Why Judges Can Tell a Machine Wrote Your Poem
Generated poems are flooding contest inboxes, and judges catch them without any detector. A working judge explains the seven tells, what the score sheets reveal, and how to protect your own voice.
Last autumn, three entries arrived in the same judging cycle from three different countries, and all three ended on the same move: an abstract noun being reborn in the light of another abstract noun. Hope reborn in the light of dawn. A soul reborn in the light of love. Different names, different continents, one author. That author was software, and this is the open secret of generated poems in contests right now: judges do not need a detector to catch them. The poems confess. After scoring thousands of entries, I can usually feel a machine-made poem by line four, and I want to show you exactly what gives it away, because the answer will make your own writing better even if you have never touched a text generator in your life.
The Tell Is Sameness, Not Badness
Here is what most people get wrong. They assume judges catch generated poems because the poems are bad. The opposite is closer to true: the poems are suspiciously competent. Clean grammar, tidy stanzas, consistent tone, a bow-tie ending. What they lack is friction. Human poems, even brilliant ones, have burrs: an image that arrives from nowhere and gets defended anyway, a word choice slightly wrong in a way that turns out right, a private reference the poet did not bother to explain. Generated verse averages those burrs away, because averaging is what the technology fundamentally does. It produces the most probable poem, and the most probable poem is the one a judge has already read nine hundred times.
The result reads like a greeting card with a thesaurus: fluent, warm, and completely uninhabited. Nobody's actual kitchen is in it. Nobody's dead brother. No street with a name.
A poem is evidence that a particular life happened. Counterfeit evidence convicts itself.
Seven Tells Judges Notice First
Ask any working judge and you will hear versions of the same list. Mine, in the order I usually notice them:
- Interchangeable abstractions: souls, echoes, whispers, shadows, light, doing things to each other with no object from the physical world in sight.
- Uniform line lengths and rhythm: every line the same weight, no rush, no stumble, no held breath. Real poems breathe unevenly.
- The explained metaphor: an image immediately followed by its meaning, as if the poem were grading itself. Like a bird, my heart takes flight, meaning I felt free.
- Symmetrical endings: the last stanza circles back to the first with mechanical neatness, resolution guaranteed, nothing left open.
- No proper nouns: no towns, no brand names, no month of the year, no sister called by name. Specificity is risk, and the software is built to avoid risk.
- Emotion announced, never earned: the words grief, joy, and longing appear, doing the job that images should be doing.
- Era-blind vocabulary: a poem supposedly about a modern breakup reaching for gilded, ethereal, and forevermore, because the training diet skews toward the archive.
Any one of these can appear in an honest human draft. Four or more together, in a poem with no friction anywhere, and every judge at the table is already exchanging looks.
What the Score Sheet Shows
Numbers back up the instinct. Writory's monthly contest scores every entry across five criteria, creativity, emotion, structure, language, and theme, and machine-flavored entries produce a signature so consistent I could sketch it blind: language scores in the high range, structure respectable, and creativity and emotion sagging fifteen to twenty points below them. Fluency without invention. A human beginner shows the reverse profile, raw language mechanics with some image nobody else on earth would have chosen. I trust the second profile with a future. The first one has no future, because the ceiling is built in: you cannot revise your way to originality inside a system engineered to be typical.
That gap matters for entrants who never cheat, too. If your honest poem leans on stock phrases and announced feelings, it drifts toward the same profile, and it gets read with the same fatigue. The bar for specificity has been raised for everyone. I consider that the one gift this mess has given the art form.
One Stanza, Two Authors: a Worked Example
Let me make the difference physical. Here is a stanza I generated in the house style of the machine, on the theme of losing a brother:
In the silence of my aching heart, / whispers of your memory dance like shadows, / and though the darkness tries to keep us apart, / your light lives on within me always.
Fluent. Rhymed, even. And empty: no brother exists in it. Now here is the stanza a human wrote in my workshop on the same theme, four drafts in:
Your boots are still in the hallway. / Size eleven. Nobody moves them. / The dog checks them every morning, / then eats.
No grief announced, no light, no heart. Just boots, a shoe size, and a dog that has adjusted faster than the family. The second stanza cannot be generated, not because the software lacks skill, but because it lacks the hallway. It was not there. Specificity is not a style. It is a receipt, and judges are trained to check receipts. If your drafts live closer to the first stanza than the second, the fastest repair I know is studying how images carry feeling without naming it, which is the entire craft of the objective correlative, and then testing yourself against a theme as worn as grief, where only the specific survives.
The Rules, the Risk, and the Honor System
Now the blunt policy part. Nearly every reputable contest, ours included, requires that entries be your own original composition, and machine-generated text fails that test by definition. Getting caught usually means disqualification, sometimes a ban, and in prize situations, returned money and a name quietly shared among organizers. It happens more than entrants think, because the failure modes are so repeatable: two entrants in the same cycle submitting near-identical poems is not a coincidence, it is two people typing the same prompt.
But enforcement is honestly the smaller stake. The larger one is this: contests exist to find and encourage writers. If the poem is not yours, winning gets you nothing you can build on. There is no second book in a prompt. The certificate would certify someone else, and some part of every entrant who cheats already knows that, which may be why so many of the generated entries arrive with no cover note at all, like a package nobody wants to sign for.
Five Ways to Keep Your Poem Unmistakably Yours
Whether you draft with software assistance and revise heavily, or never touch the stuff, the same discipline protects you, because the real enemy is genericness from any source. Before you enter anything, run this sequence:
- Count your proper nouns and objects. A poem with zero named places, people, or physical things is running on fumes. Add the hallway, the month, the brand of tea.
- Circle every abstraction. Soul, hope, darkness, heart. Replace at least half with something you could photograph.
- Break one pattern on purpose. If every line is the same length, shorten one to three words and let it land hard.
- Cut the explanation after your best image. Trust the boots. Delete the line that tells the reader what the boots mean.
- Read it for the burr. Find the one strange, private, defensible choice in the poem. If there isn't one, the poem is not finished, whoever wrote it.
The Machine Raised the Bar. Clear It.
I will end with the opinion I opened this piece to defend. The flood of generated verse has not ruined contests. It has clarified them. For years, competent-but-hollow poems could place on polish alone. That lane is now closed, because polish is free and everybody can smell it. What still cannot be manufactured is the poem only you could have written: the shoe size, the dog, the thing your grandmother said wrong in two languages. That was always the good stuff. Now it is also the only stuff.
So the strategy for the current era is almost embarrassingly simple. Write from your actual life with your actual nouns, revise until the burrs are load-bearing, and submit to rooms where a human reads every line. If you want proof that yours reads as lived, enter a contest that shows you the score breakdown and look at the gap between your language number and your emotion number. When emotion and creativity run level with language, no judge on earth will wonder who wrote your poem. The evidence will be all over it, the way it should be.
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.