Why Large Language Models Seem to Overuse Em-Dashes

Abstract

This review examines the popular claim that large language models (LLMs) overuse em-dashes. The wording is intentionally loose: an LLM has no taste for punctuation. It produces a distribution over possible next tokens conditioned on its prompt, training, post-training, and decoding settings. The useful question is therefore narrower: under what conditions do some model outputs use the mark more often than a matched human baseline, and what could explain that result?

The current literature supports four careful conclusions. An em-dash pattern has been measured in some model outputs and in a large post-ChatGPT scientific corpus. Its rate varies substantially by model, instruction, and genre. It can be a salient cue that people choose to edit away. It is not, on its own, evidence that any particular passage was AI-written. Autoregressive continuation, training-register effects, and post-training for helpful explanatory prose are plausible contributors, but no study establishes one universal cause.

Keywords: large language models, stylometry, punctuation, em-dashes, post-training, AI writing, authorship attribution

What Has Been Measured

The direct evidence base is young. It combines a population-level corpus study, a targeted multi-model preprint, a peer-reviewed post-editing study, and broader stylometric research. Each source supports a different claim, so the distinctions matter.

A population-level change is not a causal attribution

Czuma analyzed first versions of medRxiv preprints deposited from 2020 through 2025. The preregistered primary cohort comprised 69,632 papers with sufficiently long Discussion sections. The prevalence of a Discussion section containing at least one em-dash increased from 4.23% before 30 November 2022 to 11.58% afterward, an absolute difference of 7.35 percentage points (95% confidence interval 6.94 to 7.77) [1]. The study also reports a delayed rather than immediate rise and several sensitivity checks.

That is the strongest directly relevant population-level result, and its own conclusion is restrained. It detects a change in writing, not the authorship of a particular paper. Editorial shifts, author demographics, writing software, or other factors could contribute. The paper is a preprint, so peer review may still revise its interpretation.

Model variation rules out a universal trait

Freeburg reports a targeted experiment across twelve instruction-tuned models from five providers. Across ten essay prompts, reported free-generation rates ranged from zero for the tested Llama models to 10.62 em-dashes per 1,000 words for GPT-4.1; the small human comparison averaged 3.23. Under a request to avoid Markdown, overt formatting was largely suppressed while em-dashes persisted in several models, reaching 9.10 per 1,000 words for GPT-4.1 in that condition [2].

The useful inference is heterogeneity, not a universal signature. The paper's Markdown account is a specific and testable proposal, not a demonstrated mechanism. A peer-reviewed counterexample reinforces the point: Gude et al. found that GPT-4o produced very few em-dashes or semicolons in a New York Times-style lead-generation task [9].

A Cautious Mechanistic Account

No study currently proves a single cause of em-dash frequency across deployed LLMs. The following account is therefore a synthesis. It explains why the pattern is plausible while reserving causal claims for controlled experiments.

Autoregressive continuation favors flexible bridges

Modern language models generate one token at a time. In compact form, a model estimates P(next token | previous text, prompt, trained parameters, decoding). The mark is unusually flexible in English prose: it can introduce clarification, contrast, apposition, or a late qualification without requiring a writer to commit early to a full stop, colon, semicolon, or parenthetical construction. In a local continuation distribution, that flexibility can make it a competitive bridge after a clause has already begun [6, 7].

This is an inference about generation, not proof that a particular punctuation token is universally high probability. Human writers benefit from the same rhetorical flexibility. The meaningful comparison is whether a defined model condition chooses the mark more often than matched human prose after controlling for genre, length, prompt, and editing stage.

Punctuation is computationally meaningful, but model-specific

Punctuation is not just decorative noise to a transformer. Chauhan et al. used interventions in GPT-2, DeepSeek, and Gemma and found substantially different roles for punctuation across architectures: it was necessary and sufficient in multiple GPT-2 layers, less so in DeepSeek, and not so in Gemma [4]. The study does not concern em-dashes specifically, but it rejects the simplistic claim that all models handle punctuation the same way.

Post-training can reshape surface style

Instruction tuning and reinforcement learning from human feedback change which responses a model produces for a prompt. Ouyang et al. showed that a 1.3B InstructGPT model was preferred by evaluators to a much larger untuned GPT-3 model on their prompt distribution [5]. Such procedures do not optimize a punctuation quota. Yet if evaluators favor fluent, explanatory, caveated prose, post-training can indirectly increase conventions common in that register.

This remains a plausible route, not an em-dash-specific causal finding. The alignment literature does show that small amounts of format-biased preference data can induce broad presentation preferences, although existing experiments have not tested this punctuation mark [11].

Markdown is a hypothesis worth testing

The most specific current account is that the em-dash can survive formatting suppression because it occupies both a prose and a structural register. Freeburg calls this Markdown leaking into prose [2]. The claim yields clear predictions: compare matched base and instruction models, vary Markdown exposure in controlled fine-tuning, hold prompts and decoding fixed, and observe whether the mark shifts independently of headings, lists, and bold text. Until such work is replicated, it should be presented as a hypothesis rather than a fact.

Why Readers Notice the Pattern

Readers do not need a frequency counter to notice a visual cue. An em-dash is larger and more salient than a comma, so repetition is conspicuous. It also creates a recognizable cadence: assertion, interruption, qualification. In short explanatory prose, repeated use can make the rhythm feel smooth but templated.

Baumler et al. provide a useful human counterpart. In a preregistered study of 81 people post-editing LLM drafts, editing moved text toward each writer's own unassisted style, yet the edited outputs remained closer to LLM text and less diverse than unaided human writing [3]. In an exploratory feature analysis, participants removed 23% of 254 em-dashes in the drafts. That is evidence that the mark can be salient during revision; it does not prove that the mark alone causes a reader to judge text as AI-generated.

What the Mark Cannot Tell Us

An em-dash is not an AI detector. Human writers have used it for centuries, and punctuation norms vary across genre, publication, and editor. A single instance is nearly uninformative; several instances may reflect a house style, a copy editor, accessibility software, or deliberate rhetoric. Even a corpus-level trend cannot attribute a particular document to a model [1].

Stylometry works best as a controlled multifeature analysis, not as a folk heuristic. Przystalski et al. separated defined human and LLM samples using lexical, grammatical, syntactic, and punctuation features together [8]. That does not license reverse inference from one symbol to authorship. Earlier work on machine-generated news likewise cautions that stylometry has important limits outside the exact provenance question and dataset on which it is evaluated [10].

The practical rule is straightforward: use the mark as a revision cue, not an accusation. If it makes a passage feel repetitive, edit for rhythm. Do not use it to claim that a student, colleague, scientist, or applicant used AI.

A Decisive Next Study

A strong causal study would preregister a factorial design. It would sample multiple base and instruction-tuned models, several versions per family where available, a matched human writing panel, and several genres. It would hold prompts, maximum length, temperature, top-p, and output-cleaning rules fixed. It would count U+2014 occurrences per 1,000 words and estimate a model with random effects for family, version, prompt, and genre.

The study should then manipulate one factor at a time: Markdown exposure in fine-tuning, instruction-tuning data, and decoding policy. It should publish prompts, tokenizers, seeds, inference dates, model identifiers, and raw outputs. Its primary conclusion should be relative: a confidence interval for the difference between a defined model condition and a defined human baseline. It should not claim that a passage containing an em-dash is AI-generated.

Conclusion

The strongest defensible answer is not that LLMs love em-dashes. Some models and configurations appear to produce them at notable rates, and a large scientific corpus shows a contemporaneous rise in their use. The evidence also shows substantial heterogeneity and serious attribution limits.

Autoregressive continuation, training registers, and post-training for helpful prose are plausible contributors, but targeted experiments are needed to apportion causality. Until then, the mark is best understood as a useful but over-salient style signal: worth editing for voice, not worth treating as forensic proof.

References

1. Czuma, P. (2026). Em-ergence of the em-dash: A population-level rise in em-dash frequency in medRxiv preprints at the dawn of the large-language-model era. arXiv:2606.29540. Preprint.
2. Freeburg, E. M. (2026). The Last Fingerprint: How Markdown Training Shapes LLM Prose. arXiv:2603.27006. Preprint.
3. Baumler, C., Bao, C., Nghiem, H., Yang, X., Carpuat, M., & Daume III, H. (2026). Can You Make It Sound Like You? Post-Editing LLM-Generated Text for Personal Style. Proceedings of ACL 2026, 43867-43895. https://doi.org/10.18653/v1/2026.acl-long.2030.
4. Chauhan, S., Chaudhary, M., Kiu, C. K., Nellessen, S., & Schoots, N. (2026). Punctuations and Predicates in Language Models. Findings of EACL 2026, 5622-5636. https://doi.org/10.18653/v1/2026.findings-eacl.297.
5. Ouyang, L., Wu, J., Jiang, X., et al. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35.
6. Brown, T. B., Mann, B., Ryder, N., et al. (2020). Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems, 33, 1877-1901.
7. Holtzman, A., Buys, J., Du, L., Forbes, M., & Choi, Y. (2020). The Curious Case of Neural Text Degeneration. International Conference on Learning Representations.
8. Przystalski, K., Argasinski, J. K., Grabska-Gradzinska, I., & Ochab, J. K. (2026). Stylometry recognizes human and LLM-generated texts in short samples. Expert Systems with Applications, 296, 129001.
9. Gude, A., Santos-Rios, R., Bond, F., Flickinger, D., Gomez-Rodriguez, C., & Zamaraeva, O. (2026). More Aligned, Less Diverse? Analyzing the Grammar and Lexicon of Two Generations of LLMs. Proceedings of ACL 2026, 38900-38911.
10. Schuster, T., Schuster, R., Shah, D. J., & Barzilay, R. (2020). The Limitations of Stylometry for Detecting Machine-Generated Fake News. Computational Linguistics, 46(2), 499-510.
11. Zhang, X., Xiong, W., Chen, L., Zhou, T., Huang, H., & Zhang, T. (2025). From Lists to Emojis: How Format Bias Affects Model Alignment. Proceedings of ACL 2025, 26940-26961.

Source-status note: references 1 and 2 are preprints at the time of writing. Their results are included because they directly address the question, but they are not treated here as settled consensus.