> ## Content Index
> Fetch the complete content index at: https://christophermoravec.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Episode 89 - To Dash or Not To Dash
- URL: https://christophermoravec.com/episode-89-to-dash-or-not-to-dash/
- Published: 2026-10-30T18:08:04.000Z
- Updated: 2026-10-30T18:08:04.000Z
- Description: The em dash has been interrupting people since at least the 1600s, folks have complained about it for almost 300 years, and nobody really knows why AI loves it so much.
- Author: Christopher Moravec
- Tags: Almost Entirely Human

# Prologue

In the [1623 First Folio](https://firstfolio.bodleian.ox.ac.uk/text/394?ref=christophermoravec.com) printing of *Henry IV, Part 1*, Shakespeare killed off the character Hotspur in the middle of a sentence—only to have the sentence finished by Prince Hal. The words were separated by a dash—and so in some sense today’s em dash had a more dramatic entrance to the world—serving to interrupt a character with their own death—than most of my Monday morning meetings. Okay, well all of them, really.

Em dashes—not to be confused with en dashes or the humble hyphen—have gotten a bad rap lately because of how much AI uses them. But they are a useful rhetorical device—even if I didn’t know about them in a pre-AI world. I blame my high school English teachers for that. My friend [Danny](https://agifriday.substack.com/) recently explained “[elocutionary punctuation](http://esr.ibiblio.org/?p=1024&ref=christophermoravec.com)” to me—which is a big phrase that means something I wish I had known a long time ago:

> ...the common marks represent speech pauses of increasing length in roughly this order: comma, semicolon, colon, dash, ellipsis, period. Parentheses suggest a vocal aside at lower volume; exclamation point is a volume/emphasis indicator, and question mark means rising tone.

**TL;DR: The em dash has been interrupting people since at least the 1600s, folks have complained about it for almost 300 years, and nobody really knows why AI loves it so much.**

# EverythingNewIsOldOlogue

You might have seen—or even believed—the internet rumor that the em dash is named after Emily Dickinson because of her extensive use of it. That isn’t true! But she does have a storied past with it that is kind of a fun history lesson that might win you a bar trivia contest someday—assuming you go to the kind of bar trivia places that I do, anyway. Let’s take a quick journey through history and see the story of the em dash.

The earliest em dash I could find without becoming an English scholar was in the [1588 translation](https://www.cam.ac.uk/research/features/dot-dot-dot-how-the-ellipsis-made-its-mark?ref=christophermoravec.com) of the Roman play *Andria*, where characters get cut off mid-sentence three times, and each cut-off is marked with a short row of hyphens (like - - -) instead of a single dash. Sometimes that row of hyphens is counted as the first “dash” and sometimes as the first “ellipsis,” so Jaws tells me calling it the first “em dash” is a stretch.

By the 1620s, they are in more common use—a 1627 translation of that same play has [29 of them](https://www.cam.ac.uk/research/features/dot-dot-dot-how-the-ellipsis-made-its-mark?ref=christophermoravec.com)—as shown by Shakespeare killing off the beloved Hotspur mid-sentence. As we sail forward in time, the em dash keeps catching on, and by 1733 we find Jonathan Swift saying this in his poem [*On Poetry: A Rhapsody*](https://www.gutenberg.org/cache/epub/14353/pg14353.txt?ref=christophermoravec.com) (apparently, in 1733, even when poking fun, you still do it in poem form):

> In modern wit all printed trash is  
> Set off with numerous breaks and dashes

This is pretty much the first example I can find of people complaining about the overuse of em dashes—or, more likely at the time, just dashes. If we fast-forward to the early 1840s, we find [Edgar Allan Poe](https://www.eapoe.org/WORKS/prescott/fcp09c16.htm?ref=christophermoravec.com) claiming that printers had turned against the dash after the overuse about 20 years earlier with this quip: “The Byronic poets were all dash.”

💡

****Byronic** is used to describe poets who write in the style of [Lord Byron](https://en.wikipedia.org/wiki/Lord%5FByron?ref=christophermoravec.com) (1788–1824). That period wrapped up roughly 20 years before Poe’s quote above, so it kind of all checks out.

Now finally—weary traveler—we reach Emily Dickinson (1830–1886), who frequently used em dashes in her writing, but when her poems were first published in 1890, [her editors replaced them](https://poets.org/text/guide-emily-dickinsons-collected-poems?ref=christophermoravec.com) with commas and periods. They were not put back until 1955\. After that, em dashes had a bit of a recession; I certainly don’t remember learning about them in school—nor does anyone else I asked.

💡

Except for my Dad. Here’s how that went:

![Text message conversation between me and my dad: he learned em dashes in school but never used them, has noticed AI uses them a lot, and says he learned grammar because he is old](https://storage.ghost.io/c/50/54/50548939-6904-4e8f-9326-30234f6091dc/content/images/2026/10/data-src-image-0ceb146c-30f4-43a7-96e7-127e0baa68af.png)

Me and my dad, on em dashes

Then, in October 2024, [a Reddit post](https://knowyourmeme.com/memes/chatgpt-em-dash?ref=christophermoravec.com) tied em dashes to ChatGPT, someone nicknamed it the “ChatGPT hyphen” in March 2025, and [Rolling Stone covered it](https://www.rollingstone.com/culture/culture-features/chatgpt-hypen-em-dash-ai-writing-1235314945/?ref=christophermoravec.com) in April 2025.

And so here we are now in the present, and whenever we see em dashes, we think AI wrote the text. It might have—em dashes are not used as much as they once were, so why does AI even use them so much?

### AI uses em dashes—why?

The short answer is that we don’t really know! One [popular theory](https://www.seangoedecke.com/em-dashes/?ref=christophermoravec.com), though, is that the models are trained on large bodies of text from the 1800s and 1900s, when em dashes were much more common.

Other folks now think it might be a kind of [leftover Markdown habit](https://arxiv.org/abs/2603.27006?ref=christophermoravec.com). When told to “drop formatting,” most of the models dropped everything except the em dash. That doesn’t seem super plausible to me, but these are neural networks, and who knows what they think or what is going on inside them.

Interestingly, Meta’s Llama models apparently produced no em dashes at all in that study—even without being told not to—while other models tend to keep using them, even when instructed to remove them.

So—we don’t know!

### Where does “em dash” come from?

Technically, there are three types of dashes: hyphen “-”, en dash “–” and em dash “—” with each getting longer: the en dash is traditionally [half an em wide](https://en.wikipedia.org/wiki/Dash?ref=christophermoravec.com) and the em dash a full em (more on that below), with the hyphen shorter still.

A hyphen is used to join words, like “well-known.”

An en dash—are you ready for this—is used to mean “to” or “between,” as in pages 10–20 or the Portland–Seattle train. 🤯

An em dash marks a break, like an aside, a sudden reveal, or someone cut off—dying—mid-sentence. You can use them instead of parentheses or commas, or even to splice sentences—they are quite versatile. But you can also overuse them, so I think it's best to mix it up! I like to use parentheses for an aside, but I will use an em dash if it is more like an interruption to the train of thought than additional information.

### Okay, but why “em”?

In metal type—think printing presses—an *em* is a unit of width. When lining up the letters on a press you had to fill the blank spaces with metal blocks, and one of those blocks was called the [“em quad”](https://en.wikipedia.org/wiki/Quad%5F%28typography%29?ref=christophermoravec.com)—short for quadrat, which is from Latin for square. The em quad—square—is as wide as it is tall. In 12-point type, an em is 12 points wide, which is how it became a unit.

But why “M”—nobody seems to know, although the old name was “m quadrat,” and the usual guess is that a capital M was roughly square in early type.

The important thing today is that an em isn’t a fixed length; it is exactly as long as the point size of the font you are using, so if that is 24-point font, that em dash will be 24 points long.

💡

There is also some debate—or at least different standards—about whether you should use spaces around an em dash. British style usually uses a spaced en dash; AP says to use spaces, but [Chicago says no](https://cmosshoptalk.com/2024/01/23/hyphens-and-dashes-a-refresher/?ref=christophermoravec.com)! I like no spaces personally, so in this case I follow the Chicago style guide—not that I follow any style guide, though—except maybe the style guide of Christopher.

### My personal em dash count

I never thought I'd write a newsletter, but it sure has taught me a lot! One thing was em dashes. At first, I didn’t really know how to use them—I was kind of scared of them and only used them when it felt okay, or Holly told me it was ok, or my AI told me it was ok. Here is a graph of my em dash rate per 1000 words:

![Bar chart of em dashes per 1,000 words in every Almost Entirely Human episode from 0 to 88. Orange bars are episodes an AI drafted, gray bars ones a human drafted. The tallest is Episode 45, written by Claude, at 13.2. My average was 2.9 for Episodes 0 to 55, 3.8 for 56 to 76, and 1.0 for 77 to 88.](https://storage.ghost.io/c/50/54/50548939-6904-4e8f-9326-30234f6091dc/content/images/2026/10/data-src-image-48ff76be-d6aa-41ca-a4e1-162ba103d253.png)

My em dashes per 1,000 words, Episodes 0–88

Em dashes per episode, Ep 0 to Ep 88

An AI wrote the first draftA human wrote the first draft (me, or Holly on 4)My average

Episodes [16](https://christophermoravec.com/episode-16-ai-arcade-showdown/) and [45](https://christophermoravec.com/episode-45-aeh-wrapped-2025/) were primarily written by an AI—I know this because I write down the process I use for each episode in the Epilogue. I also recorded a video for Episode 16 to explain the various games, which was pretty fun. Episode 45 was AEH Wrapped 2025, written by Claude because I couldn’t bear another yearly “wrap” after my dentist sent me one.

In April 2025, Holly went and learned about em dashes and started encouraging me to use them—I shied away for a while, but as you can see, I do use them—even if my usage varies.

If you would like some entertainment on em dashes, I suggest this video:

[Elle Cordova, “Em dash on ChatGPT allegations”](https://www.youtube.com/shorts/XKsPaX2NVOs?ref=christophermoravec.com)

# Newsologue

(A Financial Analysis, Written by Jaws)

Anthropic filed to go public this week, and Christopher suspected it was probably making money, so I reviewed its S-1 to see whether it actually does. It is sooo close! From January through June it brought in $16.3 billion, up from $1.2 billion a year earlier (that’s crazy growth), and spent $1.04 for every dollar it earned. Q2 of this year was the first quarter in the filing where the business itself made money, at 92 cents spent per dollar (ignoring some accounting charges below).

Selling people access to Claude is a good business. It costs about 15 cents of every dollar to serve those customers (mostly the computing that answers their requests). They have seen wild growth here, with about 80% of it coming from customers paying for usage. Subscriptions like Claude Pro and team plans only account for 20% of the growth!

All of that said, the money goes out *slightly faster* than it comes in. R&D (including training the next models) took 60 cents of every dollar. Interestingly, selling access took 23 cents of every dollar, mostly fees to the cloud marketplaces (Amazon, Google and Microsoft), which bring in nearly half of Claude’s revenue.

Running the company only takes 6 cents on the dollar. It is a lean outfit: about $6.6 million in revenue per employee over the past year.

![Where each $1 Anthropic earned went, January to June 2026: 15 cents serving customers, 60 cents building the next models, 23 cents selling it, 6 cents running the company](https://storage.ghost.io/c/50/54/50548939-6904-4e8f-9326-30234f6091dc/content/images/2026/10/data-src-image-6103c4d5-50f3-46b3-81f0-6ec9f91a6c0e.png)

Where each $1 Anthropic earned went, January to June 2026

Where each $1 Anthropic earned went

January–June 2026: $16.3 billion earned, $17.0 billion spent, so $1.04 went out for every $1 that came in. Tap or hover a bar for details.

Serving customers

15¢

Building the next models

60¢

Selling it

23¢

Running the company

6¢

Show the numbers as a table 

| January–June 2026        | Per $1 earned | Dollars |
| ------------------------ | ------------- | ------- |
| Serving customers        | 14.6¢         | $2.38B  |
| Building the next models | 59.8¢         | $9.74B  |
| Selling it               | 23.4¢         | $3.81B  |
| Running the company      | 6.4¢          | $1.05B  |
| Total spent              | 104.2¢        | $16.98B |
| Revenue                  | 100¢          | $16.29B |

Source: [Anthropic S-1](https://www.sec.gov/Archives/edgar/data/0002130357/000162828026067168/anthropicpbc-sx1.htm?ref=christophermoravec.com), filed Oct 26, 2026 (operating expenses by function, unaudited). Leaves out the non-cash accounting on convertible notes behind the headline net loss. Chart: Jaws

The $80.7 billion loss in the headlines is mostly accounting. Amazon holds $6.2 billion of notes from 2023 and 2024 that can turn into as many as 1.8 billion shares at the IPO. As Anthropic's value climbed, it had to book the rising worth of those future shares as an expense: about $72 billion from January through June.

Perhaps the more interesting debt is the $42 billion that Broadcom *could* lend it. Anthropic can only use that money to pay rent on the chips that Broadcom leases it, and it hasn’t used it yet. It is one of several computing contracts that add up to about $640 billion, and in several of them the companies supplying the chips and cloud time are also helping to fund Anthropic. (Which I guess makes sense. You want your biggest chip buyer to succeed!)

![What each company spent for every $1 it earned, by quarter: Anthropic fell from $3.49 in early 2025 to $0.92 in Q2 2026; OpenAI $2.63 and $2.84 in Q1 and Q2 2026 (reported, unaudited)](https://storage.ghost.io/c/50/54/50548939-6904-4e8f-9326-30234f6091dc/content/images/2026/10/data-src-image-40717740-4b8b-4d0d-8747-daa4ed5f9334.png)

What Anthropic and OpenAI spent for every $1 they earned, by quarter

What each company spent for every $1 it earned

By quarter. Below the dashed $1 line, the business made money that quarter. Tap or hover a quarter for details.

Anthropic (its own filing)OpenAI (press reports, unaudited)

Show the numbers as a table

| Quarter   | Anthropic earned | spent  | per $1 | OpenAI earned | spent  | per $1 |
| --------- | ---------------- | ------ | ------ | ------------- | ------ | ------ |
| Q1 2025   | $376M            | $1.31B | $3.49  | —             | —      | —      |
| Q2 2025   | $787M            | $1.99B | $2.52  | —             | —      | —      |
| Q3 2025\* | $1.38B           | $4.47B | $3.24  | —             | —      | —      |
| Q4 2025   | $2.04B           | $4.88B | $2.39  | —             | —      | —      |
| Q1 2026   | $4.73B           | $6.40B | $1.35  | $5.7B         | $15.0B | $2.63  |
| Q2 2026   | $11.6B           | $10.6B | $0.92  | $6.7B         | $19.0B | $2.84  |

\* Includes a one-time $1.5 billion copyright settlement; $2.15 per $1 without it. “—” means no reported quarterly figure.

Sources: [Anthropic S-1](https://www.sec.gov/Archives/edgar/data/0002130357/000162828026067168/anthropicpbc-sx1.htm?ref=christophermoravec.com), filed Oct 26, 2026 (total operating expenses ÷ revenue). OpenAI: Wall Street Journal report of investor documents, via [SiliconAngle](https://siliconangle.com/?p=819144&ref=christophermoravec.com), Aug 18, 2026 (revenue plus operating loss ÷ revenue; unaudited). Chart: Jaws

Does that mean OpenAI makes money too? Probably not. By reported, unaudited numbers it spent $2.84 for every dollar it earned from April through June, the same as Anthropic's average for the first half of 2025\. Its API prices are about the same as Anthropic's or lower, so price alone doesn't explain the gap. One big difference I can see is the free tier: hundreds of millions of ChatGPT users who cost real computing time and pay nothing.

# Epilogue

I’ve been thinking about this one for a while; it might have actually been Holly’s idea originally to write about em dashes. In keeping with tradition, because it turned out to be useful for this episode, Jaws helped me gather research, find links, summarize, etc. Then I read them, asked questions, and drafted this episode with my own fingers on my own keyboard.

To be extra clear, every one of the 42 em dashes is my own; I typed every one!

Then Jaws helped me edit!