Artificial intelligence (AI) and its impact on academic writing with Sandie Elsom

In this episode, we explore the fascinating world of artificial intelligence (AI) and its impact on academic writing. Our guest, Sandie Elsom, Lecturer in Technology Education and AI enthusiast, discusses the challenges of distinguishing AI-generated text from human writing and explores the implications for academic integrity. We'll uncover the linguistic features that often differentiate AI-generated text and examine the effectiveness of current detection tools. Join us as we navigate the complexities of AI in education and discuss strategies for encouraging academic honesty in the age of AI.

Berber Sardinha, T. (2024). AI-generated vs human-authored texts: A multidimensional comparison. Applied Corpus Linguistics, 4(1), 100083-. https://doi.org/10.1016/j.acorp.2023.100083

Liang, W., Izzo, Z., Zhang, Y., Lepp, H., Cao, H., Zhao, X., Chen, L., Ye, H., Liu, S., Huang, Z., McFarland, D. A., & Zou, J. Y. (2024). Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews. arXiv.Org. https://doi.org/10.48550/arxiv.2403.07183

Muñoz-Ortiz, A., Gómez-Rodríguez, C., & Vilares, D. (2023). Contrasting Linguistic Patterns in Human and LLM-Generated Text. arXiv.Org. https://doi.org/10.48550/arxiv.2308.09067

Turnitin AI Technical Staff. (2023). Turnitin’s AI writing detection model architecture and testing protocol. Turnitin. https://www.turnitin.com/

Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., Popoola, O., Šigut, P., & Waddington, L. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19(1), 26–39. https://doi.org/10.1007/s40979-023-00146-z

Transcript

Natalie McMaster: Welcome to our technology and learning podcast. I'm Nat, McMaster, and I'm a member of the AARE Technology and Learning Special Interest Group. Now, today, we're chatting with Sandy Elsom about identifying AI writing in academic content. Hi, Sandy, it's fantastic to have you on the podcast today.

Sandie Elsom: Hi, Nat. Thanks for inviting me in.

Natalie McMaster: The use of AI in academic work is an area that many of our listeners are interested in. Can you start by telling us what the main differences are between AI generated text and human written text highlighted in the research on this topic?

Sandie Elsom: Well, the main thing, I think, is that AI text can't adjust to different context. So they only know as much about the context as the user tells them, and if users aren't using AI effectively, they'll likely tell them very little. So for example, if a student copies an assignment question into a Gen. AI tool and asks it to complete the task. It won't know anything about the content that's actually been taught in the class. So, I did try this myself. I tried to generate a 600 word, 5 paragraph essay to use as an exemplar. And I was using ChatGPT. ChatGPT couldn't do it, because it doesn't know the very specific way that we teach students to write a 5 paragraph essay at UniSC. So if if I were a student trying to offload my essay, I'd have to teach the AI the particular style that I've been learning in class. And if I'm capable of doing that, I'm probably capable of writing the essay, or at the very least I know what the correct style is. So even in the process of trying to cheat, I would actually be learning.

But AIs also just can't replicate those subtle variations that exist in academic and conversational writing. So you know, when you and I have a conversation, for example, we have a shared language because we study and research in similar fields. We can refer to particular events. Maybe we use some jargon that other teacher researchers maybe, wouldn't even be familiar with. Perhaps we follow similar procedures because we both learned the same way. So AI writing is less varied than that in terms of the sentence structure, the language that it's using and overall it's just less engaging.

Natalie McMaster: They are great tips for recognising AI text, but what are the linguistic features that typically distinguish this AI generated text from human writing?

Sandie Elsom: Yeah, absolutely. So AI texts tend to use more formal and a less varied sentence structure and the vocabulary of an AI is more predictable. So there was one study that found that certain adjectives appear more frequently in AI generated text, and some of the words that AI tends to use are commendable, meticulous, innovative, intricate, and versatile.

So when 'Turnitin' are trying to identify AI generated writing, they use identifiers. Some of their identifiers would not necessarily be obvious to a human reader. So they use a thing called long range statistical dependencies. Which means that they look at the way that later parts of a text refer back to earlier sections. So for example, when we're structuring an essay or a paper, the conclusion will refer directly back to the thesis or the key points raised in the introduction. Even though there might be many words or phrases in between. Or we refer to a person by name in one sentence, and then a few sentences later, we refer to that person by a pronoun, like he or she, or they. And this comes naturally to us as human writers. But it's more difficult for Gen. AI. And so AI text might become less coherent as the passages get longer.

There are a couple of other features that turnitin users that we might be able to recognise. So one's called perplexity. And that's about the predictability of word sequences. So AI texts are prediction machines, and they're more predictable in their word choices. The other one is called burstiness, and that measures variations in sentence structures. So human writing naturally shows more variability. So a smooth, predictable text is more likely to be AI, than that sort of general messiness of human writing where we can normally find some clumsiness and some errors.

I can actually give you an example of this. So my students have to write reflections on their learning, and they start by describing the work that they did that week, and then writing a reflection about it. So I've got a reflection here written by student Jenny, who gave me permission to use her work. Jenny had been working through online tutorials, learning how to use a video editor. And she wrote in her reflection:

'Due to academic writing restraints, my reflection for this learning journey cannot ideally voice how I feel about this week's tutorials. Phrases like, "Oh my goodness", "holy doly" and "what the", will need to be left out.

So we can easily recognise that as human, generated because of Jenny's use of informal expressions. The way she refers to academic writing restraints which suggests that she knows that academic writing restraints exist, and what they are, and also we can hear her emotion in her writing. As an experiment, I took Jenny's description of the work that she had done, so the description that informed her reflection and I asked ChatGPT to write a reflection based on it. And this is part of what ChatGPT gave me back.

'Looking forward, I am eager to dive into introduction to editing Part 2 to further refine the trailer I have started. This initial part of the tutorial, despite its challenges, has laid a strong foundation for my editing skills, and has adjusted my expectations for the time I might need to allocate to future sessions. This reflection serves as a reminder of the importance of flexibility and patience in the learning process, especially when acquiring complex new skills.'

So I think it's pretty easy to hear the difference between the 2 styles. Jenny's natural style is less formal and more playful, and she talks about her emotions. Her language is less polished. But the AI text is more general and really could apply to anyone. But Jenny's writing is clearly about Jenny.

Now, just to make this story a bit more interesting. I pasted that reflection written by ChatGPT, into ChatGPT, and I asked it to assess whether it was written by a human or an AI. Now ChatGPT wasn't quite sure, but it thought it was likely written by human. So, despite language which screams AI generated to me as an AI user. It was tricked by the contextual information that I gave it, in the form of Jenny's description of the work that she had done. So adding context to the prompt, can make AI writing harder to detect by software. But you, as a human might still be able to recognise it.

Natalie McMaster: Those examples you shared are very helpful. How effective are current tools at detecting AI generated text?

Sandie Elsom: They are not very effective. At this stage detection tools generally show poor reliability, especially when the texts are subtly modified in some way. There's research to show that at this stage commercial tools like 'turnitin' do not significantly outperform free tools in detection accuracy. That said, however, if the user hasn't made any effort to obscure their use of AI, so they pretty much just paste it in the assignment question. Hit, enter, and then submitted the output. There's a pretty good chance that a detection tool is going to pick that up.

Natalie McMaster: What are the challenges in effectively detecting AI generated content in academic settings?

Sandie Elsom: Well. Obviously slight modifications in the text can greatly reduce the efficacy of the detection tools. So if the users, having a back and forth conversation with the Gen. AI tool and giving input, and working their way towards an answer. And then putting that answer at least somewhat into their own words. Well, commercial detection tools are not going to pick that up. You might be able to recognise it yourself if you're familiar with what AI sounds like, but it's going to be really hard to prove. And something that I think is really interesting, is that question of, at what point is the student just learning the material? You know, if they're learning the content. But if they're having an AI, help them with the structure and language of the essay, does that necessarily matter? You know what's the purpose of the essay? Research is meant to be shared. So it needs to be understandable. What if AI can actually help us disseminate the work that we're doing more effectively than not using AI? You know it doesn't matter if the paragraph is less bouncy or less complex, or are we just sticking to rules because they're there? I think that there's a really fine distinction between using AI as a learning tool and using AI as a not learning tool. And it's entirely possible that the only person who really understands what's going on is the user themselves.

Natalie McMaster: That's a really good point Sandy. Something for a lot of us to think about in teaching and academia as well.

Sandie Elsom: I'm not really sure how controversial that is. At this time. I sort of keep just throwing that out there.

Natalie McMaster: Yeah. Well, I guess if I talk from my perspective, I utilise AI in the courses that I teach. I guide students in how to use AI for learning and I have clear guidelines for use. They need to ensure that they provide me with the prompts they have used and also an explanation of how the output generated by AI has been used. I have a thing about banning things. The minute you ban something, people still use it but try and hide their use. I want students to be open and transparent about how the use of AI has expanded their learning, because we also know that AI tools have limitations.  

Sandie Elsom: Absolutely. And I allow my students to use AI as well. And I give them suggestions for how they can use it ethically. And what I'm trying to achieve there is that they enjoy using AI in the way that they're permitted to use it, and they don't feel it necessary to try and cheat with it. So I've yet to know whether that's effective or not. But I don't see a lot of AI plagiarism in my course. I'm not saying I see none, but it's a fairly minor problem.

Natalie McMaster: My final question for today is how can educators ensure academic integrity in this forthcoming era of AI?

Sandie Elsom: Yeah. So obviously, as the AI models get better, then the AI detection is going to get even harder. It's gonna get more challenging. And I don't think that we're gonna be able to rely on tool assisted detection. We need human oversight. So our assessment design as teachers, academic integrity staff. We all need to be using AI ourselves, so that we have a feel for what AI looks like and sounds like.

And it's not hard to use AI in your life. There are myriad uses for AI in your life, and the more you use it the better you'll understand it. If you're thinking you're just waiting for a good training course to come along and teach you how to use it. Then you need to stop that right now. It's the easiest technology I've ever used. You just need to try things out and develop your confidence. So for me personally, for example, I've got a custom bot that I've created. A little custom GPT, that understands my food intolerances, and it creates recipes for me, based on what I currently have in my cupboard. And I've to taught it how to do that. I used AI to create a role playing mystery game for my family at Christmas, in including costumes and backstories and hidden secrets and everything. And recently, when my cat seemed a little bit off colour, I used an AI vet to review a photo of him to help me decide if I needed to take him to the real vet or not. Which I did, by the way. As learning designers, when we're starting to create our courses and our assessments, I think that we have to require context in the assessments that we set. So create tasks that asks your students to draw on something that happened in class by, you know, reporting on a presentation for example, or responding to a provocative piece of writing that you created.

And and I think the students have to be using AI as well, so that they get used to its strengths and limitations. I want them to see how harmful poor use of AI can be, and what an absolute treasure it can be when it's being used properly to aid learning.

Natalie McMaster: Well that concludes our podcast today. Thanks for sharing your research and experiences with us 

Sandie Elsom: Oh, you're welcome. Thanks Nat. It is one of my favorite topics to talk about, and I love to hear as well what people are doing with AI. Because you know, I think that we can all inspire one another and have a really active, enjoyable learning community around AI and higher. Ed.

Natalie McMaster: I'm sure our listeners are going to come away with some food for thought on the subject of AI to share with their colleagues.

Sandie Elsom: Thanks, Nat.