Why AI poetry rhymes too well

Ask a language model for a rhymed quatrain about snow and you will get one, instantly, and it will be correct: four lines, alternating rhyme, iambic enough to tap your foot to. That correctness is the tell. Poems by working poets are almost never this obedient, because meter and rhyme are instruments of pressure, not of decoration — the form is there to be strained against, and the strain is where the poem lives.

Read Dickinson and count the crimes: slant rhymes that barely rhyme, grammar bent until it cracks (“say how many dew”), abstractions capitalized like criminal defendants, poems that stop rather than end. Her hymn meter is a cage she rattles. A model imitating her keeps the cage and releases the animal — you get the dashes, the capital letters, the churchy cadence, and none of the danger. We run this experiment daily in the puzzle, and players learn to feel the difference within a week: the imitation consoles, the original unsettles.

The image parade

Free verse has its own failure mode. Because contemporary models learned poetry substantially from the internet's most-shared poems, they write in the house style of virality: a procession of striking images, one per line, each a self-contained metaphor — hands as countries, snow filing reports, grief as an unsigned letter. No image is bad. That's the problem. A human free-verse poem almost always exhales somewhere: a flat line, a spoken aside, a piece of plain information that lets the striking lines strike. The machine never exhales. Its poems are all peaks, which is a kind of flatness.

Whitman is the stress test here. His catalogues look mechanical — list after list — but inspect any one of them and you find items that earn their place only because he liked the sound, plus grammar that goes feral under pressure: “the lone singer wonderful causing tears.” An adjective shoved after its noun, a participle dangling off the end of the line. Models smooth this wildness out automatically, the way autocorrect fixes a typo. What gets fixed is the poem.

Risk as a metric

The deepest difference is appetite for embarrassment. Real poems risk being ridiculous — Whitman apostrophizes his own throat, Dickinson rhymes “strategy” with “physiognomy,” Frost lets a couplet go corny because the thought needed it. Machine poems are tasteful to a fault. They grieve beautifully, praise appropriately, and land endings that would survive a workshop. If you finish a poem and feel gently improved rather than slightly implicated, check the byline.

A practical checklist, then. Exact rhymes everywhere: suspicious. Meter that never stumbles: suspicious. Every image striking, no line flat: suspicious. Closing line that reframes the whole poem as wisdom: very suspicious. One genuinely broken moment that a careful editor would have queried: probably human. The limp is the fingerprint.