Neural network poetry is so bad we think it’s written by humans
Can a machine incapable of experiencing emotion write poetry that stirs the soul?
Source: Matt Reynolds
A neural network trained on thousands of lines of poetry has tried its hand at penning its own rhymes that mimic certain forms of verse. Its best efforts even fool people into thinking they’re reading the words of a human poet, rather than the algorithmic output of a cold-hearted AI.
The poetic bot is fully tunable, says Jack Hopkins, who developed the system while he was a researcher at the University of Cambridge. It can be programmed to write in a particular rhythm or pen poems on specific themes.
Set the theme to “desolation”, for example, and the angst-ridden AI comes up with the following snippet of verse:
The frozen waters that are dead are now
black as the rain to freeze a boundless sky,
and frozen ode of our terrors with
the grisly lady shall be free to cry
The AI can be endlessly tweaked to produce different flavours of poetry. It could write about Brexit in the style of a Greek epic, or rewrite snippets of Romeo and Juliet while mimicking Eminem, Hopkins says.
But flesh-and-blood poet Rishi Dastidar suspects that the AI is all surface and no subtext. Real poems explore ideas that might not be immediately apparent in the text, he says. But an AI doesn’t deal in ideas, it just puts one word after another.
Although it might be short on ideas of its own, the AI poet did have plenty of examples to draw inspiration from. It was trained on over 7 million words of 20th-century English poetry, most of it from poetry books found online.
This poetic education gave the neural network the ability to write lines of poetry one letter at a time. But rather than let the network freestyle, Hopkins added another element that encouraged it to write in particular styles or about certain themes.
Tell the neural network to write about fire, for example, and it will keep checking to make sure some of the words in the line it is writing concern fire. If the result isn’t fiery enough, the neural network scraps that part of the poem and starts again in the hope of picking more appropriate words.
Hopkins employed a similar mechanism to persuade the AI poet to write lines that rhymed or followed a particular rhythm. For example, Hopkins could make the AI write poetry in iambic pentameter – the poetic rhythm common in Shakespeare’s plays and sonnets.
Neverthless, the AI poet is pretty behind the times, says Dastidar. “The art form and the craft stopped thinking about these things seventy years ago,” he says. Modern poets deliberately choose when to follow or depart from formal constraints, but this AI is a slave to them.
This is inevitable, Dastidar says, because the AI poet is only trained on poems from decades ago and has no innate urge to write anything that’s truly novel. You can’t be truly creative, he says, if your template is only what has already been written.
Lack of creativity aside, the neural network still managed to fool some people who thought the poetry was written by a human. Hopkins asked 70 people to guess who’d written a fragment of poetry – a computer or a living, breathing poet – you can try the test for yourself here. The most human poem, it turned out, was actually written by AI.
Not everything about the AI poems were bad, says Dastidar. For example in this poem:
The crow crooked on more beautiful and free,
He journeyed off into the quarter sea.
his radiant ribs girdled empty and very –
least beautiful as dignified to see.
He thought the use of “crooked” as a verb was interesting, and might be the sort of thing you teach beginners in a creative writing class.
And while Dastidar isn’t convinced in general, he did write the poem below in response to Hopkins’s neural network. Perhaps the AI poet is better suited to life as a muse.
Non automatically-generated human response verse
The neural network has picked up its pen;
the programmed darling wants to be a poet.
Why it wants to is beyond our ken,
but the neural network has picked up its pen.
Maybe it’ll create a perfect digital madeleine
if the constraining models are set to “intuit”.
The neural network has picked up its pen;
the programmed darling wants to be a poet