The Machine Poetry Debate: What Software Cannot Want
A contest judge explains what generated verse gets right, where it goes hollow, and the five-step practice that makes your poems impossible for any program to imitate.
A poem crossed my scoring desk last spring that did everything right and made me feel nothing. Clean tercets. A competent extended metaphor about the sea and memory. Not one line you could flag as wrong, and not one line that had cost anybody anything. I gave it middling marks and could not stop thinking about why. That poem is my way into the machine poetry debate, because whether that entry came from a person imitating poems or a program predicting them, the diagnosis was identical: fluent language, absent wanting. The whole argument about software and the soul of poetry comes down to that one word. Wanting.
I am not going to spend two thousand words wringing my hands about the end of literature. Judges do not have that luxury; we have stacks to score. What I can offer is a working answer from the reading chair: what the programs genuinely do well, where they go hollow, how experienced readers spot the hollowness, and, most usefully, how to write the kind of poem no amount of prediction can fake.
The Poetry Debate Is Really a Debate About Wanting
Strip away the panic and the hype and the question underneath is old. It is the same question we asked about rhyming dictionaries, about creative writing degrees, about the sonnet-generating parlor games of the eighteenth century: is a poem the arrangement of words, or the pressure behind the arrangement? If a poem is only its arrangement, then a sufficiently good arranging machine writes real poems, full stop. If a poem is pressure made visible, then the machine has a problem no update fixes: nothing is pressing.
A program has no dead father. It has never waited for a biopsy result or held a grudge for eleven years or loved someone it was ashamed of. It can produce the sentence "my father's hands were rivers" because ten thousand poems taught it that fathers, hands, and rivers keep company. But it does not mean the sentence, because there is no father. Emily Dickinson wrote in 1862, "The Soul selects her own Society, / Then shuts the Door." Selection is the soul's signature. Software does not select. It averages. And an average, however elegant, has never once shut a door.
What the Programs Genuinely Do Well
Honesty requires giving the machines their due, because pretending they produce only gibberish loses the argument before it starts. Generated verse is often grammatically flawless, structurally tidy, and stocked with serviceable images. It handles received forms competently; ask for fourteen lines with a turn and you will get them. For beginners, that competence can even teach something real about scaffolding, the way a player piano can teach you what a waltz is shaped like.
Here is the uncomfortable part: that level of competence already beats a portion of any open contest's slush. Poems assembled from familiar parts, hitting expected beats, taking no risks. The machines did not invent that kind of poem. They industrialized it. If your writing was already an average of the poems you have read, you now have a very fast competitor. I consider that a useful alarm, not a tragedy. The bar for being worth reading just moved to exactly where it always should have been: above the average.
Where Generated Poems Go Hollow: Stakes
After a season of reading suspiciously frictionless entries, I can tell you the hollowness has a consistent shape. Three tells, in order of reliability:
- Nothing is at risk. The poem admires its subject from a safe distance. No confession that could embarrass, no detail so specific it must be checked against a real life, no sentence the writer visibly hesitated over.
- The images are category-true, not memory-true. Grandmothers have flour on their hands. Oceans are grief. Autumn means loss. Each image is plausible for the category and impossible to trace to one Tuesday in one kitchen.
- The poem never surprises itself. Line eight already lives comfortably inside line one. Human poems worth keeping usually swerve somewhere, because the writer discovered something mid-draft and let it wreck the plan. Prediction cannot swerve. Swerving is precisely what prediction exists to prevent.
Notice that all three tells describe plenty of human-written poems too. That is the point the whole debate keeps missing. The enemy was never silicon. The enemy is writing without stakes, and it predates electricity.
A poem is proof that its words cost somebody something. No cost, no poem, whoever or whatever typed it.
A Worked Example: Competence Versus Cost
Let me make the difference visible. First, a stanza I wrote to imitate the frictionless mode, the kind of thing prediction produces and tired poets default to:
Grief is an ocean, endless and deep, / its waves wash over the shores of my heart, / and memories drift like ships in the night / carrying pieces of you apart.
Nothing is wrong. Everything is borrowed. Ocean, waves, shores, ships: each image is on loan from the category called grief, and the heart pays no interest. Now the same feeling, rebuilt from one checkable, embarrassing, specific fact:
Your number is still second on my speed dial. / I demoted no one. I just stopped / calling anybody first. The phone / learns nothing. It offers you every day.
The second stanza risks something: it admits the speaker maintains a small shrine inside a phone, which is slightly shameful and completely human. It contains a fact a machine would never bother to invent because it is not statistically associated with grief; it is associated with one person's Tuesday. And it swerves, blaming the phone, then forgiving it, then accusing it of the cruelest thing of all, optimism. That swerve is voice: not a style, but the audible record of a particular mind changing in real time.
How Judges Spot the Hollow Poem
People keep asking whether judges can detect generated entries. Mostly we do not have to, because scoring for quality does the job automatically. On a five-criteria sheet like ours, hollow poems bleed points in emotion and creativity even when structure and language score high, and that lopsided profile is itself the fingerprint. The deeper truth: a hollow poem loses to a flawed, alive one every time I hold the two in my hands. I will forgive a broken meter for one image I have never seen before. I will not forgive perfection that wants nothing from me. Writing about loneliness, the most common theme in any month's stack, the alive poem names its specific room; the hollow one names the concept and decorates it.
A Five-Step Practice for Writing What a Machine Cannot
Complaining about the machines changes nothing about your next poem. This practice does. Run it on tonight's draft:
- Start from a fact only you possess. Not "loss" but the speed dial, the coat that still smells wrong, the joke nobody else at the funeral understood. If the detail could be googled, dig deeper.
- Include one line that embarrasses you slightly. Embarrassment is the body's proof of stakes. If every line is safe to read at a dinner party, the poem is a dinner party.
- Let the draft swerve once. Midway through, ask the poem a question you do not know the answer to, and write past your plan. Keep the wreckage if it is truer than the blueprint.
- Cut every image you have already read elsewhere. Be brutal. Oceans of grief, weight of the world, shattered hearts: if it arrived pre-assembled, it leaves.
- Read it aloud and mark where your own breath changes. Where nothing changes, nothing is happening. Rewrite those lines from the body, not the thesaurus.
A poet I mentored ran her drowned-village poem through exactly this sequence last winter. The first version was competent and interchangeable. The fifth version contained her grandfather refusing to name the underwater church, and it placed. The five-criteria breakdown showed her exactly where the risk paid: creativity and emotion jumped while structure held steady. Every entrant in our monthly contest gets that same breakdown, which is the fastest way I know to learn whether your poems are alive or merely correct. When yours is ready, put it in front of us between the 1st and the 25th and read the numbers honestly.
The Soul Question, Answered Plainly
So does the machine have a soul, and does poetry need one? I can only answer as a reader. When I finish a real poem, I have met someone. A specific, unrepeatable someone was lonely or furious or amazed, and pressed that state into language so hard the shape held. Generated verse can imitate the shape. It cannot supply the someone. There is nobody home, and readers, given enough lines, always end up knocking.
Which leaves us, the ones with dead fathers and speed dials and shame, holding the only advantage that ever mattered. The machines have finished proving that fluency is cheap. Good. Fluency was never the job. The job is to want something on paper so precisely that a stranger wants it too. Software cannot apply for that position. It has no wants to bring. You do. That is not a consolation prize in this debate. It is the entire game, and it was always yours to lose.
By Team Writory, for Writory.
Enter a Writory poetry contest
Enter the September contest and have your own poem read and scored. That first entry makes you a Writorian.