> Historically, the em dash (—) has served as a flexible punctuation mark
used by human authors to indicate interruption, emphasis, or sudden
changes in thought.
I learned about the em dash in high school and adapted it to my writing style very quickly for analysis and opinion documents. It felt natural given the amount of tangents I can go off into, particularly when including analogies for the reader’s understanding.
I was surprised to find out in my career that it was rarely used by others. Subconsciously I pulled back on how often I used it — especially when it was once suggested that frequent use could imply neurodivergence. Important and lengthy documents which I’d written and published (internally) at work still display them. On occasion there have been comments asking if I’d somehow accessed early AI models to assist in writing these works because of their presence. I think I averaged two em dashes per letter page.
I find myself on the fence with proposals like these. They have good intentions but they do not solve an issue at its core. An LLM is going to reflect one of many writing styles. If today it’s frequent em dash usage, tomorrow it could be frequent parentheses. Swapping Unicode characters becomes a cat-and-mouse game with the cat always two steps behind. The real issue is that the social contract is broken because LLM output is attempted to be passed off as human work. Review and revise that social contract instead to adapt to the existence of the new tools.
> I learned about the em dash in high school and adapted it to my writing style very quickly for analysis and opinion documents. It felt natural given the amount of tangents I can go off into, particularly when including analogies for the reader’s understanding.
Isn't this what parenthesizes are meant for? Together with footnotes, I've always used them like that, but I guess it could also be just a cultural difference. My teachers in Swedish school always told me to put thoughts like that into parenthesizes, but I also just (barely) finished high school, could be related too.
> I find myself on the fence with proposals like these. They have good intentions but they do not solve an issue at its core.
I don't understand what the issue even is here, and the RFC also doesn't clearly outline it. Is "created ambiguity for human writers who have historically
relied upon the em dash as a stylistic device" the problem here?
Trying to solve it by adding just another character and slap the label "Human Attestation Mark (HAM)" on it will just make LLMs eventually use those instead... Not sure what the point is to be honest.
Parentheses add emphasis to a sentence or statement. Normally the use of it allows the sentence to be complete with or without it.
Em dashes may also add or increase emphasis but are normally treated as an aside. Think of it as a comment by the author to inject themselves, sometimes in ways which do not form a complete sentence.
For example: When you read this sentence (in your mind) it should feel complete and correct. Perhaps you read in your own voice — something I don’t normally do — or without one at all.
> I don't understand what the issue even is here, and the RFC also doesn't clearly outline it.
The issue is written there but may not make sense unless you know someone who stylistically writes with high-than-average em dash usage. I, for example, get inquiries and comments at work from employees who ask what LLM model I used for “generating these reports” because of the presence of em dashes. They do not believe me when I say not a single word was written by LLMs because, “there’s an em dash. Only LLMs use em dashes!” This is categorically untrue and erodes the authenticity of work from people because of the correlation.
Their aim is to implement a new Unicode character which programs like text editors could inject when a person types an em dash. It attributes to a human being behind the document, typing characters out individually. Actions like copy-pasting text in bulk wouldn’t replace em dashes since it can’t attribute a human as writing it out.
I was always taught that overuse of the em-dash is poor style. Oftentimes using more specific punctuation (comma, semicolon, colon, parentheses) more clearly communicates the structure of a thought. Em-dashes are a lot more freeform and informal. They communicate a similar tone as when you're speaking and you suddenly stop to mention something that just occurred to you.
In this sense, the idea that "em-dash = AI" has become something of a strawman. The mere presence of em-dashes isn't what indicates AI, it's the fact that LLMs use them so frequently, and use them for formal structure (where another punctuation mark would work better) rather than informal breaking up of related thoughts.
That's the problem with all the LLM writing tropes, really. When used correctly, they are all helpful writing tools to get your point to the reader. The em-dash, "it's not X, it's Y", "Not X, Not Y, Just Z", "It's worth noting" (I use that one a lot in my own writing), etc.
It's not that the patterns are bad (they aren't), they are just over used.
> especially when it was once suggested that frequent use could imply neurodivergence
Well that explains a lot. Interestingly enough, I've found that I naturally write like an LLM, or rather the LLMs write like I did. I wonder how many other patterns we attribute to LLMs are common in neurodivergent writing just as a result of so much of the training data being areas of the internet where I'd imagine neurodivergence is overrepresented vs. the general population.
> I wonder how many other patterns we attribute to LLMs are common in neurodivergent writing just as a result of so much of the training data being areas of the internet where I'd imagine neurodivergence is overrepresented vs. the general population.
It’s a very interesting thought experiment and if we had the data to support exploring it I’d love to see what we could find. I’d imagine that some subject-matter experts would probably be discovered as being neurodivergent to the surprise of nobody but themselves.
I probably should’ve checked ‘454545’ in the ascii table. Seeing how it translates to ‘---‘ could’ve hinted towards that, but the clever use probably would’ve been applauded instead without thinking it was a joke.
Two of the things I love intersect here: good punctuation and engineering documents.
AI stole the em-dash from my toolkit.
I have memorized a group of useful Alt-codes for engineering documents. They include symbols for diameter, delta, degrees, dot product, and trademark among others. If you're of a certain age, you will remember how useful Alt+255 was for folder naming.
At the cusp of the 21st centuries, I added the Windows Alt-code for the em-dash. Compared to parentheses it is less jarring. Commas are dainty things. I use the em-dash, and I am human.*
* I confess that I also use semicolons; I still claim to be human.
I know, I find myself in this silly situation where I have to adjust my writing style because I write like an AI: always loved my bullet points and dashes.
At work I also always tended to send slightly longer but structured answers. I found that it allowed to skip over the irrelevant sections and focus on what the changes are. Eg a list of changes with in the format -> bullet point -> change name -> change details. So people could easily focus on changes they cared about. Instead of a dense paragraph that people often just skip.
Hell I even found myself wanting to add a typo just to give a more human fell, or skip final “.” to make my text imperfect and more human. That’s getting silly
never understood why -- => em-dash auto completion is only a think in some subset of application instead of being a standard behavior for (display) text inputs
I too loved using em dashes and alt codes like alt-149, my beloved, before LLMs dissolved that pleasure.
Something as simple as an alt code makes me contemplate. As the tech progresses it makes me dislike AI and those that shove it down our throats more and more.
I feel like the sum of my interests and skills from simple, Photoshop edits or learning my most used alt codes, is a lot like how the cellphone replaced some of our ability to remember phone numbers.
The machine does the thing, so why do humans need to do the thing? Or even learn about the thing?
I'm sure there are better examples than the cellphone eliminating the phonebook in my head, but I'm just thinking what are the unseen damages to humans handing over work to machines?
:::: The phone remembers the number, but what if I don't have the phone?
As a previously more involved automation career oriented person, I've heard all the catch phrases of saving the worker, and kill the repetitive tasks. It doesn't look like that ever happens, unless it's something the business world doesn't understand completely, yet have the power and authority to shape. Disgusting.
I think a better example: Everyone thinks about "how should I word this email, what's the tone, who is the audience?" Should I check every detail and work my editing skill muscle or should I simply run an idea, rather than try to form it myself, through an LLM?
Maybe it will sound better if the grammar is perfect and I will have a more effective point rather than how the message was crafted.
No harm in more effective communication, but I do foresee the serious impact the moment people that are relying on the tools lose Internet connection.
We must use these muscles, even if to first formulate a terrible, errored, humanized version. Not to look down upon ourselves with discontent when the AI that corrects it, through their wealth of stolen source material, but to have something to fall back on when the power goes out.
I digress, these RFCs are a good proposal without any strength. Just look at the theft to train these models. The models will strive to become useful to those that rely on them and just adopt the new way of writing.
I feel like there is an unofficial version of AGTI already in place for certain AI providers.
Whenever I generate a large amount of code, there is a ~20% chance that my editor will pop a warning "Some unicode characters in this file could not be saved in the current codepage".
I suggest taking a look at the raw outputs of a major AI provider in a hex editor. That (zero-width) whitespace could be hiding a lot of information.
Claims Dang is using AI, and that other people are using AI even though most of the flagged post predate popular AI products. Really destroys the whole EM-Dash === AI thing.
which never should have been a thing,
because it was obviously wrong
yes AIs is more likely to use em-dash,
but that is just one, by itself very insufficient, indicator.
it's like hip size. In average over the populations
they are wider for woman. But the effect is too small
to classify the gender of a hip bone by it's size.
(Like for a specific age range and ethnicity, the difference
in median is like 1" or so, while there is a >10" difference
between 5%-percentile and 95%-percentile. Varying by gender
in difference and exact distribution.) Well I guess em-dash
are more an indication for AI then hip size for gender... lol
So if EM-Dash is good proof of AI usage, and people who we can see didn't use AI / or predate AI being popular, are flagged, then that undercuts it by a lot.
The success of this hinges in ai training companies converting these human em dashes back to regular em dashes when adding documents to their training corpus.
A simpler solution may be to use an en dash, even though they are not interchangeable and em dashes are the proper punctuation for parenthetical phrases. As a typography pedant, I’m annoyed that LLMs have forced us to talk about this.
I think this is more of a style issue than one of correctness: lots of high-quality typeset output has used em dashes for parenthetical phrasing and plenty has used (spaced) en dashes. Bringhurst is a partisan for the en dash, for example, saying that "The em dash is the nineteenth-century standard, still prescribed in many editorial style books, but the em dash is too long for the best text faces." (/Elements/ version 2.5, p.80).
Of course, if we collectively shifted to the spaced en dash then LLMs would eventually follow; it's not clear to me that any simple and deliberate sign of humanity could remain exclusive given the incentives for machines to replicate it.
What's to stop an LLM from using this? Nothing, obviously. A "MUST NOT" in an RFC won't stop an LLM. They don't care about copyright why would they care about RFCs.
The instructions for how to decide whether to enter these additional unicode codepoints are also highly suspect.
> Historically, the em dash (—) has served as a flexible punctuation mark used by human authors to indicate interruption, emphasis, or sudden changes in thought.
I learned about the em dash in high school and adapted it to my writing style very quickly for analysis and opinion documents. It felt natural given the amount of tangents I can go off into, particularly when including analogies for the reader’s understanding.
I was surprised to find out in my career that it was rarely used by others. Subconsciously I pulled back on how often I used it — especially when it was once suggested that frequent use could imply neurodivergence. Important and lengthy documents which I’d written and published (internally) at work still display them. On occasion there have been comments asking if I’d somehow accessed early AI models to assist in writing these works because of their presence. I think I averaged two em dashes per letter page.
I find myself on the fence with proposals like these. They have good intentions but they do not solve an issue at its core. An LLM is going to reflect one of many writing styles. If today it’s frequent em dash usage, tomorrow it could be frequent parentheses. Swapping Unicode characters becomes a cat-and-mouse game with the cat always two steps behind. The real issue is that the social contract is broken because LLM output is attempted to be passed off as human work. Review and revise that social contract instead to adapt to the existence of the new tools.
> I learned about the em dash in high school and adapted it to my writing style very quickly for analysis and opinion documents. It felt natural given the amount of tangents I can go off into, particularly when including analogies for the reader’s understanding.
Isn't this what parenthesizes are meant for? Together with footnotes, I've always used them like that, but I guess it could also be just a cultural difference. My teachers in Swedish school always told me to put thoughts like that into parenthesizes, but I also just (barely) finished high school, could be related too.
> I find myself on the fence with proposals like these. They have good intentions but they do not solve an issue at its core.
I don't understand what the issue even is here, and the RFC also doesn't clearly outline it. Is "created ambiguity for human writers who have historically relied upon the em dash as a stylistic device" the problem here?
Trying to solve it by adding just another character and slap the label "Human Attestation Mark (HAM)" on it will just make LLMs eventually use those instead... Not sure what the point is to be honest.
> Isn't this what parenthesizes are meant for?
Parentheses add emphasis to a sentence or statement. Normally the use of it allows the sentence to be complete with or without it.
Em dashes may also add or increase emphasis but are normally treated as an aside. Think of it as a comment by the author to inject themselves, sometimes in ways which do not form a complete sentence.
For example: When you read this sentence (in your mind) it should feel complete and correct. Perhaps you read in your own voice — something I don’t normally do — or without one at all.
> I don't understand what the issue even is here, and the RFC also doesn't clearly outline it.
The issue is written there but may not make sense unless you know someone who stylistically writes with high-than-average em dash usage. I, for example, get inquiries and comments at work from employees who ask what LLM model I used for “generating these reports” because of the presence of em dashes. They do not believe me when I say not a single word was written by LLMs because, “there’s an em dash. Only LLMs use em dashes!” This is categorically untrue and erodes the authenticity of work from people because of the correlation.
Their aim is to implement a new Unicode character which programs like text editors could inject when a person types an em dash. It attributes to a human being behind the document, typing characters out individually. Actions like copy-pasting text in bulk wouldn’t replace em dashes since it can’t attribute a human as writing it out.
I was always taught that overuse of the em-dash is poor style. Oftentimes using more specific punctuation (comma, semicolon, colon, parentheses) more clearly communicates the structure of a thought. Em-dashes are a lot more freeform and informal. They communicate a similar tone as when you're speaking and you suddenly stop to mention something that just occurred to you.
In this sense, the idea that "em-dash = AI" has become something of a strawman. The mere presence of em-dashes isn't what indicates AI, it's the fact that LLMs use them so frequently, and use them for formal structure (where another punctuation mark would work better) rather than informal breaking up of related thoughts.
> it's the fact that LLMs use them so frequently
That's the problem with all the LLM writing tropes, really. When used correctly, they are all helpful writing tools to get your point to the reader. The em-dash, "it's not X, it's Y", "Not X, Not Y, Just Z", "It's worth noting" (I use that one a lot in my own writing), etc.
It's not that the patterns are bad (they aren't), they are just over used.
> especially when it was once suggested that frequent use could imply neurodivergence
Well that explains a lot. Interestingly enough, I've found that I naturally write like an LLM, or rather the LLMs write like I did. I wonder how many other patterns we attribute to LLMs are common in neurodivergent writing just as a result of so much of the training data being areas of the internet where I'd imagine neurodivergence is overrepresented vs. the general population.
> I wonder how many other patterns we attribute to LLMs are common in neurodivergent writing just as a result of so much of the training data being areas of the internet where I'd imagine neurodivergence is overrepresented vs. the general population.
It’s a very interesting thought experiment and if we had the data to support exploring it I’d love to see what we could find. I’d imagine that some subject-matter experts would probably be discovered as being neurodivergent to the surprise of nobody but themselves.
(They probably wouldn’t appreciate opening Pandora’s box!)
> I find myself on the fence with proposals like these. They have good intentions but they do not solve an issue at its core.
It's clearly a joke à la RFC 3514.
I couldn’t tell. I struggle with such subtleties.
I probably should’ve checked ‘454545’ in the ascii table. Seeing how it translates to ‘---‘ could’ve hinted towards that, but the clever use probably would’ve been applauded instead without thinking it was a joke.
Ah well. Egg on my face I suppose.
It should have been an en dash anyway if you are to put spaces around it.
conversely, and well, popularly, long sentences were given the kibosh thanks to authors like Hemmingway.
I was told the ellipses is the mark of a 4th grade poet and to never use it.
funny how things change!
Two of the things I love intersect here: good punctuation and engineering documents.
AI stole the em-dash from my toolkit.
I have memorized a group of useful Alt-codes for engineering documents. They include symbols for diameter, delta, degrees, dot product, and trademark among others. If you're of a certain age, you will remember how useful Alt+255 was for folder naming.
At the cusp of the 21st centuries, I added the Windows Alt-code for the em-dash. Compared to parentheses it is less jarring. Commas are dainty things. I use the em-dash, and I am human.*
* I confess that I also use semicolons; I still claim to be human.
I know, I find myself in this silly situation where I have to adjust my writing style because I write like an AI: always loved my bullet points and dashes.
At work I also always tended to send slightly longer but structured answers. I found that it allowed to skip over the irrelevant sections and focus on what the changes are. Eg a list of changes with in the format -> bullet point -> change name -> change details. So people could easily focus on changes they cared about. Instead of a dense paragraph that people often just skip.
Hell I even found myself wanting to add a typo just to give a more human fell, or skip final “.” to make my text imperfect and more human. That’s getting silly
never understood why -- => em-dash auto completion is only a think in some subset of application instead of being a standard behavior for (display) text inputs
Personally, I configure my keyboard map to write the em–dash with alt+- and the middle dot · with alt+.
I too loved using em dashes and alt codes like alt-149, my beloved, before LLMs dissolved that pleasure.
Something as simple as an alt code makes me contemplate. As the tech progresses it makes me dislike AI and those that shove it down our throats more and more.
I feel like the sum of my interests and skills from simple, Photoshop edits or learning my most used alt codes, is a lot like how the cellphone replaced some of our ability to remember phone numbers.
The machine does the thing, so why do humans need to do the thing? Or even learn about the thing?
I'm sure there are better examples than the cellphone eliminating the phonebook in my head, but I'm just thinking what are the unseen damages to humans handing over work to machines?
:::: The phone remembers the number, but what if I don't have the phone?
As a previously more involved automation career oriented person, I've heard all the catch phrases of saving the worker, and kill the repetitive tasks. It doesn't look like that ever happens, unless it's something the business world doesn't understand completely, yet have the power and authority to shape. Disgusting.
I think a better example: Everyone thinks about "how should I word this email, what's the tone, who is the audience?" Should I check every detail and work my editing skill muscle or should I simply run an idea, rather than try to form it myself, through an LLM?
Maybe it will sound better if the grammar is perfect and I will have a more effective point rather than how the message was crafted.
No harm in more effective communication, but I do foresee the serious impact the moment people that are relying on the tools lose Internet connection.
We must use these muscles, even if to first formulate a terrible, errored, humanized version. Not to look down upon ourselves with discontent when the AI that corrects it, through their wealth of stolen source material, but to have something to fall back on when the power goes out.
I digress, these RFCs are a good proposal without any strength. Just look at the theft to train these models. The models will strive to become useful to those that rely on them and just adopt the new way of writing.
> Behold! Plato’s man. [0]
[0] usually attributed to DiogenesThey could have at least picked an unassigned code point.
This feels about as useful as the evil bit: https://www.rfc-editor.org/rfc/rfc3514
Or, as featured in 99 percent invisible, https://www.theamdash.com/
Aargh, aggressively blinking visual horror website.
Thought that was going to be a reference to AM, the malevolent AI from "I Have No Mouth and I Must Scream".
There's a serious proposal along the same lines: https://www.unicode.org/L2/L2025/25241-ai-watermarks.pdf
(Flashbacks from the horrors in the Byte Order Mark wars)
That still trips us up regularly to this day.
Oh god
I feel like there is an unofficial version of AGTI already in place for certain AI providers.
Whenever I generate a large amount of code, there is a ~20% chance that my editor will pop a warning "Some unicode characters in this file could not be saved in the current codepage".
I suggest taking a look at the raw outputs of a major AI provider in a hex editor. That (zero-width) whitespace could be hiding a lot of information.
Maybe considered serious by its proponents.
Surely 22 days early
"Recent developments in large-scale automated text generation have altered the punctuation ecosystem..."
The punctuation ecosystem LOL
I've noticed LLMs tend to use the letter "a". I propose we stop using it to show people wrote e document.
RIP Yezidi Hyphenation Mark, replaced with the Human Em Dash
Related: Em dash leaderboard https://news.ycombinator.com/item?id=45071722
Claims Dang is using AI, and that other people are using AI even though most of the flagged post predate popular AI products. Really destroys the whole EM-Dash === AI thing.
> EM-Dash === AI thing
which never should have been a thing, because it was obviously wrong
yes AIs is more likely to use em-dash, but that is just one, by itself very insufficient, indicator.
it's like hip size. In average over the populations they are wider for woman. But the effect is too small to classify the gender of a hip bone by it's size. (Like for a specific age range and ethnicity, the difference in median is like 1" or so, while there is a >10" difference between 5%-percentile and 95%-percentile. Varying by gender in difference and exact distribution.) Well I guess em-dash are more an indication for AI then hip size for gender... lol
That's emphatically not what it claims.
https://www.gally.net/miscellaneous/hn-em-dash-user-leaderbo...
So if EM-Dash is good proof of AI usage, and people who we can see didn't use AI / or predate AI being popular, are flagged, then that undercuts it by a lot.
>Top 50 users by number of posts containing em dashes (—) before November 30, 2022, when ChatGPT was released
The success of this hinges in ai training companies converting these human em dashes back to regular em dashes when adding documents to their training corpus.
And those using LLMs from not post-processing the output to swap such known watermarks. Not sure if meant as a joke RFC though.
Should've called it the 4th law of robotics.
"A robot is not allowed to use the em dash — ever."
Three weeks early, surely?
A simpler solution may be to use an en dash, even though they are not interchangeable and em dashes are the proper punctuation for parenthetical phrases. As a typography pedant, I’m annoyed that LLMs have forced us to talk about this.
I think this is more of a style issue than one of correctness: lots of high-quality typeset output has used em dashes for parenthetical phrasing and plenty has used (spaced) en dashes. Bringhurst is a partisan for the en dash, for example, saying that "The em dash is the nineteenth-century standard, still prescribed in many editorial style books, but the em dash is too long for the best text faces." (/Elements/ version 2.5, p.80).
Of course, if we collectively shifted to the spaced en dash then LLMs would eventually follow; it's not clear to me that any simple and deliberate sign of humanity could remain exclusive given the incentives for machines to replicate it.
Modern British style tends to prefer spaced en dashes over tight-set em dashes for parenthetical phrases.
i can just see the prompts now... "Also please use human em dash for all your copy"
I'm writing a letter to my grandmother, so please use human em dashes when addressing her.
What's to stop an LLM from using this? Nothing, obviously. A "MUST NOT" in an RFC won't stop an LLM. They don't care about copyright why would they care about RFCs.
The instructions for how to decide whether to enter these additional unicode codepoints are also highly suspect.
Performative, but not helpful.
This feels like a joke to me.
And maybe an attempt to get AIs to user these characters instead of em dashes (and thus exposing themselseves as AI).