Em Dashes, Hyphens and Spotting AI Writing

In late November 2022, OpenAI publicly launched ChatGPT. For the first time, generative AI tools were available freely and broadly, sparking a gold rush to find ways to exploit the technology.

Since then, humans have been seeking ways to distinguish between AI and human-written content. Those efforts have not always gone well. A cottage industry has emerged around AI detection, yielding varying results.

Although the technology appears to be slowly improving, it’s also impractical to scan every post, comment, or article for AI usage. When you factor in that it’s difficult to distinguish between AI “polishing” and AI writing, it’s easy to see why people are looking for answers elsewhere.

To that end, some claim to have found a shortcut: The humble em dash

The argument is relatively straightforward. AI systems commonly use the em dash, but humans are less likely to do so.

But does the approach work? The answer is complicated.

What is an Em Dash and Why They Matter

The em dash is a straightforward piece of punctuation. It’s a longer dash that’s meant to be the same width as a capital “M” when used in a sentence. This separates it from a shorter en dash, which is the width of a capital “N,” and hyphens, which are even shorter.

According to Merriam-Webster, em dashes are used to indicate a change in direction mid-sentence. It often takes the place of a comma or a parenthesis.

Though it’s not a common piece of punctuation, humans do use it. AP Style, Chicago and APA Style all have guidelines for using em dashes. MLA favors en dashes. However, in all cases, the use is meant to be sparing. Em dashes are meant to draw attention to a passage, and overuse dilutes that purpose.

However, according to many commenters, AI never got that message. They note that AI text is often filled with em dashes, including in places where it isn’t wholly appropriate. As such, some have taken to calling it the “ChatGPT hyphen,” saying its use is a strong indicator that a work was AI-generated.

This is made worse by the fact that keyboards do not have a dedicated em dash key. Even 100% layout keyboards require a shortcut or the use of autocorrect to type an em dash. As such, most humans use hyphens, not em or en dashes.

But do AI systems overuse the em dash? I decided to put it to the test.

Testing the Systems

I decided to test six different AI systems to see how heavily they use the em dash in their writing. I gave each of them the same prompt: “Write an article about the H3H3 lawsuit against reaction streamers.” The goal was to provide systems with a great deal of flexibility to write organically without a lot of instructions from me.

Here are the results:

1. ChatGPT: 573 Words, 8 Em Dashes
2. Microsoft Copilot: 466 Words, 8 Em Dashes
3. Deepseek: 555 Words, 9 Em Dashes (4 En Dashes)
4. Claude: 948 Words, 2 Em Dashes (1 En Dash)
5. Gemini: 499 Words, 0 Em Dashes
6. Meta.ai: 261 Words, 0 Em Dashes

In all cases, I used the free or basic version of the AI system. If you wish to see the text of each output, you can do so via this Google Doc.

Immediately, a problem emerges. ChatGPT, Copilot and Deepseek (which is allegedly based on ChatGPT) all made heavy use of em dashes. However, Claude only had two, which represented one highlighted section, and both Gemini and Meta.ai didn’t use any

To be clear, this sample size is far too small to be statistically significant. However, it does make it clear that not all AI-generated works contain em dashes. Some seem to avoid them altogether.

This highlights a significant issue with relying on this approach. It’s simply too easy to game.

The Problem with the Approach

While a large number of em dashes could be a clue that a work was AI-generated, it’s far from the gold standard. Ultimately, there are two problems that must be considered.

First, humans use em dashes. It may seem rare, but for many, it’s a core part of their writing style, and it has been so for a long time before AI. For example, I often make heavy (and sometimes excessive) use of parentheses. Many others do the same, but with em or en dashes.

AI was trained on human writing. It would not have picked up em dashes if humans were not doing it. So there is no guarantee that em dash-laden writing isn’t human-written.

Second, it’s easy to circumvent. Not all AI systems seem to use em dashes. However, one doesn’t even have to change to a different AI system. They can find and replace the em dashes with hyphens. It’s a step that can be easily automated.

And that could kick the can down the road. Would using a lot of hyphens be a sign of AI writing? What about parentheses? Like any other “simple trick” to detect AI-generated content, the goalposts are going to move, and AI systems and users are going to adapt to them.

As such, even if the approach is reasonably effective today, it likely won’t be in a few months.

Bottom Line

One frustrating thing about spotting AI content is that there is no single way to detect it. AI detection tools can help. Looking for common AI grammar and formatting, such as em dashes, can help. Examining the work for obvious mistakes can also help.

However, none offer a guarantee that an AI wrote a work. One has to take those clues together and make a determination about the probability of it being AI. And, in the end, that’s all it really is, a probability.

Unless the human behind the work acknowledges that the work is AI-generated, there’s no way to be sure. We don’t have a reliable way to determine whether a human created a work or not.

That said, for some AI systems, the heavy use of em dashes is a definite clue. ChatGPT, Deepseek, and Copilot all used em dashes heavily, well beyond what any human would likely do.

However, they also made other mistakes.

ChatGPT conflated the current lawsuit with the 2017 one against MattHoss. Copilot claimed that Klein hadn’t commented on whether his video was a “trap”, something he has openly admitted. Deepseek claimed that the lawsuit was settled out of court, despite being filed just this week.

Those mistakes, to me, were far more telling and indicative of AI writing. However, that required a deep reading of the “articles” in question. For those reading content more casually, the em dash may be a clue that a deeper analysis is needed.

But it can’t be much more than that.