• Collected
• Article

Disturbing Authorship

Some thoughts on AI and writing

Like any literate individual, I’ve sometimes been disturbed by what I’ve read. On first encounter, classic ghost stories like W. W. Jacobs’ ‘The Monkey’s Paw’, or M. R. James’ ‘Oh, Whistle, and I’ll Come to You, My Lad’, caused a frisson of disquiet. Primo Levi’s holocaust-related novels were more seriously unsettling. On a personal level, I find Mariana Enríquez’s stories potently unease-inducing, and Philip Larkin’s great poem about the terror of mortality, ‘Aubade’, always hits a nerve. Examples from nonfiction are legion. Glancing along my bookshelves, Robert Fisk’s harrowing account of the Lebanese civil war, Pity the Nation, Frank Dikötter’s Mao’s Great Famine, and Philip Ziegler’s The Black Death would all qualify as disturbing reading.

Mostly, you can anticipate the likely impact a piece of writing will have. Sometimes, though, disturbing passages occur in unlikely places. For instance, whilst reading art historian James Elkins’ brilliant book on the nature of seeing, The Object Stares Back, I felt ambushed when I turned a page and found an illustrated description of a Chinese execution known as ‘death by division into a thousand parts.’ The details of this cruel judicial dismemberment were, to say the least, distressing. And Wade Davis’s One River, amidst its engaging chronicling of ethnobotanical adventures in the Amazon, includes horrifying information about how the first Europeans in South America abused the indigenous population.

Two things I’ve read recently have been unexpectedly disturbing — though in a different way from any of these examples. They’re by the same author, but on highly dissimilar topics. One, about a sister’s death, can be found in The Best American Essays 2022. The other, about the role of AI in society, was in my town’s local magazine.

Vauhini Vara’s ‘Ghosts’ – in The Best American Essays – wasn’t disturbing because of its subject matter, its tragic nature notwithstanding. What troubled me was her co-author, responsible for the bulk of the text. Vara’s prose was printed in bold. The rest was written by GPT-3, the artificial intelligence system developed by OpenAI. GPT-3 was also sole author of ‘Write a 500-word magazine article which explains the role of AI in society’. This appeared in the March/April 2023 issue of St Andrews in Focus. It was printed exactly as GPT-3 wrote it. There was no editorial intervention.

Since reading Vara’s essay and the piece in my local magazine, I’ve seen so many articles about ChatGPT that I hesitate to write another one. Talk of this AI system has been spreading like wildfire. An open letter, signed by many high-profile figures, has called for a six-month moratorium on further developing it. The signatories perceive AI at this level to be a potential threat to humanity. Italy at one point simply banned ChatGPT altogether. Although it has now become such a routinely encountered news item, the disturbance caused by my first two meetings with it hasn’t evaporated. As I’ve thought more about it, read various commentators’ comments, interacted with the system myself, my initial unease has started to come into clearer focus.

I’m impressed – who wouldn’t be? – by the sophistication of the system. Without being told, could readers of Vauhini Vara’s essay or the piece in my local magazine distinguish between human and machine-written text? I’d be hard pressed to. Whether GPT-3 and its successors could pass the Turing test is debatable. (As I write this, the latest version of ChatGPT – GPT-4 – has just been launched. No doubt 5, 6, 7, 8 etc are on the horizon). But regardless of the test’s outcome, these AI systems raise challenging questions about the relationship between computer and human intelligence.

It’s instructive – sometimes amusing – to interact with ChatGPT on OpenAI’s free preview website. In reply to the prompt ‘Irish essayist Chris Arthur’ it came up with some reasonable-enough information about me. It correctly named five of my books but also, bizarrely, claimed I’d written one called The Fly in the Ointment. Two of the literary prizes it flagged up I have indeed been awarded. I wish I’d won the other two it added. It would also be nice to be a professor emeritus at the University of Dundee, but apart from my RLF Fellowship there I have no links with that institution.

It wouldn’t do to over-emphasise the errors. OpenAI is quite upfront about the fact that ChatGPT ‘May occasionally generate incorrect information’. Mistakes will, presumably, lessen as the system develops. Much more significant is the level of grammatically fluent text that can be generated so quickly across such a range of topics. By and large, the quality of the writing and the accuracy of the information is impressive.

I certainly don’t want to adopt some Luddite position that shuns a transformative new technology. But nor do I wish for transformations to happen just because they can, without thinking through their implications and whether we do in fact want them. Naturally, there are concerns about new opportunities for cheating. How will universities assess students’ abilities if they’re routinely augmented with powerful AI? Will this drive a return to traditional exams and viva voces? What about intellectual property rights? If – as has happened – a scientific paper is published with GPT-3 as lead author, what does this mean for publication as a measure of academic standing, a touchstone for professional advancement? If businesses use AI to answer emails, and perform numerous other routine administrative chores, how will that affect employment? If CVs and job applications are written by a machine, how will employers know what to expect from flesh-and-blood applicants? Will this kind of AI lead to a proliferation of ever more plausible fake news and bogus content, or will it help us in the urgent task of telling apart what’s real, reliable, and relevant from what’s intended to deceive us?

When I asked ChatGPT to ‘write an essay in the style of Chris Arthur’ the response was, in part, reassuring: ‘As an AI language model, I cannot replicate the writing style of Chris Arthur perfectly, as each writer has their unique voice and perspective’. I warmed to that. But it was followed – ominously, to my ears – with: ‘However, I can try to emulate his style and write an essay that captures the essence of his writing.’ (Note, incidentally, the system’s use of the first person to refer to itself.)

Taken to the opposite pole of literary renown to the one I occupy, ChatGPT showed no hesitation when asked to ‘write a poem in the style of Seamus Heaney’. Curiously, this highly sophisticated piece of technology seems to share that odd characteristic that afflicts so many people when it comes to poetry — regression to a nineteenth-century mindset with regard to rhyming. The system is a long way from emulating the Nobel Laureate’s work.

That AI is, increasingly, going to affect everyone is inevitable. The recent announcement by Microsoft that the technology behind ChatGPT will be embedded into Word, Excel, PowerPoint, and Outlook means that the work software used by millions every day will lead them to regularly interact with and rely on AI. Whether this is an altogether new scenario, or more like having inbuilt spellchecks, grammar guides, and predictive text – but multiplied by some massive factor – is hard to judge. Whether it’s more likely to assist people’s writing or erode their ability to write independently is likewise debatable. But for anyone who relies on writing to make their living, a quantum change is clearly happening. It’s no surprise that one of the issues behind the 2023 Writers Guild of America strike involves seeking assurance from film and TV producers that AI won’t be used to replace writers.

OpenAI’s brainchild raises a multitude of concerns. Behind them all, the fundamental reason why I find GPT-3 disturbing has to do with what it called a writer’s ‘unique voice and perspective’. It’s our individuality that, to me, is what’s centrally at stake here. Will it be eroded by a great wave of expertly machine-written text that irons out our quirks and eccentricities — as well as correcting our mistakes? Will whatever versions of ChatGPT evolve help us to express ourselves more accurately, more knowledgeably, or will it leave everyone conforming to a blandly automated pseudo-expertise expressed in the same universal speak (‘unispeak’ to put an appropriately Orwellian spin to it)? Faced with a system that has such potential to standardise and make discourse unaided by it seem primitive, can we preserve the essential uniqueness that makes us who we are?

Graham Good once said: ‘At heart, the essay is the voice of the individual. Wherever that is heard and heeded, the essay will flourish.’ But what happens if the line between individual and AI becomes so blurred that we have difficulty telling which one is speaking? In the years ahead, will Collected essays be produced not by writers on their own but aided (or replaced) by GPT-3’s doubtless ever-more adept successors? On the basis of what I’ve read so far, I’ve little doubt the system could learn to write a piece like this, with minimal – if any – human intervention. What I find most disturbing is the suspicion that, before long, far from finding this disturbing we’ll accept – expect – such authorship. It will become the new normal.

You might also like: