A chatbot rhymes a poem but can't count its syllables

Ask for a haiku, then count the syllables

A chatbot asked for a haiku hands back three tidy lines about cherry blossoms. Count the syllables and the form falls apart: often five, then eight, then six, where a haiku wants five, seven, five. Ask for a limerick and the rhymes usually land while the beat stumbles through the middle.

The gap is measured, not just felt. A 2024 benchmark, PhonologyBench, put text-only models through three tasks: counting the syllables in a sentence, producing a word that rhymes with a given one, and turning spelling into pronunciation. On syllable counting, people scored 90% and the models trailed them by 45 points. On rhyming, the same models trailed by only 17.¹ Same systems, same input, and the task that turns on counting is the one they are worst at.

To see why counting is the hard part, follow a line of verse backward from the page into the model, one step at a time, and stop at the step below which nothing changes the answer.

The model never receives a syllable

Start with what the model is handed. A line of verse does not reach it as letters, and it does not reach it as sound. A step called the tokenizer first chops the text into tokens, the sub-word chunks a model actually reads, each one an entry in a fixed vocabulary. The chunks are chosen for a single reason: frequency. The method most models use, byte-pair encoding, begins with single characters and repeatedly merges whichever adjacent pair occurs most often across the training text, stopping when the vocabulary reaches a set size.² A common word like rabbit survives as one token; a rare one breaks into several pieces. Nothing in that procedure ever consults how a word sounds.

So the two things a poet counts are simply not in the input. A syllable is a unit of sound, one beat of a spoken word. Stress is which syllable you lean on when you say it. The word record is really two words: stressed on the first syllable, REC-ord, it is a noun; stressed on the second, re-CORD, it is a verb. The seven letters are identical. Only the beat differs, and the beat is nowhere on the page. A person hears it. The model receives record as a token or two, with no channel that carries the beat at all.

Meter is built from exactly those two things. A haiku counts syllables, five then seven then five. Iambic pentameter counts them too, ten to a line, and adds a stress pattern: an unstressed beat followed by a stressed one, five times over. To hold any of it a writer has to count a kind of sound the model was never given.

Could the spelling fill the gap?

One reasonable hope is that the model reconstructs the sound from the spelling. For English, that is a poor bet, and the reason is a property of the language itself. English has a deep orthography, meaning its spelling maps only loosely onto its sounds. The pair ea says one thing in beat and another in head, with no mark on the page to tell them apart.³ The letters underdetermine the sound.

The size of the bet is measurable. PhonologyBench also scored grapheme-to-phoneme conversion, the job of turning spelling into a phonetic transcription. A dedicated tool built for nothing else topped out near 62% on common English words, and the best model came in under 53%.¹ When the purpose-built converter misses more than a third of the time, a general chatbot inferring sound from letters on the side has little firm ground to stand on.

Why the rhymes land and the meter limps

If sound is missing, why does rhyme come out so much better than meter? For two reasons, and neither one restores the missing sense of sound.

The first is that rhyme is thick in the training text. Human poems rhyme most of the time: across large collections of human verse, fewer than about 15% of four-line stanzas use no rhyme at all.⁴ A model trained to predict the next token has seen flame close lines that end in name and same countless times. It can reproduce that surface pattern without knowing what any of the words sound like, the way it can finish a familiar phrase it has never heard aloud.

The second is a head start the model gives itself on rhyme. Interpretability work that traced a model composing a rhyming couplet found it settling on the final rhyme word before it wrote the line, then bending the words in between toward that ending; suppressing the chosen word made the model pick a different one and rewrite the whole line to reach it.⁵ That is the forward planning an earlier lesson covers, and the result surprised the researchers, who had assumed the model chose one word at a time.⁶ Meter gets no equivalent. There is no single syllable-count the model can fix on at the line break and build the line toward.

None of this makes rhyme reliable. On PhonologyBench, the strongest model rhymed common words about 69% of the time but rare words only 46%, where people managed 86% and 60%.¹ Rhyme success tracks how common a word is, which is what a system leaning on how often word-endings co-occur would produce, rather than one hearing the endings. Rhyme is better than meter. It is not solved.

The floor beneath the failure

Keep going down and the steps run out. The part of the model that writes the poem has no syllable counter and no stress map to consult. Nothing in the path from text to output was built to hold a beat, so when a form demands an exact count the model approximates from surface statistics and often misses. That much is settled: tokens carry no direct syllable or stress information, and on the tasks that need it, the models measurably fail.

Two questions sit above that floor, and both are genuinely open. The first is whether the sound is truly absent or merely unused. Recent probing work points to unused. A simple classifier reading a single-token word's internal representation can recover which phonemes the word contains about 96% of the time, against 42% for a random-embedding control, and nudging that representation along a learned "vowel" direction changes which words the model rhymes.⁷ Some phonetic structure is latently present in what the model learned. It is just not reliably deployed for exact counting.

The second open question is why some models scan so much better than others. On the same syllable-counting test, with the same text-only input, one strong model reached 55% while another sat at 23% and a third at 20%.¹

That is a spread of more than 30 points between two capable closed models on identical input, and none of the sources explains it. One clue does sharpen the picture. Give a model finer-grained input and the numbers move: feeding characters instead of whole-word tokens raises rhyming and syllable scores, and inserting character-level delimiters lifted one model's rhyme-awareness from 66% to 77%.⁸ The structure is recoverable when the tokenization stops hiding it, which points at the tokenizer, not the text medium, as the bottleneck.