Using Large Language Models in Learning and Teaching
AI in Teaching and Learning series |
|---|
Image by Stable Diffusion (another tool for generative content)
Using AI in education
There's been a lot of fear and uncertainty around using an LLM in education, mostly on the side of educators worried that their students will use the LLM to generate their assignments. While this kind of academic misconduct may occur for a small number of cases, it shouldn't be expected that the majority of your cohort is about to start using ChatGPT to write all of their assignments. To paraphrase something I heard recently, we need to teach students to use LLMs like they would a calculator - that is, get the LLM to do a basic sub-task and use their own character, creativity, personality and intellect to refine the output into a well-considered body of work. Charles Sevigny from the Department of Anatomy and Physiology in our own School has also written a great piece on why we shouldn't fear LLMs in learning and teaching, especially if we're already following recommendations from online teaching during the pandemic.
Another academic example of using LLMs in higher education is provided by Ethan Mollick from the University of Pennsylvania. He describes his processes in a number of places on his blog which is well worth browsing, but specifically the posts about using it specifically as a tool, and using it to improve writing style which also includes a number of suggestions about prompt crafting and a reminder of some of the limits of these tools.
Closer to home, Open Universities Australia has published a good guide for student use of LLMs. Briefly, they recommend using it to write jumping-off points or proof-reading, but citing the use of the LLM as well, and obviously not to use it to write the entirety of the assignment. Flinders University also recommends using LLMs for explaining complex concepts in simpler terms and for generating quizzes and flash cards to use for revision (something we could also do, especially for fomative assessment items), and also provides a guide on how to reference the use of these AI tools. Deakin University has information on basic prompt crafting and a list of AI tools, their limitations, and ways that they can help students to study.
In the future, LLMs might also be able to offer more personalised feedback for students on their (written) assessment, allowing academics more time as well as reducing the need for casual staff to mark assignments. This will initially require both a training period and a transitory period to ensure the accuracy of the marking, and that the mark reflects the rubric for the assessment. Eventually it may be possible to generate an LLM trained specifically for each assignment, offering a consistent mark over years of assessment. The ethics of this as a standardised practice are most certainly going to be a hotly debated topic over the next few years!
Where do we start?
Getting comfortable with the tools is an obvious inital starting point. Here's a very short list to whet your appetite:
- OpenAI is the original ChatGPT. There are no-login instances that you can try as well.
- Bing Chat is ChatGPT integrated with Web searching, and is now available by default in Microsoft Edge browser. A great example of use is at Ethan Mollick's blog where he asked it to generate an outline for a new paper based on gaps in existing knowledge and suggesting methods consistent with his previous papers.
- BLOOM, an LLM created as multilingual (46 natural and 13 programming languages) from the start. This is still in it's early days, but should improve rapidly as tools for community expansion are added.
- Stable Diffusion is a text-to-image model that generates images based on a prompt. As an example, the image at the top of the page was created with the positive prompt "futuristic, university in space, robot (lecturing:1.1) to (human students:1.7), window to space with stars" and "(text:1.3), (hands:1.1), fingers" as a negative prompt.
NB: The numbers indicate weights that were applied; a slightly increased weight to lecturing and a heavy weight to human students because it kept producing robot students a lot of the time.
Using these tools, we can start playing around with prompts to see what kind of output we can get. LLMs can be used as they are to generate starting points for new ideas and concepts, and using internet-linked LLMs can help with your deeper questions that might need links to papers or companies that offer the types of solutions generated to get a sense of how you might implement them. Stable Diffusion is a great method for producing images that are royalty-free, however they can take some time to generate a result that meets your needs, no matter the prompt crafting - as a case in point, these were all generated from the same prompt as above:
Images generated by Stable diffusion for the positive prompt "futuristic, university in space, robot (lecturing:1.1) to (human students:1.7), window to space with stars" and "(text:1.3), (hands:1.1), fingers" as a negative prompt.
Applied example: MCQ distractors
Given a question (as stem or lead-in), ChatGPT will generate a number of distractors for you. Obviously these will need to be verified by you or the subject matter expert, but as a way to generate the extra one or two that you're struggling to come up with something that seems plausible but not obvious it can be fantastic.
Let's see a very basic example for the purpose of this article:
| > I'd like to write a multiple choice question. Given the question "What is the molecule that binds oxygen in a red blood cell?" give me 4 distractors that seem plausible. |
|---|
Certainly! Here are four plausible distractors for the multiple-choice question based on the given stem: A) Hemoglobin: This option serves as a distractor since hemoglobin is actually the correct answer to the question. Hemoglobin is the protein found in red blood cells that binds oxygen, allowing for its transport throughout the body. B) Myoglobin: This option suggests that myoglobin is the molecule that binds oxygen in a red blood cell. While myoglobin is indeed a molecule that binds oxygen, it is found primarily in muscle cells and is responsible for storing and releasing oxygen within muscle tissues, rather than in red blood cells. C) Erythropoietin: This option implies that erythropoietin is the molecule that binds oxygen in a red blood cell. However, erythropoietin is a hormone produced by the kidneys that stimulates the production of red blood cells in the bone marrow, but it does not directly bind to oxygen. D) Albumin: This option suggests that albumin is the molecule responsible for binding oxygen in a red blood cell. However, albumin is a protein found in blood plasma and plays a role in maintaining osmotic pressure and transporting various substances, but it does not bind oxygen in red blood cells. These distractors provide plausible alternatives to the correct answer while incorporating related molecules or concepts. However, the correct answer, which is not included among the distractors, is hemoglobin. |
Here we see some of the odd quirks of ChatGPT, including the repetition of information, and logically/factually incorrect statements even within its own output ('This option serves as a distractor since hemoglobin is actually the correct answer...' with 'the correct answer, which is not included among the distractors, is hemoglobin').
Applied example: Curriculum development
Need to totally refresh a tired subject, or create a wholly new subject but don't know where to start? Let's try a first year biology subject to minimise the amount of prompt-crafting for this example:
| > Create a curriculum outline for a first year university biology course. Include formative and summative assessment items, and rubrics for the assessment items. |
|---|
Course Title: Introduction to Biology Course Description: Course Goals: Assessment: Formative Assessment:
Rubric: Rubric: Rubric: Summative Assessment:
Rubric: Rubric: Rubric: Rubric: Course Outline: Week 1: Introduction to Biology Week 2: The Cellular Basis of Life Week 3: Genetics Week 4: Evolution Week 5: Ecology Week 6: Review and Midterm Exam Week 7: Plants Week 8: Animals Week 9: Human Biology |
Okay, not a bad start. Looking at those assessment items, it seems to put a lot of importance on organisation and clarity, grammar and spelling, and formatting since they are in nearly every one. I'm not sure how well that would align with your subject's learning outcomes, but again as a starting point that's not terrible. Also really impressive that it spits all of that out in seconds, where it would take quite some time to type it manually (NB: formatting that nicely for this page template took me nearly an hour - maybe I should have used the LLM to troubleshoot that!).
In your own subject, the weekly structure will obviously need your own input and consideration. Using LLMs as a tool to prompt you is an excellent way to see different aspects or new teaching methods in a subject you know very well, and allow you to incorporate new content or delivery.
As we mentioned before, we can (and should!) encourage students to use these tools ethically, and we should of course hold ourselves to the same standard. Generating an outline like this is a fine timesaver, and can prompt you in different directions that you may not have considered. Provided we keep adding all the creativity, critical thinking and passion to our work, then everyone can benefit from using these tools.
AI in Teaching and Learning series |
|---|
Written by Jairus Bowne, Academic Specialist in the School of Biomedical Science's Digital Learning Hub and ChatGPT. |