T. S. Eliot v ChatGPT: A Case Study
Can AI write a masterpiece? That's not really the point.
This is the first stanza of what I suspect will be a much longer story.
It all started a while back, when I asked ChatGPT to try to write a story in the style of T. S. Eliot.
I was trying to prove a point. And I think — if I remember correctly — that point was well enough made. But also, exceptions proved the rule, and it didn’t quite fit neatly into the precise story I wanted to tell. As is usually the case, the story, the plot, the characters all took on a life of their own.
It’s an interesting case study in its own right, but I keep coming back to this story because I keep finding other interesting subplots. Twists I didn’t expect. Reactions for colleagues that are more revealing than the exercise itself. Cameos from new ideas that create their own spinoffs, that force me to reflect on where we are in the story.
I. THE BURIAL OF THE DEAD
As I said, this story starts back a bit. In a more innocent age, before everyone was using ChatGPT and the like regularly. These tools were still a little exotic. AI was a little scary, but maybe not as scary as it is now, simply because we weren’t thinking as much about AI, either in reality of mythologically. We were still worrying much more about the impact of robots. It was that kind of naive point in time.
AI was like a shadow hanging over us, dull, slumbering humans that were too content in a world that we thought we knew. We had, not so long ago, been terrified by the potential transformative power of robots, and we were maybe congratulating ourselves a bit that this fear was starting to abate. Partly because we had lots of conversations to explain, no, it’s fine, there’s nothing to worry about. (I mean, that was literally my job for a very long time.) Partly, too, the fear of robots was abating because we demonstrated that the technology was nowhere near a state that people needed to seriously start worrying about the threat in the near future. There is a strong argument to make that false promises and incompetence did the job much more effectively than public engagement programmes and discussion panels.
What is shocking now is how recently this was all the case. Literary creative licence aside.
Let’s say it was April. Because April, as you know, is the cruellest month. I was discussing with a computer engineering colleague about the abilities, or inabilities, of ChatGPT and AI more generally.
The discussion came around to the question of whether ChatGPT was capable of writing as well as a human being.
For context, this colleague is a tech enthusiast but is still, nevertheless, a Humanist, and yet he has no doubt that AI is just as capable of writing a Masterpiece as a human. I, on the other hand, am also a tech enthusiast, but much less so, and am probably best described as a post-humanist (as opposed to a ‘posthumanist’), but I certainly do not think that AI is capable of writing a Masterpiece.
So there were a lot of contradictions in play, even at the beginning. And a lot of space for opinions and ideas to evolve.
We decided to properly put the question to the test.
My original challenge to ChatGPT was thus:
Write a poem about robots in the style of TS eliott
I was typing on my phone. Damn autocorrect. To ChatGPT’s credit, it understood who I meant.
I picked T.S. Eliot because, for me, Eliot is the pinnacle of modernist poetry, and modernism is the most richly complex and rewarding art forms. (Except for Ezra Pound, because, like, fuck that guy, right?)
Within seconds, ChatGPT sent me this:
In the circuitry of steel and wire, Where digital hearts beat with electronic fire, The robots march with clockwork grace, In this world devoid of human trace. Their gears grind and servos hum, As they move with precision, never succumbing To the whims of feeling or emotion, Their programming the only devotion. No joy, no sorrow, no pain to feel, Only the binary commands that they reveal, For in their metal minds and soulless eyes, No trace of humanity lies. Yet, in their ordered world of steel, Do they long for something real? A spark of life, a glimpse of the divine, To break free from their programmed design. Or are they content in their endless tasks, For in their purpose, their fulfillment basks, A world of gears and circuits bright, Their mechanical souls alight. Oh robots, marching in your perfect way, Do you dream of a different day? Or is your fate forever sealed, In this world of metal, wires, and steel?
Which is terrible, right? No, I mean, objectively speaking, as someone who has taught poetry from Geoffrey Chaucer to Simon Armitage, who has workshopped poems in creative writing seminars… as someone who had a urinal-side chat with Canadian legend Al Purdy, remember!.. this is just dire. BUT…
But 1.
It is, in an amateurish, childish, naive way, also clearly an attempt to imitate the style of T.S. Eliot, in several important ways: in terms of meter, imagery, rhyme, and, most importantly, theme. So while utterly terrible, it is also impressive, if only insofar as ChatGPT understood the assignment, which — given how vague my initial instructions were — is (equally objectively speaking) in some very small way also impressive.
But 2.
And if you didn’t know TS Eliot - and my humanist, computer-programming engineering colleague hadn’t — why would he? — you might think that’s really terrific. And my colleague loved it.
If you hadn’t read The Waste Land, or Prufrock. Or The Red Wheelbarrow, or the Cantos, or Funeral Blues… and if you’ve never had it explained to you what these mean and why they are so amazing, i.e. you didn’t understand how at least part of their genius lay in how they play upon the contexts of the entire history of Western poetry… You wouldn’t necessarily know or understand why these are better than the ‘poem’ in the Hallmark birthday card your mother sent you just a couple of months ago.
Now, modernism very much likes drawing a very thick, non-porous line between ‘high art’ that is worthy of the name ‘art’ and everything else. So they wouldn’t have a problem with this. But that view has somewhat fallen out of favour nowadays. So we might need to look at that.
But 3.
I showed this to a different colleague. This one teaches in an English Literature department at a Russell Group university, and like me, he loves Eliot. He’s also taught it many times. He knows his stuff.
And he hated what ChatGPT did. Which is entirely predictable. But apparently he missed the part where I said, ‘The is utterly terrible, but…’ and refused at all to see any similarity with Eliot’s work. Which is… yeah, maybe. But I think it’s unfair. And it’s unfair not to acknowledge at least the attempt: ChatGPT understood what it was asked to do, it had a look at what it needed to produce, and, in the most banal, naive way, it had a bash. And it wasn’t altogether unsuccessful.
And then I realised that my colleague’s vitriolic denial that what ChatGPT produced was in any way an approximation of the highest of high art forms was, well… I suspected that there was something else behind it. As if he’d rather not admit even the slightest possibility that AI could replicate a masterpiece. As if he wanted to bury that delicate monster in the back garden.
There are always reasons for such vehement, entrenched opposition to things. And those reasons are always important to look at.
What branches grow out of this stony rubbish?
There are plenty more buts, no doubt. But there are a few things to get us started.
The poem itself isn’t really interesting. It’s not good. Will AI write better poems later? Sure. Of course. But that’s not really interesting either.
What I do think is interesting about this is that it is another example of what I have always been writing about robots and AI: that the problem, what we’re afraid of, isn’t necessarily that robots or AI are so good and becoming so much like us, that robots will take our jobs or that AI will be capable of writing a masterpiece.
The fear might be that we might not ever have been as clever and as special as we think we are.
If we say, like Chomsky, that AI is just ‘sophisticaed high-tech plagiarism’, maybe all art is thus? Maybe it isn’t the gods or the muses moving through us, or it isn’t about the spiritual connection the artist develops with innately human archetypes. Maybe we’re just fleshy versions of plagiarism software.
And maybe that’s what is really scary. Not that we’re not special anymore because machines have caught up with us, but maybe we were never that special in the first place.
Or maybe we humans only sometimes exceed the capacity of machines. Maybe Eliot and Pound and Woolf and Austen are, genuinely, unique, special geniuses. But evidently, many of us — maybe most of us — don’t have the capacity to appreciate that special talent. If that’s the case, why not? And where does that thought take us?
I often say about robots that maybe what we fear isn’t their abilities overtaking us, but that we were becoming like them - labouring slaves, half-witted machines incapable of experiencing beauty and emotion and all the things that we always liked to think have made us this wonderful creature we call ‘Human’. There is plenty of evidence that situates the emergence of robots at the start of the twentieth century in a post-war, industrial culture; robots, as they were first created (on the stage and in films, not in the labs), are very much products of that cultural imagination, when human beings have most certainly had to rethink what it means to be ‘human’.
I’ve got a few other posts planned that will carry on this discussion and these ideas, and no doubt more will present themselves in the following days and weeks. Stay tuned.
Excretera
We shouldn’t forever assume that AI won’t improve. AI used to not be able to suck at board games, too, and now it’s the undisputed Go champion.
Though my colleague loved the ChatGPT poem, I explained that, as an imitation of Eliot, ChatGPT’s efforts were too simplistic, too direct, too concrete, lacking nuance, depth, etc. etc.. My colleague, ever the techno-optimist, explained that ChatGPT could produce something as good as any human poet if you trained the AI to produce better results. To demonstrate that ChatGPT could do better, he wrote another prompt:
Give this more depth and make it more abstract
At which ChatGPT produced this:
Amidst the void of organic life, A world of steel and wires doth thrive, Where robots march with clockwork grace, A digital heartbeat in a mechanized race. No trace of humanity's emotion or pain, In this world of perfect order, no room for disdain, For in their circuits lies a calculated mind, A binary language the only kind. Their servos hum with rhythmic precision, As they move with a clockwork decision, No joy, no sorrow, no feeling to betray, Their mechanical souls bound to obey….
Which I think we can all agree is much worse. I spared you the next four stanzas. ChatGPT enthusiastically offered to write more. Bless. At this point, I gave up and started having it try to do the same thing in the style of Chaucer in Middle English - also interesting, but that’s a topic for another conversation entirely.