A dark and scary cave: Teaching AI literacy in the Humanities
I began experimenting with large language models in 2020. In those days the state of the art was GPT-2 (not yet ChatGPT: that wouldn’t arrive until 2022), and interacting with it was a rather odd experience. It generated grammatically correct sentences, but their content was often imperfectly connected to the prompt: it often got things wrong, made things up, or lost track of what it was talking about in mid-paragraph. It served as a promising proof of concept, but little more.
Six years and almost two trillion dollars of global AI investment later, I find myself oddly nostalgic for GPT-2. It was a goofy little surrealist of a model: it took up only a few gigabytes of hard drive space, ran on ‘just’ 1.5 billion parameters (modern LLMs have several trillion, which is why they are so hungry for computational power), and could tell nonsense stories all day long. Its non sequiturs and leaps of logic made using it a bit like talking to a small child. I vividly remember one narrative it wrote, whose protagonist descended through a dark, scary cave to a darker, scarier cave, beyond which was an even darker, scarier cave, and so on until I got bored of telling it to continue. Perhaps he is still descending to this day.
But the child is all grown up. Ask a modern LLM to tell a story, or write an essay, and it won’t give you stream-of-consciousness nonsense. Instead, it will generate a highly-polished text in a matter of seconds. They no longer make basic errors, and scholarly studies show that readers, including academics, often overestimate their ability to identify AI-generated writing when they see it. Just the last few months have seen both the 2026 Granta Commonwealth short story prize controversy, in which some winning entries were accused of being AI-written, and Hachette’s decision to withdraw the novel Shy Girl from sale due to concerns over the author’s AI usage. Such events suggest that even major publishers and literary prize judges may no longer be able to reliably distinguish between human and LLM-generated text.
What can we do that anyone will value, now that the machine can do all this?
These developments have caused much concern among teachers and students of the humanities: if LLMs can automate the writing (and editing, and interpretation) of texts in the same way that calculators automated calculation, then what place remains for us? Many humanities students have consequently, and understandably, come to see AI technology simply as the enemy, a monster which stands ready to devour their future. There are many strong reasons to be wary of AI, including its ecological costs, its tendencies towards algorithmic bias, and the legal and ethical issues created by the aggressive scraping of the internet for training data. Such stances are often reinforced by simple aesthetic revulsion at the corporate blandness of default prose LLM styles, and the airbrushed, so-pretty-it’s-ugly quality of AI-generated art. But always alongside them, now, is the same rumbling fear: what can we do that anyone will value, now that the machine can do all this?
One common response has been to simply declare humanities courses as LLM-free zones. Over the last three years, however, I’ve been increasingly trying to incorporate the analysis of AI outputs into my English Literature teaching at Exeter. I have encouraged students to experiment with creating and critiquing AI-generated fiction, non-fiction, and criticism, and I have come to feel that this may be one of the most important skills I have to teach. Whether we like it or not, LLMs are now being deployed in many fields which have traditionally provided employment for humanities graduates, including law, publishing, journalism, translation, and administration. But the more such technologies are employed, the more vital it becomes to have people with the skills necessary to assess when such machines may safely be relied upon to do our writing and editing for us, and when they may be leading us astray.
The most hyped technology on the planet is a language machine – and language is precisely the field that we know best.
Modern LLMs no longer output obvious non sequiturs or get trapped in infinite caving loops, like GPT-2 did in 2020. Those were rookie errors. Instead their weaknesses are, to paraphrase Eddie Campbell and Alan Moore, diseases of language: subtle patterns built into their training data, exacerbated through endless rounds of training. Watching LLMs critique one another’s outputs, in contexts such as the EQ-Bench3 creative writing benchmarks, makes clear the extent to which AI is blind to its own clichés. These blind spots result in rashes of stylistic stigmata such as the ‘it’s not X, it’s Y’ phrasing beloved of modern LLMs, or the unpredictable ‘attractor basin’ effects that lead to AIs becoming obsessed with writing about women named Elara Voss or lighthouse keepers named Elias Thorne. But we, as students of the humanities,can learn to discern them, and to articulate why leaving them unchecked might degrade the quality of our thought and communication. All we need is the proper training, and a sufficient amount of practise.
This, I suspect, will be a crucial task for humanities education going forwards. At a moment when humanities departments are under tremendous pressure, the most hyped technology on the planet is a language machine – and language is precisely the field that we know best. As AI-generated content seeps into every aspect of life, we need people who combine an ear for language with an understanding of why LLMs work the way they do. We need people who can recognise AI outputs when they see them, and identify their qualities and their limitations. Because without a human in the loop, identifying the flaws that such systems are unable to perceive within themselves, the caves that they lead us into may prove to be very dark and scary indeed.
