Em-dash overuse
Em-dashes deployed as a universal connective tissue — between clauses, inside lists, instead of commas, colons, and parentheses.
What it is
The em-dash (—) shows up several times per paragraph, doing the work of commas, colons, parentheses, and semicolons indiscriminately. The cadence becomes recognizable: clause — qualifier — restatement — close.
Why models do it (first principles)
Instruction-tuned chat models were trained on a heavy diet of polished web prose and editorial writing where the em-dash is fashionable. Reward models learned to associate em-dash rhythm with 'fluent' or 'sophisticated' output. At inference time the next-token distribution over '—' is unnaturally peaked compared to a human writer who reaches for a comma or full stop first. There is also no dispreference signal during RLHF for using the same punctuation mark twice in one sentence.
How to think about it
This is overfitting to a stylistic surface feature. A human writer chooses punctuation to reflect a mental model of how the reader will parse the sentence — pause, aside, abrupt pivot. The model has no such parsing model; it has a learned association between certain rhythms and reward. So it reproduces the silhouette of careful prose without the underlying decisions that produced it. Once you see it you cannot unsee it: the punctuation is a tell that no one is home behind the cadence.
Examples
The framework is powerful — and flexible — but it requires care — especially around state — to avoid bugs.
The framework is powerful and flexible, but it requires care around state to avoid bugs.
Three em-dashes in one sentence is almost always a tell. Commas and a single conjunction do the same job without performing 'thoughtful prose'.
Coffee — the kind that smells like a roastery — is what got me through.
Coffee, the kind that smells like a roastery, is what got me through.
Fix prompt
Choose punctuation to track how the reader will parse the sentence, not to perform fluent rhythm. Strong dashes are for genuine breaks the reader needs help noticing; reaching for them as default cadence flattens the prose into a recognizable silhouette and tells the reader that the writing was shaped by feel rather than by thought about how it lands.
Watch for
Concrete phrasings this pattern usually shows up as. These are not part of the copyable prompt. The prompt teaches the principle so the model can recognize the move even when the exact phrasing differs. Use this list to self-audit your own writing or to test a model.
- — used between clauses where a comma would do
- — used as a parenthetical aside
- multiple em-dashes in a single sentence
- — used to attach a closing flourish