Recently, Artificial Intelligence (AI) software like Chat Generative Pre-trained Transformer (ChatGPT) has become a common tool used by students in higher-education.
ChatGPT is an online AI software that was released in November 2022 by OpenAI. It has gained popularity by quickly providing lengthy and detailed responses in a variety of different knowledge fields. However, it has also been found to produce uneven factual evidence.
Ethical issues have arisen as to whether this software can be considered plagiarism or a tool for students, and how it will influence the future of education in different fields.
The Gateway sat down with Matthew Guzdial, an assistant professor in the department of computing science, to discuss how AI software can be utilized in the education system, outlining its positives and negatives.
Responses have been edited for brevity and clarity.
Q: What are your thoughts on ChatGPT or similar software?
ChatGPT follows on with a number of other earlier GPTs that were also transformer models. ChatGPT is a little bit more interesting than some of these earlier ones because of how it was trained. In addition to a sort of supervised learning, it tries to generalize the data. It was also trained via reinforcement learning.
Q: What could the implementation of software that detects AI-generated responses look like?
The way that these transformer models work is that they take in one line and then predict the next. The way that they’re trying to do that is by processing the line of text and trying to figure out “what should I pay attention to in this input text?”
That means that if I feed the same transformer two different lines of text, but the words that it has learned to pay attention to are similar, it will give me a similar response. This means that we can figure out if the text is likely to have been generated by something like a transformer model.
Q: If a student uses AI, is it always plagiarism? What makes it plagiarism?
That depends on how you use it. Transformer models like ChatGPT are currently being used in the [tech] industry for writing texts, but still with someone editing the final text right before it’s sent off somewhere. Technically, all the words that ChatGPT has ever seen have already been written by somebody else, but it’s not exactly quoting.
You can make it quote exactly from its training data. If a student is copying and pasting a large amount of text output from ChatGPT, the student might be plagiarizing ChatGPT, because they’re not using their own words. They’re trying to pass off somebody else’s words as their own. But, is ChatGPT plagiarizing? Because how many of those sequences of words are just things that ChatGPT got from this massive text scraping?
Q: Do you think there are any possible benefits that can come from AI, such as ChatGPT, when used in an educational context? What about the disadvantages?
There’s lots of stuff happening in this area in terms of how AI can figure out the correct order of things, in terms of the order of problems to give you so that you learn the best. Or, how can we use AI for question answering? Or as an automated teaching assistant (TA)? There’s tons of stuff happening in this space.
If a student is just learning to copy and paste from ChatGPT, this is just as bad as a student learning to copy and paste from the internet. Maybe worse, because the sources they find on the internet are probably accurate versus ChatGPT, where sources are going to be fictional much of the time.
Students are always going to try to find ways to not have to learn if they can avoid it. Learning is hard, and they’re not always taught particularly well. If students are interested in their own education, I wouldn’t recommend engaging with ChatGPT as an instructional tool.
Q: Are educators, as of now, able to distinguish original responses from AI-generated responses?
We can’t say for educators at large. People who are used to looking at transformer-generated text can distinguish it pretty easily. There are set things that transformer texts tends to have, but that assumes that you have that experience of looking at this kind of text.
If we’re talking about an English literature professor, they’re not necessarily going to be able to distinguish it. You can’t just go and educate a whole bunch of professors. They don’t have the time or the interest.
Q: Have you noticed any changes in assignments and coursework after ChatGPT came out?
Not with ChatGPT, because that was fairly recent, but with Codepilot on GitHub. Codepilot is like ChatGPT, but for computer code instead. So some of the professors in computing science have even been encouraging students to use it because unlike language, computer code has to compile.
If it gives you bad outputs, it’s immediately obvious. Some professors have changed how they set up assignments based on that. I know some professors, not here at the U of A, but elsewhere, have changed from more essay-based things to more short answers that refer back specifically to things that occurred in a lecture.
ChatGPT did not attend that lecture — there’s no way it could possibly know to reference these things in this way.
Q: Is there anything you would like to add?
Another thing that students are likely to be aware of is Mid-Journey or Stable Diffusion, the art generation AI. It’s quite a bit different from ChatGPT, because it uses a diffusion model. Diffusion models are much more likely to plagiarize, to directly take things or even just replicate stuff from their training data — things that they’ve seen before.
There are also legal issues around these. It’s very possible that right now, as it is, you can’t copyright something that was generated by one of these AI systems. That’s possibly going to be true for quite a while. Using them in certain cases might open somebody up to be sued.
One big thing for students in the faculty of arts, you probably don’t need to worry too much about one of these things stealing your job. Industries where people are using these things are using them to give an artist or a writer a starting point and then let them edit from there.
These things are brittle and liable to break in weird ways. Nobody’s using them live. Or, nobody who knows better is using them live. These things are potentially useful tools to help with the process to get you through a first draft faster, but they’re not all-powerful.
Every year, The Gateway publishes hundreds of articles like the one you just read that are free for everyone to access. But The Gateway needs your support to continue publishing its award-winning journalism. Please consider donating today, even a small amount can help the University of Alberta’s only newspaper continue serving the campus community. Thank you.