I asked ChatGPT to provide ten possible completions to this line, taken from Gerard Manley Hopkins's poem "God's Grandeur"
It will flame out like ...
Here is what it came up with:
a comet streaking through the night.
a candle in the wind.
a phoenix rising from the ashes.
the last rays of a dying sun.
fireworks bursting in the sky.
a shooting star's brief flight.
the embers of a forgotten fire.
lightning splitting the heavens.
a spark in the dark, then vanish.
the final flare of a burning match.
These phrases are all well-known, and not exactly unexpected; they are bland and literal. Compare that with the completion in Hopkins's poem. If you don't already know the poem this will be completely unexpected:
It will flame out like shining from shook foil.
Why is it that chatGPT can't produce stunningly original completions like that? The reason lies in statistics. In generating continuations, it uses probability, and a random number generator that is tuned (like loaded dice) to be heavily biased towards more likely (and hence less surprising) continuations. Also, if it chose an unlikely continuation, it probably won't make sense at all. There are vastly more nonsensical unexpected continuations than ones that make sense.
What can we learn from this and apply to our own poems? A lot of our writing can be automatic, sometimes creative, but sometimes commonplace.
So next time you write a poem, go through the first draft and examine each word or phrase and see if you can make a better one, instead of the expected word, make it unexpected. Make it a metaphor, rather than literal.
Here's an example from one of my poems, describing the grid system of roads in
Washington DC:
Geometric city. Anonymous streetsLabelled with letters and numbers????? the shops into blocks.
Here are chatGPT's 10 boring suggestions:
Neatly dividing
Methodically carving
Carefully segmenting
Precisely arranging
Artfully separating
Rigorously partitioning
Strategically placing
Systematically organizing
Elegantly splitting
Thoughtfully grouping
Here's mine: Cheese-wire.
I'll bet you didn't see that coming! It was based not on poems I'd seen before but
lived experience; how as a kid I was fascinated to see cheese cut up with a cheese-wire.
Iain Strachan
Image from Wikimedia Commons
Well done Iain! The repetitive algorithms make AI redundantly brilliant don’t they? My girlfriend put one of my recent poems through the Chat GPT interpreter and it came out frighteningly close to what I had intended. For interpretation yes… for unique innovation… no way!! Thanks for the wonderful article my friend! I remember when you challenged me some time ago to come up with phrases to compare with AI for that cheese grid poem… such a fine work of originally superb effort… you are indeed scientist AND poet sir!
ReplyDeleteYes, I have noticed with my own poems, that it is frighteningly good at apparently understanding what I meant. But after a few you start to see how it works - in particular how it only goes line by line and doesn't consider the message of the whole poem. This experience made me sensitive to spotting that some people on the page make comments on peoples poems that were obviously AI - the favourite words come up over and over, like "transformative", "resilient", "in summary". To do that makes you appear very clever and intellectual with zero intellectual input.
DeleteGreat blog, Iain. Very informative and to the point. I try to bear these techniques in mind when writing, and not to be commonplace, but it’s difficult!
ReplyDeleteThanks for taking the time and effort to inform us 😎💯
Thank you Iain. This is great information as AI is infiltrating sites at times. Brilliant explanation. Thank you 🙏🏼 Universal Peace & Love 🪷
ReplyDeleteIn practice Iain is right. In theory he isn't.
ReplyDelete"Why AI can't be creative" is a title with a presupposition "AI can't be creative", and the presupposition isn't true. In reality, AI can be creative, but the practical realities of systems like ChatGPT mean that this will hardly ever happen. It possible, in the same way that (as Douglas Adam's said) it is possible for all the molecules of all your clothes to randomly move three feet to the left, leaving you naked. Possible, but vanishingly unlikely.
There are two parts to these AI systems. One of them defines what is possible. Hopkins' lovely continuation is possible, but vanishingly unlikely.The other defines a strategy for exploring the possibilities. In practice, OpenAI only lets you explore the top few (boring, impersonal, generalized) possibilities, so you aren't in practice going to get anything interesting, personal or surprising..
Why am I being pedantic about this? Firstly, Iain and I have been competing on pedantry since 1971. Secondly, I work for a company that uses AI models, and needs to give legal guarantees, We have spent the last two years understanding this, and our conclusion is that the best we can do is to make bad things very unlikely, then still carefully check the output.
There is another aspect to the "vanishingly unlikely" answer. In principle you could force chatGPT to favour unlikely answers over likely ones, once in a while. I like to think of it as a roulette wheel with trillions of slots - the likely (boring) ones will have billions of slots, the Hopkins example might get one. But suppose you arbitrarily picked some of the ones with only a few slots and gave them a billion, so it became more likely to be chosen.
DeleteThe problem there is that the "space" of unlikely continuations is much vaster than the "space of likely ones, and vanishingly small numbers of those will make any sense. E.G. from Prufrock:
I have measured out my life in ????
Here's what chatGPT gives:
I have measured out my life in whispered dreams.
I have measured out my life in fleeting moments.
I have measured out my life in silent tears.
I have measured out my life in shadowed fears.
I have measured out my life in stolen glances.
I have measured out my life in endless waits.
I have measured out my life in moonlit nights.
I have measured out my life in gentle sighs.
I have measured out my life in passing days.
I have measured out my life in broken dreams.
boring boring boring - a billion slots each.
but Eliot gives: COFFEE SPOONS. (one or two slots on the roulette wheel)?
but think of the enormous number of equally unlikely answers that just don't work:
Brontosauruses
Purple mist
Boeing 747s
Cheese
... etc
So the chance, even if you forced it to pick unlikely continuations, of getting one that works like "Coffee Spoons" is again vanishingly unlikely. The "fix" will just produce meaningless nonsense. Or at least is overwhelmingly likely to.
Eliots Coffee Spoons doesn't come from the statistical properties of pre-existing poems - it comes from life's experience of the shallow intellectual life in Boston, or continual tea parties, and pretentious talk about Michelangelo. It was his boredom and disillusionment with life that produced this incredible metaphor.
Better call a truce on the pedantry war now, or we'll start doing geek-speak with terms like "the curse of dimension"!