Deconstructing an AI Sonnet
(Rusty Niall is now hosted on Buttondown instead of Substack)
About a week before Open AI stunned the world, and by the world I mean tech vloggers, with its uncanny AI videos created from text prompts, I had my own little moment with Google's latest Language Learning Model, Gemini (I will henceforth refer to these specific, text-producing AIs as Language Learning Models or LLMs for short as they aren't really Artificial Intelligence).
My barometer for literary competency from the 2023 crop of LLMs such as Bard or ChatGPT was in gauging how well they did when requested to write a sonnet. The LLM would then churn out some doggerel, but that doggerel was never a sonnet.
I should add here that there is no absolute definition of what a sonnet should be. There are sonnets that follow rhyme schemes and there are unrhymned sonnets. There are sixteen-line sonnets and there are the curtal sonnets of Gerald Manley Hopkins that reduce the sonnet to 10 1/2 lines with its two segments maintaining fidelity to the 8/6 ratio of the traditional sonnet. Many sonnets are in iambic pentameter but others feature lines of varying length. In his 101 Sonnets anthology, Don Paterson settles on the definition of a fourteen-line poem. In his introduction, he gives an even more forgiving definition of a sonnet as a "wee square poem".
I tend to be a stickler for the old iambic pentameter though. There's something about those fourteen lines of i.p. that make a poem sound like a sonnet when read aloud, especially in how that reading takes about a minute. So, in the past, when I asked ChatGPT or Bard to write me a sonnet, they would often serve me up fourteen lines or jaunty, ballad-like iambic tetrameter.
Being that the rules for writing a sonnet are not carved in stone, I would often say "Thank you, could you try again but this time write it in iambic pentameter?" the LLM would say "Sure!" and then serve me another fourteen lines of jaunty tetrameter. I would then remind the LLM that it still wasn't in iambic pentameter and the LLM would give that same self-flagellating reply, "Yes, you are right, that wasn't in iambic pentameter. I apologise, I am just a Language Learning Model." It would then go on to repeat the same mistake, despite this seeming recognition of its error.
This is an important point to recognise. LLMs and other programs that we mislabel as AIs do not learn from their experiences with us. They work from massive repositories of information and serve up combinations of that information that they "think" best answer our request. So, I could have kept on insisting on my definition of i.p. but it would not have made a difference. The LLM would not have learned a thing from our exchange as that isn't the way in which they collect information.
That all changed this week when I had my own little watershed moment with Gemini, Google's rebrand of Bard. I opened it up and asked it to write me a sonnet and it finally happened, fourteen lines of iambic pentameter that followed a Shakespearean sonnet rhyme scheme (ababcdcdeefefgg). I was genuinely shocked, especially in how it was able to knock the thing out in about two seconds. I can knock out some artless doggerel too, but not in two seconds.
I think that it is important to meditate on that moment of shock when an elaborate cultural object is created by something that isn't human. Most of the reporting around AI tries to reproduce this sense of shock. AI shock stories tend to spread faster and bring in more clicks than the reflections that follow on from the initial impact. These shocks produce immediate reactions about the threat that AIs pose to creators before the AI cultists jump in with their Randian takes on how we all have to move with the times and it is a good thing when useless humans are replaced with brilliant robots.
But I'm getting ahead of myself, here's one Gemini's poems that led to me having my own AI sonnet shock. The prompt, leading on from last week's essay and YouTube talk, was the Imagination:
A canvas waits, unprimed and blank and bare,
A brush untried, with colours yet unseen.
From slumbering depths, a flicker, quick and keen,
The spark of thought, imagination's flare.
It dips the brush, and on the surface paints
A world unseen, where logic's hold is frail,
Where dragons dance and unicorns unveil
Their hidden realms, where reason's light restrains.
It builds anew, with bricks of pure desire,
Castles of cloud, and lovers bold and true,
It paints the rosebuds with a richer hue,
And fans the dying embers of a fire.
This boundless realm, where dreams and truths convene,
A boundless power, both savage and serene.
The first surprise for me is that this poem combines elements of the Italian/Petrarchan and Shakespearean rhyme scheme. It uses the envelope rhyme (abba) of an Italian sonnet but also uses more rhyming sounds across the opening 8 lines (octave) and uses the quatrain-couplet sestet (final six lines) structure of a Shakespearean sonnet. This is just the kind of hyrbid sonnet that the Romantics were fond of when they tinkered with poetic forms so in that sense it is still a traditional one.
The poem handles the rhyme well enough and scans well too, something I haven't picked up from my LLM sonnet commissions with ChatGPT, Bing and Bard so far.
At the same time, there are plenty of amateurish hallmarks – the two "and"s in the first line aren't particularly elegant. Removing the first "and" after "unprimed" might have lopped off a syllable but a caesura (a kind of implied space) would have held in its space and kept the meter ticking over. Many of the sonnets that Gemini wrote for me were littered with inversions, or "Yoda speak" as I like to tell my students. This one isn't as frequent an offender but there are a few such as "a brush untried" rather than "an untried brush" or "lovers bold and true" rather than "bold and true lovers". Inversions are common in many classic sonnets but contemporary and modern sonneteers tend to avoid inversions in order to keep a more natural, speaking tone to the language.
Here, the reasonable argument could be that Gemini shouldn't have to create sonnets with the skill of an accomplished sonneteer. An amateurish sonnet is humanlike enough. And I would agree. However, in the same way that AI images and videos are at the point of being passable until you start counting the number of fingers on each hand, the AI sonnets seem full of their own "AI hand" moments.
While LLMs are notorious for their essays that have a tendency for nonsense and embellishments while maintaining an authoritative and critical writing style that often convinces nonspecialists without immediate recall of the facts, we don't necessarily expect the same rigour from a poem. At the same time, with a typical sonnet we expect the arguments and images to cohere, for a kind of sense to hold the sonnet together even if it is abstract or surreal.
This sonnet has that "AI hands" quality in the images it clubs together, without much coherence in how they follow from each other. The blank canvas at the beginning of the poem is a reliable enough staple of this fare, the "brush untried" overplays the previous point, while "colours yet unseen" is really beyond the pale.
In what way are the colours unseen? There are two ways we can interpret it. In the first, we as readers don't know what particular palette the artist wants to use. Okay. That's pretty mundane, nobody is on tenterhooks about this kind of thing. The second reading might imply that the artist is going to use colours that have never been seen by anything. Maybe there is no biological eye or camera technology that can discern the colours that the artist is about to paint with? Wow! What a potentially interesting idea. Except the idea is abandoned immediately. The word "unseen" is used a few lines later in reference to the world that is about to be created, but this repetition doesn't seem to serve any purpose and simply comes across as clumsy.
Next, imagination is finally invoked as a flare, which serves in immediate comparison to the cliche of "The spark of thought". There's something here, if a thought is a spark, then imagination is a flare, something that can suddenly light up a night sky. Again, we might be thinking about how this image can be developed but the next line states "It dips the brush.." intimating that imagination is now the artist rather than a faculty of the artist.
The sonnet progresses to detail the world that imagination is painting. It is described as "an unseen world, where logic's grip is frail" and the features of this world include the familiar trope of dancing dragons (the scraper must have swallowed a few George RR Martin titles) followed by the truly absurd sight of unicorns unveiling hidden worlds.
How can a unicorn unveil something? Of all the animal kingdom's variations of the hand, a horse's hooves seem the least practical for the act of removing a veil. Maybe the unicorn pokes its horn through the veil, dragging it away from the hidden world before shaking it off to the floor? Again, the sonnet doesn't hang around to clarify or develop the image and instead tells us that these hidden realms are where "reason's light restrains". So, according to an earlier line, the grip of logic is frail but the light of reason keeps all things in their place. I guess the grip of logic is frail in that respect.
Imagination now "builds anew with with bricks of pure desire,/Castles of cloud, and lovers bold and true,". It appears that imagination has got bored of the painting and has raided the lego box instead. It is not clear whether the cloud castles and lovers are what the bricks are being used to build or are being used as building materials. The punctuation appears to imply the latter. Does the "bold and true" quality of the lovers help them to stack up better? I'm sure the next line will clarify things ...
It paints the rosebuds with a richer hue,
And fans the dying embers of a fire.
Oh, it's gone back to painting. Also, the fire metaphor introduced by sparks and flares is back for no particular reason. I actually think these two lines are among the better lines in this poem, but they don't really follow logically from what came before.
Onto the final couplet:
This boundless realm, where dreams and truths convene,
A boundless power, both savage and serene.
The repetition of "boundless" brings a dull thud to what is meant to be the sonnet's crowning moment. The use of the verb "convene" is particularly uninspiring. It makes the coming together of "dreams and truths" sound like something on the itinerary of a work conference. The power of imagination is then spoken of as boundless, savage and serene. Well, it can't be that boundless if it's already been restrained by the light of reason. And while I'll buy "serene", I've not seen much evidence of its savagery thus far. Here the poem hints at a darkness that has not been in evidence at any point.
The lack of any darkness or savagery is probably down to the LLM following its own design to not produce answers that might potentially disturb the user. But I think it also points us to the fact that the LLM has no unconscious mind, no sudden unexpected imagery that arrives unbidden when the artist is occupying a state of flow.
The sonnet reads like the LLM has summoned a series of images and phrases and then arranged them in a way that follows the rules of the form. It hasn't thought about how these images and phrases hold together, how the poem coheres. This isn't always the case with a poem – Modernist poets played around with disparate fragments and Postmodern Language poets waged war against the need for coherence itself. But this is obviously not presented as either, with its inversions, archaic language, iambics and rhyme scheme.
There is also the sense of uncanniness. Here I think of the use of de-aged or deepfake reconstructions of known actors in movies or TV shows, like the young Mark Hamill that pops up in some episodes of The Mandalorian. While digital human faces have come on leaps and bounds, and can be barely noticeable when not front and centre in a scene, there remains that sense of things not being quite right. The more convincing the model gets, the more pronounced little unconvincing details become. The uncanniness of this poem comes from how the sloppiness of the imagery and coherence of argument don't chime with its tighter rhyme and metre. An amateurish sonnet is normally amateurish across the board, with some metrical and rhyming anomalies to match up with the cliches and inversions.
It also makes me wonder how the LLM learned to write a sonnet. Was it fed the works of Wyatt, Spenser, Shakespeare, Keats, Barrett Browning and other canonical poets or did it trawl enthusiast websites and vanity publishers?
When I used to write sonnets I often felt the the poem was completing itself. The form gradually limits the options in terms of the words that you can use at the end of each line. The possibilities seem to shrink and an eerie feeling of determinism can take over. In this sense, it seems that the sonnet is so finely-honed a construction that maybe a machine could take over the job?
That would be the case for an okay sonnet but every great sonnet seems to find a means of escape. Shakespeare's best sonnets often lurch in a different direction in the final couplet, confounding the familiarity that might have crept in, such as the ways in which the sustained, vindictive rant of Sonnet 34 switches in the final couplet to forgiveness, albeit a forgiveness that demands the remorseful beloved's tears. While the AI sonnet might fail at holding disparate images together, the human sonnet seems to forge the contradictions and absurdities of the human condition into a satisfying whole.
Similarly, the best sonnets often break at least one rule, be they the unrhymed sonnets of Terrance Hayes or the irregular rhymes of Shelley's Ozymandias. Again, it is the force of the poet's vision that pushes against these rules and the seeming mathematical determinism of the sonnet form.
I don't think there is much for the artist to fear within AI art itself, though that is not the threat. The threat is from other humans that have fallen for the hype and decide that the work or artists is now redundant. The bean counters and strategists that cannot get beyond that initial shock to count the fingers on the generated hand or ponder the logic of unicorn unveilings.
Thanks for reading this
I had a bit of a moment last week. My wife took our kids out shopping and I had the flat to myself. I had plenty to get on with, I had two lectures to give the following Monday.
So, I started procrastinating. Luckily, I saw the procrastination coming and steered it down the "reorganise drawers" path rather than the "doomscroll" path. Partway through reorganising my drawers, I became convinced that I had lost my mini tripod and started to turn the flat upside down looking for it. I then found it in the very drawer I thought it had gone missing from. Then, after tidying up the aftereffects of my lost tripod frenzy, I didn't put the tripod back in the drawer, I set it up with a camera and mic.
You see, in the midst of that frenzy, I started to think about what I would do if I found the tripod. In the midst of that thought I made a promise to the universe that if I found the tripod, I would make a video and post it to YouTube. That video would fill the niche that my podcast used to fill. The cool thing was that I went from idea-to-execution-to-editing over a matter of hours and posted the video the next day.
I didn't start the day thinking that I would do this. A lot of my previous videos had been the result of months of planning, scripting and rehearsing. This one was conceived and filmed in the space of an afternoon while my wife and kids were out shopping.
I really like doing things this way. In most cases with art and creativity, especially in arts that aren't a specialism, it's better to churn out work and gradually get better than it is to try to attain perfection from the first try.
I'll be making another video to go with this essay this weekend. For now you can watch the last one here. The next one, which will feature me talking around the ideas explored in this essay, should hopefully pop up over the weekend, family and tripods permitting.
My weekly essays are given freely, there's no pressure to give back but if you'd like to, any of these would help:
Subscribe – A new subscriber always gives me a warm feeling. There's no better indication that you're creating something valuable than seeing your audience grow.
Spread the word – Much like the above, if you enjoyed this or anything else on Rusty Niall, please consider telling someone else about it.
Upgrade to a paid plan of your choosing – If you feel that my work is worth the price of a monthly cup of coffee and it's within your means, you can click the upgrade button and choose the monthly amount you'd like to pledge. It's completely up to you whether it buys a styrofoam cup of Mellow Birds from a burger van on a trading estate or a vodka martini from a Soho members bar. It's all good, it's all up to you and you are free to adjust or cancel whenever you wish.
Say hello! – Comments are always open, I'd love to hear what you think.