3  Academic integrity and copyright

Most of the current conversations regarding LLMs in higher education environments refer to academic integrity and copyright. This is something that is increasingly in instructors’ minds and that needs careful consideration and clear communication.

In my experience, there is a wide range of perspectives from instructors and students about this. I will not side with any one particular approach or recommendation. However, I do believe that it is fundamental to provide clarity and not to assume that there is already a common sense about this. Artificial Intelligence usage by the wider population is very recent and we are still adjusting and defining expectations of what is acceptable from it.

3.1 Academic Integrity

Some fields in higher education have engaged more in conversations about academic integrity than others. This is highly dependent from context, student body, and instructional preference. However, most of the discussion regarding AI in higher education has been centered at academic integrity.

There are different viewpoints on what is and what is not allowed in a course. My focus here will not be to determine what is permissible or not in a course, but what considerations could be effective and clear to avoid issues concerning academic integrity.

3.1.1 Syllabus

A common conception for course syllabi is that they provide a type of contract between instructors and students. Similar to a user agreement, syllabi give an opportunity for instructors to clearly outline the expectations for students in a course. As such, this is a great place to clarify what is allowed or not, or expected or not with respect to LLM usage.

For students For instructors
Make sure to read the “AI Policy” and/or “Academic Integrity” section of your syllabus. If these are too generic or not clear, reach out to your instructor and request they clearly state to what extend they allow LLMs to be used in your class. Include very detailed “AI Policy” and/or “Academic Integrity” section of your syllabus. Consider giving a few explicit examples of expected usage.

Academic integrity matters can be tricky. An effective way to explore what works in your particular case is to go beyond allowed v. forbidden usage and to consider to what extend can something be used.

For example, to completely forbid LLM (or AI) usage in a class might not be realistic nor productive. On the other hand, to completely allow unrestricted usage might be detrimental for student learning. Try to see this as a spectrum. The task then becomes to figure out where in that spectrum are you located for the specific course.

SAMPLE AI POLICY

This is part of the AI policy that I am including in my most recent Introduction to Proofs course:

Usage of generative AI

The use of generative AI in academic environments must be considered with care. A good rule to keep in mind that using generative AI to replace what a human could do might conflict with general rules of academic honesty. For our course, the recommendation is to avoid the usage of generative AI tools for assignments unless explicitly stated. When this is allowed, please make sure to cite, attribute, and/or describe to what extend you used it in your work and learning.

3.1.2 The first day of class

In the spirit of communication and clarity, I have been spending a good part of my first class day to discuss about AI in our course. This allows me to explain my perspective and to hear from students about their own expectations. As an instructor, it is interesting to hear students share their expectations.

Students have enthusiastically appreciated this exercise. They value hearing not only the policies, but also the reasoning, concerns, and hopes behind them, as well as the perceptions from other students. This is also a great way to find out for new creative ways to use LLMs, or to flag usage as inappropriate.

I start my courses with some prompts to the class such as:

  • How do you typically use AI for your courses?
  • What would you recommend to other students?
  • What do you avoid or think is inappropriate?
  • What are your concerns about AI usage?

3.1.3 Assignments

Sometimes syllabi can be a bit generic with policy and could fall short on specific considerations for particular assignments.

For students For instructors
Make sure to ask your instructor about the extend to which you can use LLMs for each assignment type. When in doubt, a general good practice is to disclose and describe how did you use LLMs. Give examples of what prompts you used. When necessary, include the specific LLM and version used. Include in your assignment instructions to what extend students can use LLMs and what kind of attribution they are expected to include. Ask to include what prompts where used.

LLMs can be thought of as a very sophisticated tool. Just as referencing Wikipedia has become normal, citing or referencing LLMs is, in general, a good practice. However, LLMs can also be thought of as more than just tools. They also act like agents, which gives them an air of autonomy and decision making. In this sense, it is also useful to think about LLMs as entities, hence the idea of attribution. This might depend on the usage.

For example, if an LLM was used to find a synonym, attribution might not be necessary, however if it was used to generate an example or create a summary, this might be the case.

A good rule of thumb is to think what would be the best practice if instead using an LLM we would’ve asked a peer to do the same task. Would we had attributed their help?

Some ways in which you can attribute or disclose LLM usage is by including the following, either in assignments or class material:

  • prepared with the assistance of AI
  • AI was used for generating graphics and schematics
  • example generated with the assistance of AI

3.1.4 Community guidelines

Each course is different. Not only due to the subject matter, but also due to everyone’s background and values. Promoting discussions regarding academic integrity considerations can also be a good team-building exercise in courses. Even more, sharing the rationale behind why advocating or discouraging certain practices can also get buy-in for both instructors and students.

3.2 An example of AI Guidelines for a course

This is an AI Guideline resource that I share with my students in class. I include this as a separate resource for class (on our Learning Management System) as a complement to my AI policy described in the syllabus.

Artificial Intelligence - specifically Large Language Models- have become more popular in recent time. These can useful or hinder your experience in class depending on the possible usage. 

Expected

Usage of LLMs is not required but could be useful depending on particular tasks or situations. 

  • Citation: cite every time you use an LLM.
  • Translation: you can use it to translate statements or questions to other languages (if English is not your first language) or to rephrase to simplify the language. Search: sometimes searching for references or concepts could be enhanced by using LLMs. As usual, be careful with hallucinations. Check for the actual references and sources.
  • Feedback: you can use LLMs to ask for feedback on your proofs and typesetting. 

Not Expected

  • Do not generate proofs using LLMs.
  • Do not copy-paste outputs of LLMs.
  • Do not solve problems using LLMs.

Useful prompts

  • i am a student in a introduction to mathematical proofs course at ucsc...

  • rephrase the following statement...

  • check my language for consistency and clarity and give me feedback...

  • help me understand this LaTex error...

General considerations

LLMs are mathematical models that generate words (tokens) based on the prior words (tokens) given to it. These are trained by performing regression (gradient decent on a particular loss function). These usually happen on high-dimensional spaces (about a trillion dimensions). As such, achievement of absolute minima is not expected. Usually the regression only achieves a neighborhood of a minimum (with a particular threshold). Therefore, LLMs have a high level of non-deterministic elements.

Training data often considers Big Data (in the order of all public internet information). As such, this requires a huge amount of computing. The latest models usually require a few months of processing in data centers that amount to the electricity usage of a small US city for an entire year. *