AI in Education - Lecturing

Since ChatGPT’s launch in 2022, being an academic involved in marking has steadily become a more painful experience. It’s quite hard sometimes pouring your energy into a module, into an assessment, and into marking, only to stumble across another report that is from a student you’ve never seen and you may never see again, who’s written a report that seems clearly written by an AI with no real understanding of your content, but in a way which you cannot prove. Instead, without a smoking gun¹ you have to accept that the student has unprovably cheated and live with the fact they have probably got an unearned first.

And each time, a small part of me chips away.

As I teach a lot of Software Engineering modules, I try to keep up with the advances in AI; A sisyphean task I would not wish upon anyone. However, this results in me often tinkering around with the latest models, the latest tools, and the latest gimmicks in the domain. As a consequence, I find myself paying for a recurring Claude subscription – Both begrudgingly and also because… well, frankly, I get the appeal.

I find AI to be a double edged sword.

On one hand, I find the industry fairly despicable. OpenAI seemingly do not care how you use their model, so long as you keep using it. xAI and Grok, led by the always awful Elon Musk, treat AI as a tool for a culture war, and will happily let you use their AI for pretty much anything despicable you can think of. I also find it quite unpleasant that these companies were quite happy raiding pretty much all of human art and literature to invent tools to sell the work they stole back to us.

On the other hand, there are a lot of cases where… unfortunately… I do understand the appeal… As a programming tool, they are utterly indispensable. Claude Code as a tool has massively increased the speed of my development process². Being able to ask Claude to find a specific feature, or where a feature could be implemented, in a codebase is very powerful, especially when trying to patch issues in a codebase you’ve never interacted with.

On the secret third other hand, each of these also have their issues – The huge proliferation of these tools is leading to the Linux kernel mailing list becoming unmaintainable, Github crashes 4-40 times a day, and AI may be making us all stupider. This isn’t even discussing the impacts of AI generated imagery, music, and more, all of which have an enormous impact on the same creatives whose work they stole to begin with.

I am conflicted.

After some discussions with my coworkers about the burnout from marking yet-another AI authored submission for an assignment, an idea occurred. This idea led to a lecture that, I’ll be honest, only left me more conflicted.

The conceit

How would students feel if, instead of submitting AI work to us, we submitted it back to them? If the roles were reversed, how would students feel if their lecturer generated AI work for them?

Instead of writing my final lecture for my module, what if I got AI to do the entire thing?

How?

To make this experiment work, I needed to get Claude to fully generate a lecture that sounds and looks comfortably like me. However, for it to be a fair test, it also needed to be automatic – The more modifications I manually make, the more unfair and inaccurate the test becomes. I needed Claude to make me a fully convincing Powerpoint that looked like I had made it, without me doing almost anything.

I decided to do this on our Computational Creativity module, a module that now naturally has to touch AI a number of times. In this module, we discuss a number of creative processes, and how these have been or could be emulated by computers. For me personally, I wanted to talk both about AI, but also about lecturing and education.

To do this, I used Claude “Cowork”. I gave Cowork access to my folder of lectures for the module, I converted my lecture template from Keynote to Powerpoint, I told Cowork the topics I wanted covered, and I let it do the work.

Prompts

I wanted to make sure there was as little human-intervention as possible, so I tried to keep the number of prompts to a minimum. I won’t post all of them, but here is the first prompt I wrote.

I want to create a new lecture - But I want you to create every single word of it. The point of the lecture is about the power of LLMs, and I want to end the lecture by demonstrating that the entire lecture was, in fact, written by you. HOWEVER, I want it to sound like I wrote it! We need to walk the line where you write it, but it sounds like I wrote it.

I want to discuss the use of LLMs in education - For example, how LLMs and AI in general could allow for customised education, courses targeted at specific people to get them from where they are to where they need to be, directed support helping people understand material when they struggle, and even the ability to create custom lectures for large-group learning. If you have any other idea of topics related to LLMs/AI in education, I’m all ears!

Perhaps the most logical first-step will be planning the sort of content we want to include, before you approach actually writing all the slides?

There was a small amount of iterating, and some hiccups along the way – The first version proposed was ~40 slides, and each had a sentence; Not my style at all.

Modifications

In order to make it fair, I wanted to make as few modifications as possible. I made 2 modifications to the slides:

1. On the title slide, I added the name and module code for the module.
2. During the reveal slide that reveals the conceit of the lecture, I removed the company name. I didn’t like that Claude has written “Claude by Anthropic” as “its” name, which I didn’t like.

Presenting

The final part of this ploy was presenting the slides. However, this is where the results start to become slightly skewed by real life.

I studied the slides ahead of the lecture to make sure I felt comfortable with the content, and presented the slides to my cohort.

The actual experience of presenting the slides was mostly fine, although presenting it felt like presenting a lecture I knew I was unhappy with. A lecture I know I would’ve liked more time to get right. Not awful, but… fine.

The biggest issue was that even with my convincing, the slides were only 16 information-packed slides. Whilst this isn’t impossible to work with, I personally do not want slides so dense.

Each slide had 3-6 bullet points, each of which was a full sentence. Whilst they all covered an interesting topic, they covered the full topic in one go. This is ok, but there’s no natural flow to the information being delivered as a result.

The content the model wrote wasn’t bad at all – I can clearly see the potential for this to be transformed into a really solid lecture with some work – but as is, presenting it was an unusual experience.

For the most part, each slide was so dense I almost had to just spend each slide fully discussing a topic[^I note that, in some ways, this does make my experience delivering the content easier, as I don’t need to be particularly careful to setup the next topic – There isn’t a connection between them anyways.], going in-depth where possible.

There were some issues, but for most part, from a presentation aspect, I feel like I did a good presentation using these slides, even if they didn’t truly feel like my voice.

At the end of the lecture, I presented a slide.

Every word of this lecture was written by an LLM.

Specifically, by Claude. I gave it a brief. I reviewed the output.

But I didn’t write it.

There were some laughs from the students as I revealed what I had done. One student laughed, got her phone out, and took a photo of the slide. For the most part, the immediate reaction of the student was exactly what I had hoped for – Amusement, intrigue, and a bit of interest.

However, I then asked the students how they felt about this?

Could they tell?

Did they feel cheated?

Did this devalue the lecture they had just watched?

Did they feel how we feel when we receive AI-generated submissions?

These were the questions that inspired the lecture, and yet… The results were not what I expected.

The students told me that they honestly did not feel cheated. The reality of lecturing means that, even if I did not write the slides, the last hour of me presenting were still my work, my words, my presentation style, my behaviour, all my quirks and my humour. Whilst some claimed they could tell something was different with the slides, they didn’t know it was AI generated, and the words from my mouth were still, in essence a creative act.

Reflection

I won’t lie, these reactions shook me a little bit. The results were not what I expected, and whilst I did assume there would be value from my side of this, I honestly thought students would be more upset by this.

This has changed my mind slightly.

I have until now held a pretty anti-AI stance when it came to lecturing. Whilst I used AI for programming, and sometimes for workshopping ideas of planning outlines, for the most part I have maintained that every word in my slides should be mine.

I don’t see this changing. Whilst I feel less stringently about this than I did a week ago, I find the process of presenting words I did not write extremely painful.

I have in the past adapted a coworker’s slides to deliver a lecture. In the process of doing this, I have always typed their slides into my own words. This was partly so I would present work that feels like it belongs with the rest of my slides – But honestly, it was also about learning the material! The process of typing up the slides myself felt like an educational experience, it familiarised me with the content I was delivering, and I always felt more confident presenting slides when I had written the slide myself.

I find myself increasingly conflicted. Students found the lecture compelling, and did not feel robbed or cheated by my active deception.

At the same time, I feel like I have robbed them of my work, and robbed myself of my voice.

Claude did an adequate job of writing slides, which overall show potential to become a good lecture but need some work. I did no work writing slides, and presented content I did not write and I feel do not sound like me.

I feel like this has the scope for misuse – A person less engaged and less honest could generate slides, not fully understand them but parrot enough of the slides to feel informed, and move on with their life. I don’t feel comfortable doing that – I put effort into my work, and want to present something I feel confident about and proud of.

In a strange way, I suppose we were all justified in our views.

Students did not feel cheated, as they see the value in the presentation itself. I feel I cheated, as I see the value in the process³.