As part of our training in Artificial Intelligence, a number of staff in the Faculty of Creative Arts and Humanities at Liverpool Hope University were given the opportunity to undertake the King’s College, London, FutureLearn course on AI in Education. It’s a subject about which I have very mixed feelings, so here are my reflections on the first week of the course.

I will be honest. I have really, really mixed feelings about AI. The first time I really came across it was when I heard folks talking about it on what was then Twitter – and the chatter was related to the hallucinated biographies it had made of them or their friends. If you asked ChatGPT about me way back then, it claimed I was a Professor in the Music Department at Huddersfield University. It got my research interests basically right, but not much more. Then, I got more involved when was Academic Integrity Officer in History at Lancaster University. Work that was produced using it was pretty easy to spot – it hallucinated footnotes, produced word salad as prose that sounded convincing until you actually sat down and thought about what it meant… that sort of stuff. But it ‘learned’ fast. There were discussions around how to get students to declare their use, RAG rating assignments, and how to grade assignments in the age of AI when the grading criteria were essentially prehistoric. Nothing was ever fully resolved, and I think the same problems still bedevil the academy, for very good reasons. On the one hand, we need to prepare students for the world of work and grown up life, and AI is now part of that whether we like it or not. But it raises significant questions for me about what it means to be a history graduate in the future. And that’s before we get to the ethical implications and environmental impact of Generative AI.

Because not long after my interest in the subject was piqued, I discovered that five of my works were pirated by LibGen, whose dataset of illegally obtained material Meta then used to train its AI. You can read more about this on The Atlantic, and check the database to see if any of your work was also included, although unfortunately it is behind a paywall – you can start a 30 day free trial if you are interested. Meta is by no means alone. The Society of Authors has reports on the Anthropic settlement and in March this year they published a book that contained only the names of those of us who signed a petition against proposed changes which would have made it easier to train AI models on copyrighted works without a licence. They note that ‘Most of the world’s leading AI companies are known to build their models by copying books without permission or payment, some even turning to shady pirate libraries to source these books’ and that ‘This has led to dozens of lawsuits from authors and publishers around the world, one of which has recently resulted in a $1.5 billion settlement between AI company Anthropic and authors, the largest copyright settlement in history.’

Then there are the biases involved, some of which have been widely reported. Others are less well known – take, for example, the LSE report that AI case summaries used to make decisions on healthcare issues downplayed medical issues faced by women, reported here in the Guardian. Attempts to fix these problems can sometimes cause unanticipated new ones, such as the racially-diverse Nazis created by Google’s AI, Gemini (again, that is a link to The Atlantic, sorry… but at least there’s that free trial I mentioned – just remember to cancel it if you don’t want the subscription!).

And the current area of concern is the environmental cost. There is an interesting podcast from BBC Radio 4’s More or Less team, considering figures for the water consumption of Generative AI, which should be available to readers in the UK and possibly elsewhere as it is part of the World Service material, but suffice it to say that the environmental impact of the huge servers themselves, and the electricity needed to power them is currently unknown and this means that we shouldn’t discount the possibility that it will be significant.

Then there are concerns about AI in learning – explicitly, I suppose, what I am trying to address by undertaking the course. It is difficult to get past the fact that the main ways I see AI being used by students is as shortcuts. I’m no fool. There will be other things going on that I know nothing about. The course did point out that AI tools that summarise lecture notes and reading materials ‘could lead to a bypassing of the cognitive processing necessary to learning’, and this is one of my worries – by encouraging the use of AI to summarise reading materials, we are undermining the in depth engagement with materials that you need in order to be a good historian. And I’m not just saying that because that’s the way I was trained and the way it has always been done. In order to build on existing knowledge in the field, we often make tiny nuances to already nuanced arguments. If all you see are the summaries, how do you tell where your research fits in?

On the other hand, there are things that I am perfectly happy to use AI for. I’m content to let it work out the best way to deal with my range anxiety regarding my new EV. It can do far more than a simple google to tell me where I should be able to get to before I need to charge, and where I need to charge at the end of my journey to ensure that I have enough power to get me around the Cotswolds (which appears to be a pretty barren area for ultra rapid EV chargers at the moment) while I’m there for a weekend.

Things in my research are even more complicated. Like transcribing thousands of documents in secretary hand – yes, I’ll be first in the queue to let a well-trained Transkribus model do the really hard, time consuming labour for me. But feed those results into an AI platform like Claude’s Projects – no thanks. It might be quicker at pulling out themes across the documents (okay, I’ve already been at this indexing a year, it’s definitely going to be), but even if I grounded the project in some documentation about my aims in the research, it’s not actually going to benefit me. It’s going to stop me making any connexions for myself, and genuinely understanding the material I work with. And what about those serendipitous finds that might not be what I set out to look for, but turn out to be the most interesting thing, or the single point that sheds a whole new light on the project, or the inspiration that sparks a follow-on project? I’d never find them. And I’ve not tried, but I suspect it would be less than competent in indexing uncorrected Transkribus transcripts – but I can do that, because I don’t need to have a perfect transcript to understand what is going on; my brain makes those corrections for me.because I don’t need to have a perfect transcript to understand what is going on; my brain makes those corrections for me.

The first activity was to think about the definitions of AI, the tools we can use and what the differences are between them. There lots of steps and short tasks to complete, and I won’t itemise each one because we would be here all day. One of the tasks was to make a song using AI. The idea was that this might be something that got the same message across in a different way to normal. In principle, I know that using songs in this way has a lot of merit – it is, after all, the basis of most of my research. But there is literally no artistry in what AI generates. There is no spark of human empathy and emotion. None of the ineffable qualities of human-ness.

So here is the power ballad I made using Suno from part of my lecture notes on the Catholic threat to Elizabeth I. Now I hold my hands up and admit that having looked at it, part of the problem here was my instruction to the AI model. I just told it to turn the notes into a song, so it took me literally and put the words to music in the genre I asked for.

But being an obedient student and actually interested in getting better, I tried again with better instructions, and this is one of its second attempts. At least this time I realised that I needed to tell it to make the words rhyme. But it gets stuff wrong, particularly when it comes to summarising historiographical debate – perhaps that’s a little unfair, as that’s probably quite difficult to do effectively in a song like this. But if music is about expressing a creative spark, then Sumo’s attempts to deal with my lecture notes are, in my opinion, unremittngly dire. I know at least 3 songwriters who could take the same set of notes and turn them into something that would be memorable, creative and have a spark of beauty.

Then we looked at the customisation of AIs so that they are based in secure and reliable information, which helps to reduce biases and hallucinations. I would be interested in using a customised, secure AI with students so that they could create their own learning materials, possibly as some form of assessment, but I would need a much stronger sense of how these things work and it’s really difficult to find the time to get to grips with these new technologies.

The final exercise in Activity 1 was quite interesting. I asked ChatGPT to consider academic integrity concerns from a JISC report on student concerns about AI, in relation to undergraduate history teaching in the UK and in line with QAA benchmarks. The response seemed to assume that the primary aim of UG history teaching is subject knowledge, in some kind of silo where it is unrelated to critical thinking – as if assessments are based primarily on factual recall. I can’t think of a single assessment in my ten years’ experience that is predicated on this basis – even the written exam (which I don’t use unless I have to) is much less about knowledge recall than interpretation. It’s possible that this says more about the student views in the report than it does about ChatGPT, but if so, that in itself is quite worrying. Still, there were some interesting suggestions as to how AI could be integrated alongside other teaching methods in what seem like fairly useful ways… I have saved a copy of the chat and will reflect on it further.

I must say, though, that even as someone who knew a bit about Gen AI before I started the course, I am finding it a bit frustrating at times. There are suggested activities that assume you know how to do things already (such as downloading the transcript from a YouTube video – which I ended up having to google how to do!); explanations that don’t fully explain the terms they use (‘Few-shot prompting:  you provide a few examples (eg previous learning materials) to guide the AI before asking it the actual question’, anyone? It might just be me but I’m not entirely sure what ‘previous learning materials’ are here, and therefore it’s quite difficult to know what sort of examples I should be giving in order to train the model to do what I want); and links to articles that are behind paywalls, so that even though I want to read them, I can’t follow it up (and yes, I am aware that this a pot calling kettle black situation, but at least the links I posted had a free trial you could use). I’m also finding it a bit uncomforable knowing that my GenAI usage has skyrocketed in order to complete the tasks, given the ethical and environmental concerns I raised above. They are covered on the course, but not until week 3. I guess using AI is pretty predictable if we want to learn about it, but it does leave me feeling slightly… unclean… guilty…