make programmers more productive at developing code, they are unlikely to have positive effects on student learning. After all, the real goal of any assignment isn’t to produce the resulting code or analysis; it’s to build knowledge, which comes as you write.

Most students understand these issues and avoid LLMs when they do work for their CS courses. Well, that’s not quite true; many still rely on LLMs when they look something up, such as How do I escape from a loop in Scheme? or What does an anonymous function look like in C++? Is that different than relying on a search engine along with sites like StackOverflow and GeeksForGeeks? Probably not; the reliability of the answers in both cases is similarly uncertain: LLMs hallucinate, while StackOverflow and GeeksForGeeks include some surprisingly uninformed answers.

However, some students focus more on their grades than on their learning. These students seem to be much more comfortable relying on LLMs more than is appropriate. It frustrates instructors. It also frustrates their fellow students, who can no longer rely on these LLM users to be good partners and who worry that the sometimes high grades that LLM users get undermine the grades that the real students get.

Using LLMs violates the class policies, at least for most assignments in most courses in computer science at Grinnell [1]. Using LLMs without citing them also violates broader policies on citation and, therefore, norms of academic integrity [2]. And, like all such violations, they should be reported and lead to both penalties and developmental education.

As I was thinking about these issues the other morning [3], I realized that there are some significant ways in which LLM violations are different than other kinds of violations. In general, when a student violates academic integrity, it’s not a regular habit. Most often, they find themselves stressed and end up asking for inappropriate help or copying from another student. At least that’s been my experience at Grinnell. One violation does not suggest that there are others. And even when it’s likely to be a habit, as in the case of students who hide their backpacks in the bathroom so that they can quickly check their phone or other materials during an exam, it’s often a habit for only one kind of work.

As I said, LLMs are different. A student who uses an LLM for one thing is likely to use it for many or most things. For example, if a course has problem sets, programming projects, and take-home exams, as a typical upper-level CS course has [4], a student who uses LLMs on a take-home exam probably used it on their problem sets and programming projects.

There’s also a huge difference between using an LLM to help you with work and, say, failing to cite something appropriately. In most cases, the former is almost always intentional, while the latter generally does not involve an intent to deceive.

These differences make me wonder whether we should impose different penalties on students who are found responsible for inappropriately using LLMs. Since it’s unlikely that this is a one-time thing, and LLM use is often difficult to detect, perhaps we should simply fail students who are found responsible for using LLMs. Wow, that sounds harsh. Perhaps we should give them the opportunity to redo all (most?) of their work, but with supervision. If they can do it in such a situation, that’s great. If they can’t, we reach a different conclusion. In both cases, they get the grade they deserve.

Perhaps not. Students don’t always do as well under stress. Unfortunately, if a student’s LLM-based work makes us reconsider their prior work, they need to show that the prior work is theirs. I wish I could suggest a better approach.

[1] Many of us feel that we need to give some assignments that explicitly involve the use of LLMs

[2] I prefer the term academic integrity to academic honesty.

[3] I drafted this musing that morning.

[4] Perhaps not take-home exams; many faculty prefer the security of an in-class exam.

Version 1.0 of 2026-05-26.