Leon Furze

If you’ve read any online news recently you’d be forgiven for thinking that we’ve already been back at school for six months, and that ChatGPT has successfully destroyed the education system, killed the essay, and turned every one of our students into a compulsive plagiarist.

I’ve written a number of posts about ChatGPT and AI in general, and I’m trying to keep a balanced perspective. Yes – the technology, and what comes next will change the education system. No – we have not rendered essays and all forms of written communication redundant.

My previous post on ChatGPT was about being creative in our uses of the technology. I believe that in order for the tech to be successfully adopted in education, teachers need to be at the helm and we need to move past the edtech narratives of the past. This means we need to look beyond “efficiency” and towards ways of working creatively with AI.

To do that, we need to find ways of working with the technology directly rather than relying on apps built and designed by non-educators. Or, we can look to novel ways of combining apps not designed for education, cobbling together our own tools which are fit-for-purpose.

One of my biggest fears of this new wave of AI is that schools will be inundated with offers for “time saving” and “workload reducing” apps that teachers could quite easily build for themselves.

By way of example, I tried to think of one of the more tedious tasks an English teacher might face. Rather than looking to lesson plans, report writing, or grading assignments, I settled on a job that all senior English teachers in Victoria must face sooner or later: finding suitably bland Section C materials.

Generic by design

For those outside of senior Victorian English, Section C is the third and final part of the Year 12 examination. More broadly, it aligns with the Areas of Study Exploring Argument and Analysing Argument. In the examination, students are presented with a written text around 700 words in length. The text – now written entirely by VCAA staff after a little plagiarism slip-up a few years back – are designed to be accessible to the 40-something-thousand students who sit the exam each year.

Unfortunately, this means that they tend towards the bland. No divisive or contentious issues, no sophisticated or nuanced arguments. Just a straightforward persuasive piece that every student has the chance to draw some analysis from. Section C is what it is.

But one of the most mind-numbing experiences I had as a senior English teacher was trying to find (or create) practice examination materials. Most previous years’ exams are available on the VCAA website, but not all. Plus, the slight changes in the 2017 study design mean that anything prior to that doesn’t really address the criteria. There are many publishers who produce Section C materials, but a keen student will quickly burn through a pack of those, and schools aren’t made of money.

Enter, ChatGPT.

Jonty Jenkins versus the Machines

The challenge for teachers in putting together Section C practice materials is this: you have to find something with the exact style of blandness – the perfect balance of predictable rhetoric devices and easy-to-follow arguments. There has to be at least one image or visual element. It might be in the form of an editorial, op-ed, letter, newsletter, website, blog post, transcript, or any other form.

The temptation is always there to use opinion pieces from online newspapers. Unfortunately, they can be paywalled (like the Herald Sun), or too academic or sophisticated for the purpose (like The Conversation or the Guardian).

Luckily, this type of generic, predictable writing is exactly where large language models excel. Rather than shopping around for suitable Section C responses, I figured it would be easy build a generator from scratch.

Here’s a short video of the whole process:

The process

There’s a PDF of three of the responses from the video at the end of this post, including images generated by copying the <image descriptions> provided by ChatGPT verbatim into Midjourney, the AI image generator.

As you can see in the video, it takes a little to-ing and fro-ing to get what you want. I’m just trying to demonstrate proof-of-concept here, so I’m not looking for perfection, but hopefully this gives you enough to go on if you’d like to try it for yourself.

Here’s a breakdown of the process:

Step one: provide the example material and a brief description

I used an actual VCE English Section C as the first prompt. To do this, I converted the 2017 exam PDF to a text file with a free online converter (to make it easier to copy/paste). Because you can’t (yet) put images into ChatGPT, I manually added brief descriptions of the images in <>. I put the whole thing into ChatGPT with the following initial prompt: The following text is an example of a style of writing called “section c”. The purpose is a simple persuasive text in which the author tries to persuade the reader to share their point of view. Image descriptions are indicated with <>. Each section c piece begins with “Background Information”. Return “OK”:

I added “return OK” because I didn’t want it to do anything else at this stage. I could have added my following prompt right at the end, but I wanted to break up the process.

Step two: Generate the first novel response

Next I used the following prompt to see what ChatGPT makes of our 2017 exam input: Write a new section c about dogs in cafes in the style of the example. Despite the Background Information referring to a response from a concerned citizen, this didn’t make its way into the actual text, so I prompted further with “And the response from a concerned citizen“.

Step three: Generating topics

Part of the frustration of finding Section C practice materials is coming up with (and locating texts for) the kind of bland topics favoured by the examination. That’s a fairly easy thing to automate with a few examples from previous years and the following prompt:

Here is a list of previous topics: primary school litter and recycling; the creation of a large watermelon tourist attraction; the rise of ebooks; the popularity of tattoos; a biodiversity conference; a review of a franchise cafe appearing in a traditional suburb; drones and the impact on agriculture; a local grocery store wanting to switch to cashless payments. Generate a list of ten possible topics for Section C style materials.

Step four: Tweaking and refining

The “cafe dogs” piece was close, but it had a few odd features it had “learned” from the example piece. For example, although it was written in the style of a letter, it featured some of the website conventions from the header of the Spire Primary piece. It’s generally pretty straightforward to address these kinds of problems.

First, I tried the following prompt: Based on the initial Spire Primary example, write a new Section C material for option 9. Include image descriptions in <> and include some form of response. Rather than a website, make this one a print magazine article.

However, as you can see in the video, it didn’t exactly work. The header links remained, even though the rest of the style had changed. The next prompt was more specific and yielded better results:

For future responses, where the material is not a website, do not include website page headings like the Spire Primary example. Follow the conventions of the form. Write a section C material on a random topic in the form of a speech transcription. Do not include a second response.

Extra tips

• ChatGPT has a tendency to become more generic/brief over time. Counteract this by “re-prompting” with the original example or with a satisfactory new example every few prompts.
• Section Cs are designed to have persuasive techniques that are easily identifiable by most students. Ask for more specifics, such as “include a personal anecdote, several authoritative quotes, an image of a relevant graph, and three emotional appeals”.
• If you plan on using this method to generate Section C materials with better quality images, make sure you instruct the model to produce good image prompts. Provide an example (e.g., from a website of Stable Diffusion tips), and ask for an image generation prompt in the main prompt.

The finished product

Here’s a PDF of the output from ChatGPT for three of the Section Cs this method generated. They’re not perfect – I didn’t follow the above tips to refine them, but that would be simple. Treat these as an example of what is possible with a methodical approach to prompting ChatGPT.

I have changed nothing from the ChatGPT output except replacing [names] with made up names, and reformatting the text. I pasted the <image descriptions> verbatim into Midjourney and included them in the materials.

Hopefully this post demonstrates that you don’t need a great deal of knowledge to get decent results from ChatGPT. I still think that there are many better uses for the technology than generating this kind of generic text, but I’m also mindful of how AI apps will be marketed into schools.

This Section C generator isn’t going to destroy the education system. Nor is it going to solve the workload problem, or the teacher shortage crisis. But it might give a handful of teachers some inspiration for playing around and experimenting with the technology. Don’t fall into the trap of paying for a PDF of Section C responses generated by an AI: make your own!

If you’ve got any questions or comments about this post, get in touch: