Despite the furor around ChatGPT at the moment, I’m not excited. I don’t think it will be the doom of academics or even current approaches to assessment and I also don’t think it is (or will be) a very good tool to incorporate into learning activities or any creative work. I think its indicative of some places where we might have new tools to help solve problems and spark creativity. In terms of learning technology we will eventually need to pivot our techniques to incorporate good AI assistance (or assistants), but I don’t think that’s where we are in January 2023.

A Biassed History of AI

In short, Artificial Intelligence is a strange field in computing. It’s one of the earliest subfields of computer science, in 1956 the Dartmouth Summer Research Project laid out a lot of what the field continues to work with today. Since then AI has had summers, when it seems like a general purpose artificial intelligence is almost here and will solve all our problems, and winters when the failure to produce those solutions means that the field as a whole has failed.

The other odd aspect of AI is that it tends to spawn other fields in computer science, some of these, like Computer Vision are obvious, but others like Human-Computer Interaction are an outcome of the question of how can we ask a computer to do things. Now we see the field of Machine Learning standing more-or-less on its own away from the rest of AI. A field tends to be AI right up until it becomes popular, and then it’s just computing.

I started working in AI during a winter, roughly in 2004 and staring a master’s degree in 2007. My introduction to AI came from the perspective of Knowledge Based Systems. We focused on how to organize information in such a way that we could find solutions to hard problems (in the computer science sense) without necessarily having to find the optimal solution. We used mechanics like And-Trees and Or-Trees (and their families) and Genetic Algorithms and when we talked about machine learning we focused on systems that tended to build knowledge representations like case-based reasoning or knowledge trees. Our work focused on building systems that experts in fields could describe and use.I liked working on these projects because there was as much interesting information coming from the field as the AI techniques.

We tended to skip over neural networks as tools. Neural networks also contain knowledge but in a different way. In a neural network, input values are fed into nodes, which calculate an output which is fed other nodes, or fed back into earlier nodes or taken as output. Knowledge is encoded in the weights given to different inputs to the nodes and by adjusting the weights you can “train” the system to recognize something in the inputs. These systems are capable of doing tasks, very well and very quickly, but it’s difficult (or impossible) to describe any of the knowledge, so it wasn’t very interesting for our purposes.

That lack of knowledge description, or the system’s ability to describe how it arrived at the answer it did, is something of a problem. If I ask a rule based system why it applied the rule it did, it can show input values and equations and how rules where processed and tested. A neural network can’t don’t that (and neither, really, can a human mind, whose neurons are the model for the neural network). This is fine, but it does mean that there’s no solutions when odd behaviour happens like when an image identifier which is trained to recognize cats also recognizes certain patterns of static as cats as well. There’s nothing we can do to sift out how that happened and even if we train the system again we can’t be sure that there are other things where the system will be absolutely 100% convinced that the picture is of a cat.

None of that is a real problem, but it was never where my interests lay. My PhD was focused on ways to build decentralised networks of agents to help do tasks (like figure out when your alarm clock needs to go off to get you to work on time based on the weather). Meanwhile computing got faster, a lot of problems were redesigned so that a GPU could solve them by doing a lot of identical math quickly. The world entered into an AI summer focused around neural network based machine learning and we had large neural network systems like AlphaGo suddenly performing complicated tasks quite well, based largely on the amount of computing we now had to throw at them. 

ChatGPT

Large language models, which include GPT (Generative Pre-trained Transformer) also snuck up on me. These contain, for a given bit of text, a probabilistic model for what text might come next. The modern versions read terabytes of information and use them to calculate those probabilities and part of what makes GPT3 the current model (excluding the chat part) work as well as it does is that it uses very large bits of text, around about 2000 words at a time. When you ask these systems to write something, they’ll take some seed text and then see what words, probabilistically, should come next.

I suspect, although I don’t know, that the “correctness” of the output depends a lot on the frequency of phases and phrase structures appearing correctly in the training set. I think this is why things like introductory computer science questions are answered fairly well, with specific answers where as things like discussion around something like research design tends to be quite vague and generic and why on some topics the text is nonsense. The modern systems at least improve the overall coherence to earlier systems like Eliza and Alice.

Still however coherent and correct, it’s important to keep in mind that there’s no knowledge in the ChatGPT mechanism and no mechanism to monitor the quality of the output. In human terms, if you’ve ever found yourself speaking out loud saying something surprising, that’s pretty much the state GPT is always in. 

It’s worth noting that the chat in ChatGPT comes from a second model which they trained on example chats. Trainers wrote examples consisting of both sides of the type interaction they would like ChatGPT to have and then trained the system on those examples to teach the system to interact with people. Again I think the larger bits of text helps ChatGPT sound less disjointed than older chat bots, but I think this layer of training also leads to the extreme confidence in its responses that makes a lot of people uncomfortable with ChatGPT. 

AI and Creativity

When I create art, a lot of what I enjoy doing is responding against randomness in the medium I’m working in, scribbles or paint splotches into pictures, or stencils creating shapes against random backdrops, so I understand the urge to use ChatGPT as a starting point to spark creativity. 

I’m not sure at the end of the day that there’s a lot of real benefits to incorporating it into teaching. I’m not sure that it’s capable (at least in it’s current form) of generating sufficiently interesting material as a starting point for creativity, or of generating sufficently correction information for a student to respond to. If you ask it to suggest some starting points for stories, the results are uninspiring and sound a lot like a lot of plots that are passe because ChatGPT is trained on literature which has already been written.

I asked ChatGPT to create prompts for short stories and mixed media applications:

All of these are reasonable, and I could imagine them being given in a class context. I had hoped for something I’d be inspired by. Certainly none of these make me excited, and while I may just need to learn the skill of prompting the AI better, I think I’d rather go write and paint.

It’s also interesting to see given the same prompt, ChatGPT produces a remarkably consistent response. 

Teaching in the age of ChatGPT

Looking at a lot of what’s been written on ChatGPT in Higher Education, there’s a general form of, “Talk about Academic Integrity with your Students,” “Talk about ChatGPT,” “Create and assessment that encourages students to intellectually engage with ChatGPT.” I like the assignment Fyfe (Fyfe, 2022) uses where he has students try to write with GPT generated text, which they submit as an undifferentiated piece along with a reflection on how the work affected them. I think there’s a benefit to getting students to engage, but that course was on engaging with these kinds of technologies. I’m not sure the effect is as good if you’re teaching a second year biology course. (If you are teaching a second year biology course, I’d love to chat!)

I have suggested something similar for teaching computer science. Have students work with GPT (or some of the more code-focused Large Language Models) to generate a solution, then have the stutents test and correct ChatGPTs work, then report to the instructor what they learned. I think this gives the student a chance to really think about why code works, but I worry that starting with a coder that’s not able to understand either what the code is supposed to do or why it works may not be good for student’s learning.

By design, ChatGPT and LLMs are knowledge free. They’re designed to take a corpus of text and determine for any particular chunk of text what other text probabilistically comes next. This produces text, and as I keep arguing, if the probable answers are correct, then the text is probably correct, but the correctness is not inherent in the model and in fact isn’t a part of the model.

I think we’ll see later versions of GPT and LLMs get better and better and produce more specific text. We may also see it produce more interesting text. The growth in image producing systems has certainly improved the quality over what was available even a year ago. I don’t think at the end of the day that we’ll see a really major change to art for a bit.

To grow, I think the system needs an underlying structure for knowledge. Understanding that a cat is a thing is going to be necessary to produce systems that are going to be able to do more than just talk about things. I expect that at some point relatively soon we’ll see some combination of an LLM with a Semantic Net model like IBM’s Watson and then we might be able to start taking these systems at least a little seriously.

From a higher education perspective, I’m very drawn to John Warner’s call to Make the work worth doing. What I think ChatGPT illustrates is that a lot of the things we ask people to do when they learn are rote and default. We pick them because they’re easy to explain and easy to mark. 

I think the real wake up call here, is to create learning experiences that are relevant to the real students in the room with you right now. Build trust with them and show that you trust them. Let them participate in defining what they need to learn and how they’re going to evaluate what they’ve learned. You’ll have to support them in that process, navigating their own learning and the things they can learn, but if they’re doing the work they want to do then they’re not thinking about how they can get around the system.

 

Fyfe, P. How to cheat on your final paper: Assigning AI for student writing. AI & Soc (2022). https://doi.org/10.1007/s00146-022-01397-z