Can AI Write a Good Poem? A Candid Assessment

The honest answer is: a generator can write a competent poem, rarely a good one, and never — so far — a necessary one. That last word is the whole argument, and it is worth unpacking rather than treating as a slogan.

What generated poetry is genuinely good at

Fluency, first. A language model has absorbed enough verse to produce lines that scan, that use enjambment sensibly, that avoid the clumsiest beginner mistakes. Ask for a haiku and you will get seventeen syllables with a seasonal image. Ask for a sonnet and you will get a volta in roughly the right place. This is not nothing; formal competence takes human beginners a year or more to acquire.

Second, combinatorial surprise. Because a model is sampling across an enormous space of association, it will occasionally hand you a pairing you would not have reached — "the lamp keeps the shape of the hour", say. Roughly one line in thirty is genuinely interesting. That is a poor hit rate for a finished poem and an excellent hit rate for a prompt to yourself.

Third, speed at scale. Twenty variations on an opening image in ten seconds is a real creative tool, in the same way a thesaurus is a real tool and for the same reason: it widens the field of options you consider before committing.

Where it consistently fails

Generated poems have no stake. Every real poem I care about was written by someone who needed something — to be remembered, to be forgiven, to hold a shape around a loss, to win an argument with a dead parent. Henley wrote "Invictus" from a hospital bed while surgeons discussed amputating his remaining leg. That circumstance is not trivia attached to the poem; it is the pressure that produced the flat, refusing tone. A model has no leg, no bed, and nothing to refuse.

The symptom of this is a specific kind of ending. Generated poems almost always close on a graceful abstraction — light, silence, becoming, the endless sky. It is a beautiful non-answer, and it is what you write when you have no particular thing to say and excellent taste in how to say nothing. Compare Dickinson's crumb. A model would not risk a crumb; it is too small, too odd, too specific to be safely lovely.

There is also the averaging problem. A model produces something near the centre of everything it has read, and poetry lives at the edges. Ask for a love poem and you will get the statistical middle of all love poems, which is exactly the territory a good poet is trying to escape.

The useful workflow

• Generate ten short pieces on your theme. Read them once and delete nine.
• From the survivor, keep only the individual lines that surprised you. Usually one or two.
• Put those lines in a blank document with no context and write toward them from your own material.
• Never keep a generated final line. That is the position where a poem has to mean something, and meaning is the part you have to supply.
• Use rewriting tools for register experiments — see what your own line sounds like as classical, as dark, as plain — then choose deliberately.

The question underneath the question

People asking whether AI can write poetry are usually asking something more pointed: is the thing I do valuable? Historically, poets have absorbed every technology thrown at them — the printing press, free verse, the typewriter, the microphone, the internet — and each one changed the distribution of poetry far more than it changed the substance. This will follow the same pattern. Vast quantities of pleasant, competent, forgettable verse will now exist. It will make genuinely felt writing easier to distinguish, not harder, in the same way that mass-produced furniture made handmade joinery more visible rather than less.

There is a real cost, and I would rather name it than pretend otherwise: the flood of adequate text raises the noise floor, and readers with limited attention may not dig past it. That is a distribution problem for poets to solve, and it is not new — obscurity has been the default condition of poetry for centuries.

My working position

I use generation the way I use a metronome. It keeps me moving, it exposes patterns I have fallen into, and it never plays the piece. When I am stuck at line eight I will happily ask for twenty possible line nines, take none of them, and discover from the shape of what I rejected what I actually wanted to say.

Write with the tool, then take the tool out of the poem. What remains — the crumb, the hoe stuck blade-up in the soil, the dash after "and I—" — is the part that was always going to be yours.