After processing more than half a million lines by human poets, experts say it is the best attempt yet to produce computer-generated poetry
Rare is the poet who has failed to tackle the glory of trees, whether it’s Joyce Kilmer (“I think that I shall never see / A poem lovely as a tree”) or Philip Larkin (“the unresting castles thresh / In fullgrown thickness every May”). Now an artificial intelligence trained by experts on more than half a million lines of poetry has had a stab, coming up with the almost-comprehensible image of a “box of light that had been a tree”.
The algorithm, which those behind it believe is the best attempt to date at training an artificial intelligence to write poetry, was fed lines from more than 100 British contemporary poets as inspiration, learning from the style of poets such as Simon Armitage and Alice Oswald. It was then given “seed words”, from which it would generate couplets based on its understanding of what poetry was. Experts from the Poetry Society, Poetry Archive and Scottish Poetry Library then filtered through tens of thousands of couplets to highlight what did, and didn’t, work. They repeated this over and over again in a five-month period, before the AI’s output began to improve.