Can AI engines write poetry?
Just taking a break from compiling my updated ‘Guide to getting your poetry published’ to share this.
One of the things I asked poetry magazine editors this time around was what they thought about poems being either composed entirely by AI, or ‘touched up’ by it. Unsurprisingly, the answer was an across the board NO! Some editors were honest enough to say they couldn’t be 100% sure they would spot it. Anecdotally, I’ve heard that passing a poem by Chat GPT before submitting it to, say, a competition, is standard procedure for Generation Alpha.
So, my conclusion is that it’s already happening, and it’s going to be very hard to hold back the floodgates.
In the interests of research I’ve been doing some investigating. Just how easy would it be to pass off an AI-generated poem as one’s own work?
So on Chat GPT I offered the prompt: “Can you write a 14 line poem with a science fiction theme in the style of Caroline Bird?” In a few seconds this is what it produced:
The British Secret Service issues me a sensible coat
and a name that tastes like static.
I learn to listen through walls politely,
to sip tea while history clears its throat.
My handler says don’t fall in love—
everything leaks eventually.
The dead drop is a park bench feeding pigeons secrets.
I memorise faces the way others memorise prayers.
At night, the truth sleeps beside me, fully clothed.
London keeps blinking like it knows.
I file my feelings under miscellaneous,
and burn the rest with the kettle on.
Once I’d tried the same exercise with four or five other poets a few commonalities started to leap out, such as heavy use of simile, often one ‘quirky’ idea per line, little or no through-references, sonic echoes or layered meanings. The poems kind of looked like poems but there wasn’t much below the surface, Which ties in with the fact that AI is only privy to what’s already been written. Or as Rachel Piercey, editor of Tyger Tyger said to me: “Each new poem works on its own terms – a highly individual, poem-by-poem house style that AI could never comprehend, because it has never been done before and AI can only look back.”
The twist in the tail is that I then ran these AI-generated poems by Originality.ai, an AI engine trained to spot AI in texts. It’s not fool-proof as is has been trained primarily on non-fiction texts, but the result was pretty comclusive. (See the header image to this post). Busted!
Gotcha!
Hi Robin
Interesting. Two years ago I submitted a poem to ChatGPT and received an enthusiastic appraisal (of course!). However it didn’t spot two obvious quotes from Housman and Marvell so has obvious gaps in its long-term memory.
Ha ha! It’ll learn!
To be fair to AI, that poem is not the worst I’ve ever read.
However, I read some AI generated haibun on a site somewhere (sorry, I’m hopeless at notes and references) and found exactly the same as you did, they look like poems but there’s nothing under the surface.
That, I fear, is probably the way forward in a world of influencers and airbrushed photos.
Could be, but maybe also a surge of interest in material created by an actual human, with the two living side by side. In a similar way to how ebooks sit alongside printed ones, when in the early days there was talk of the death of print. We can but hope!
I feel that the two things are slightly different. eBooks are merely a different way of distributing words. AI is a way of mass producing words without having to put anything human with them. People will still write, but how many people will become “poets” thanks to AI simply to see their name in print?
I’m OK, I can still write about history and numismatics, but as AI scans my work, how long before they steal all the best bits and replace me?
So what happens if you’re not, for example, a “deep image” poet and want to introduce a few quirky or jerky or simply surprising images (literal or metaphorical)? I don’t know about legitimacy, but is this an effective way to nudge your stuff into what editors say they’re looking for? Might be a fun (and naughty) experiment. (Yet for gods’ sakes, I certainly don’t see much congruity between “what we’re looking for” and what’s printed in a mag. How about “what names should you drop?”)
I think there’s a place for AI in research, play, experimentation or whatever, but I wonder if it could lead too easily to the poet losing sight of his/her own voice. I’m sure you could send an AI-tinted poem to many journals and they wouldn’t spot the AI influence. And adopting a different kind of name in order to tick certain boxes and thereby put you on the fast track to an acceptance is a similar ‘experiment’ that’s been tried to great effect. Personally I wouldn’t take the risk, but I can see how some might!