This post includes the full transcript of this week’s Beyond the Buzz episode, followed by the full evidence.
🎧INTRO
Lately, my LinkedIn feed has been warning me about the em dash. The message is that using one can make your writing look AI-generated.
That caught my attention because I’m a stickler for grammar — and not long ago, a senior editor actually added an em dash to something I had written. So I found myself wondering: should I change a perfectly legitimate punctuation choice just to make my writing look less AI-generated — or is the em dash getting more attention than the evidence warrants?
Welcome to Beyond the Buzz.
I’m Dr. Tara Moroz, scientist and communicator with decades of experience translating complex human research into clear, evidence-informed insight.
Let’s take a closer look together — starting with what’s driving the buzz.
📊THE BUZZ
The concern is not just circulating on social media. Grammarly’s current guide to em dashes says heavy use can sometimes signal AI-generated content, and it gives writers ways to prevent or remove em dashes from AI-generated text when trying to make it sound more human. [H1]
That is a pretty specific ask: change an ordinary punctuation choice because of what it might signal about who — or what — wrote the words.
There is also a reason people are paying attention to AI writing style. In a Pew Research Center analysis of English-language webpages, 35% of pages in its July 2026 crawl with post-ChatGPT publication dates showed signs of AI authorship. Those classifications were detector-based, not confirmed authorship. [H2]
But trying to identify AI writing can also catch human writing. In one study of over 14,000 scientific abstracts, an AI-text detector falsely classified up to 8% of known human-written abstracts as AI-generated. [H3]
So the pressure is real: writers are being told to watch something as specific as an em dash, while the broader task of identifying AI writing is itself imperfect. The question is how much that punctuation mark can actually tell us.
🧾RECEIPT CHECK
Let’s check the evidence — our kind of receipt check.
This is the moment to pause and ask the questions that matter — what’s the evidence,
how certain can we be, and what does it mean for your decision?
🔬WHAT THE EVIDENCE SHOWS
Here’s what the evidence shows.
The em-dash reputation is not completely invented. Pew notes that em dashes tend to appear more often in AI-generated than human writing. [E1]
And in its broader sample of webpages, em-dash use roughly doubled between 2023 and 2026. But because that sample reflects a web increasingly containing both human and AI-generated or AI-edited writing, the increase cannot tell us whether human writers themselves started using more em dashes, whether growing AI use contributed to the rise, or both. [E1]
And neither finding makes the em dash a reliable marker of AI authorship in an individual piece of writing. Pew cautioned that an em dash alone does not establish AI authorship, and its AI classifications were probabilistic rather than based on confirmed authorship. [E1]
Part of the problem is that writing style does not cleanly separate human writing from AI-generated text across all contexts. [E2,E3,E4,E5,E6,E7]
Writing style varies among human writers, and it varies across AI models too. Peer-reviewed research has found differences in grammatical and rhetorical style across AI models rather than one uniform “AI voice.” [E2] And preliminary 2026 research looking directly at em-dash use found substantial variation across 12 AI models — and across eight published human essays used for comparison. [E3]
So an em dash can be associated with AI writing without being unique to it. [E1,E3]
But the bigger question is whether an em dash actually influences how people or AI-detection tools judge who wrote something.
In a small, non-peer-reviewed study of 85 people, 12 spontaneously mentioned the em dash when explaining whether they thought writing was human- or AI-generated, and 11 of those associated it with AI. [E8]
So at least some readers do use the em dash as an AI cue. But they were comparing two different passages that differed in many ways, so the study cannot tell us whether the em dash itself caused that suspicion or how common that reaction is more broadly. [E8]
I did not identify a controlled study testing whether an em dash, with the rest of the writing held comparable, makes readers more likely to judge it as AI-generated. [G1]
Nor did I identify a peer-reviewed experiment testing whether removing or replacing em dashes makes people or detectors less likely to classify otherwise comparable writing as AI-generated. [G2]
And I did not identify a peer-reviewed study validating the em dash by itself as a reliable way to classify an individual text as human- or AI-generated. [G3]
That uncertainty is not limited to the em dash. [E4,E5,E6,E7]
Automated detectors have been tested across different AI models, writing domains, generation settings, and altered text, and their performance varies across those conditions. [E4]
Human judgments vary too. In one study of long-form nonfiction, people who frequently used ChatGPT for writing were much better at identifying AI-generated text than less-experienced evaluators. In the initial comparison, frequent users correctly identified about 93% of AI-generated articles, while the less-experienced group identified about 57% — close to chance. [E5]
And authorship is not always simply human or AI. In another small study, five blinded raters were asked to distinguish five levels of AI involvement, ranging from fully human to fully AI-generated writing. Their overall accuracy was 19% — about chance for a five-way task. [E7]
That blurry middle ground matters for automated detection too. In a separate study, 12 detectors were tested on writing ranging from fully human to writing with increasing amounts of AI editing, and classification became more challenging as the amount of AI involvement changed. [E6]
Across those studies, both human and automated judgments vary depending on who or what is doing the judging, the kind of writing involved, and how much AI assistance sits in the middle. [E4,E5,E6,E7]
Taken together, that makes it hard to justify a confident authorship judgment based on one punctuation mark. [E1,E4,E5,E6,E7,E8,G3]
🧠WHY THIS BUZZ RESONATES
So why does this resonate?
An em dash is easy to see and easy to change. If you are worried that someone might question whether your writing is really yours, avoiding one punctuation mark can feel like simple insurance.
And there is something appealing about having a visible rule for a messy problem: this looks like AI, so don’t do this.
But writing does not always cooperate with rules that make a messy question feel simple.
🧭THE TAKEAWAY
So what’s the takeaway?
You do not need to stop using em dashes simply because they have become associated with AI writing. In one small study, some readers did use them as an AI cue, and em dashes may appear more often in AI-associated text. But the evidence does not support treating an em dash as a reliable standalone sign of AI authorship. [E1,E3,E8,G3]
We also do not know how much an em dash changes reader suspicion, or whether removing one reduces it, because I did not identify controlled experiments testing those questions. [G1,G2]
The useful distinction is between a pattern and a fingerprint. An em dash may contribute to an impression, but the evidence does not support treating it as enough, on its own, to determine authorship. [E1,E4,E5,E6,E7,E8,G3]
So I’m not planning to retire the em dash that my editor was perfectly happy to add. If it is the punctuation that makes the sentence work, I’ll use it.
The em dash may be a clue about style. It is not proof of who wrote the words. [E1,E3]
💭REFLECTION PROMPT
Something to reflect on…
Have you started changing the way you naturally write because you worry it might “look AI”?
📬OUTRO & CTA
If you found this useful, follow Beyond the Buzz and share it with a friend who likes to know what actually holds up.
You can also explore the full transcript, the clarity poll, and evidence in The Evidence Edit.
Until next time, stay curious.
This is Beyond the Buzz — I do the research so you don’t have to.
Next week: Are full-body MRI scans worth it if you feel healthy?
📚REFERENCES — What’s the Hype (H1–H#) / What’s the Evidence (E1–E#)
🔓 Open Access |🔒Paywalled
HYPE SOURCES
H1
Joki, K. (2026, January 28). What is an em dash? How to use an em dash correctly. Grammarly. https://www.grammarly.com/blog/punctuation-capitalization/what-is-an-em-dash/ 🔓
H2/E1
Bestvater, S., Smith, A., TerBush, C., Baronavski, C., & Chavda, J. (2026, August 20). How much of the internet is written with AI? Pew Research Center. https://www.pewresearch.org/data-labs/2026/08/20/how-much-of-the-internet-is-written-with-ai/ 🔓
H3
Rashidi, H. H., Fennell, B. D., Albahra, S., Hu, B., & Gorbett, T. (2023). The ChatGPT conundrum: Human-generated scientific manuscripts misidentified as AI creations by AI text detection tool. Journal of Pathology Informatics, 14, Article 100342. https://doi.org/10.1016/j.jpi.2023.100342 🔓
EVIDENCE SOURCES
E1/H2
Bestvater, S., Smith, A., TerBush, C., Baronavski, C., & Chavda, J. (2026, August 20). How much of the internet is written with AI? Pew Research Center. https://www.pewresearch.org/data-labs/2026/08/20/how-much-of-the-internet-is-written-with-ai/ 🔓
E2
Reinhart, A., Markey, B., Laudenbach, M., Pantusen, K., Yurko, R., Weinberg, G., & Brown, D. W. (2025). Do LLMs write like humans? Variation in grammatical and rhetorical styles. Proceedings of the National Academy of Sciences, 122(8), Article e2422455122. https://doi.org/10.1073/pnas.2422455122 🔓
E3
Freeburg, E. M. (2026). The last fingerprint: How Markdown training shapes LLM prose [Preprint, arXiv:2603.27006v1]. arXiv. https://doi.org/10.48550/arXiv.2603.27006 🔓
E4
Dugan, L., Hwang, A., Trhlík, F., Zhu, A., Ludan, J. M., Xu, H., Ippolito, D., & Callison-Burch, C. (2024). RAID: A shared benchmark for robust evaluation of machine-generated text detectors. In L.-W. Ku, A. Martins, & V. Srikumar (Eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 12463–12492). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.acl-long.674 🔓
E5
Russell, J., Karpinska, M., & Iyyer, M. (2025). People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text. In W. Che, J. Nabende, E. Shutova, & M. T. Pilehvar (Eds.), Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 5342–5373). Association for Computational Linguistics. https://doi.org/10.18653/v1/2025.acl-long.267 🔓
E6
Saha, S., & Feizi, S. (2025). Almost AI, almost human: The challenge of detecting AI-polished writing. In W. Che, J. Nabende, E. Shutova, & M. T. Pilehvar (Eds.), Findings of the Association for Computational Linguistics: ACL 2025 (pp. 25414–25431). Association for Computational Linguistics. https://doi.org/10.18653/v1/2025.findings-acl.1303 🔓
E7
Cheng, A., Lin, Y., Reedy, G., Joseph, C., Wirkowski, S., Mallette, V., Nagesh, V., Krieser, D., & Calhoun, A. (2025). Ability of AI detection tools and humans to accurately identify different forms of AI-generated written content. Advances in Simulation, 10, Article 66. https://doi.org/10.1186/s41077-025-00396-6 🔓
E8
Liu, M. (2025). Perceiving the humanness: Reader evaluation of AI-generated versus human-written romantic texts [Master’s thesis, Erasmus University Rotterdam]. Erasmus University Thesis Repository. https://thesis.eur.nl/pub/76801/ 🔓
🔎 WHAT WE STILL DON’T KNOW (Evidence Gaps G1–G#)
Evidence search current to September 2026.
G1 — Does seeing an em dash actually make readers more likely to think a piece of writing was AI-generated?
Some readers in one small study spontaneously identified the em dash as an AI cue, but the passages they compared differed in many ways. Directly relevant controlled evidence was not identified that isolates the em dash itself, so we cannot tell how much seeing one changes a reader’s judgment.
G2 — Does removing em dashes actually make human writing less likely to be mistaken for AI?
No directly relevant peer-reviewed experiment was identified that changed only the em dashes in otherwise comparable writing and then measured human judgments or detector false positives. So the practical benefit of deleting or replacing em dashes specifically remains unquantified.
G3 — Can an em dash, by itself, reliably tell you whether a piece of writing was AI-generated?
Directly relevant validation evidence showing that em-dash presence or frequency alone can reliably classify an individual text was not identified. Population-level associations and model experiments do exist, but those are not the same as validating the punctuation mark as an individual authorship test.
🎧 Prefer to listen?
Listen to Does an Em Dash Make Your Writing Look AI-Generated? on the Beyond the Buzz website — or follow Beyond the Buzz in your favourite podcast app.
Educational content only. This publication does not provide individualized medical, psychological, or professional advice.
Full disclaimer: beyondthebuzzmedia.com/disclaimer


