50% Discount
At the supermarket, a good-looking salad bowl has a sticker on it: 50% off, probably because of the sell-by date. When presented at the check-out, the POS terminal reads it as $7.89, ignoring the discount. Normally the computer would adjust the price after scanning the sticker, but this time the human is left to their own devices. The cashier is a little flustered. The $7.89 original price is labelled right next to the 50% off. But how much is 50% of $7.89? They fumble with their phone for a while to find a calculator. I suggest, estimating, $3.95 ought to do it. Eventually, the calculator does its thing and says $3.945. After another hesitation, I get rung up for $3.94.
Does the fact that the cashier used a machine, rather than just their brain, to come up with the price, have any bearing on the sale?
Enter the world of LLM machines generating text for humans. For some this presents the same dilemma as a cashier not being able to do basic math without a machine. While using calculators has since long entered the realm of acceptable assistance, getting ChatGPT to write for you is not yet a citizen of that world, especially in education.
Imagine what’s coming when the lines between human and machine generated text blur even more or disappear altogether just like the outcome of a sum calculated by a human can be indistinguishable from one computed by a calculator.
Em-dash
You, mostly, can not tell if a calculator was used to arrive at the outcome of a sum, but for a brief while, people were sure there is a tell-tale sign, a text is generated by the machine, and thus could be suspect: Enter the venerable em-dash.
The prevalence of em dashes in AI-generated text stems from the fact that these punctuation marks are extremely common in the vast amounts of human-written material used to train artificial intelligence models.
Since AI learns by identifying patterns in its training data, and human writers frequently use em-dashes for emphasis, interruptions, or to set off parenthetical information, the AI naturally incorporates them into its own output. This is not a deliberate design choice by AI developers, but rather a consequence of the data it was trained on.
While some people believe the overuse of em dashes is a reliable indicator that text was generated by AI, this seems a misconception. The em dash is a legitimate and versatile punctuation mark used by skilled human writers for dramatic effect or to convey complex thought processes.
The issue arises when LLMs overuses the em dash, treating it as a default solution for pauses, interruptions, or adding information, often replacing simpler punctuation like commas or parentheses. This overuse can make the writing feel overly dramatic, monotonous, or “robotic,” even though the punctuation itself is grammatically correct. This pattern of overuse is a symptom of AI’s tendency to rely on statistically common patterns rather than nuanced human judgment.
Some users have reported that even when explicitly instructed to avoid em-dashes, Large Language Models will still insert them, suggesting they are deeply ingrained in the model’s default output patterns.
And consequently, humans are ditching em-dashes, not because they don’t like them, but because they’re terrified their writing will look fake. One commenter calls it out directly:
“the em dash is now a GPT-ism and is not advisable unless you want people to think your writing is the output of a LLM”
Community forums continue reporting the same pattern of too many em-dashes with no effective solutions. The issue remains unsolved.
A delightful Short by Ellen Cordova, presents a dialogue between 2 quotation marks on one side and the em-dash on the other:
Quotation marks: “Tell us about why ChatGPT is using you all the time?”
Em-dash: “I don’t want to talk about it. Instead, we could talk about how I am named after the great Emily Dickinson who used me a lot in her poems”.
Quotation marks: “Is that true? The em-dash was named after Emily Dickinson?”.
Em-dash: “No, that’s not true, but it feels like it could be true”.
“But you guys wouldn’t get it. You’ve never been in a relationship like Emily and I had”.
“Historians will say we were just desk mates, but that’s not the whole truth”.
Quotation mark: “Ah, it was a love story?”.
Em-dash: “God, I miss her…”.
“And now you invite me here to talk about ChatGPT?”.
“After the loves that I have known?”.
“After the delicious lines of poetry, prose and dialogue, that I carried on my back?”.
“ChatGPT, the monster, that gorges itself on Emily Dickenson, Virginia Wolf, Toni Morrison and all the great wielders of the em-dash,”.
“You want to blame me for getting stuck in ChatGPT’s teeth as it vomits them all back out, half digested in a pile of slop?”.
“Well, forgive me, but I don’t find that a scintillating topic of discussion”.
So far, it turns out, identifying a text as having be entirely created by LLMs, and give it perhaps some sort of watermark, is not as simple as counting em-dashes. A lot of (plagiarism and LLM detection) tools exist and are used especially in education, but they produce a high number of false positives (or false negatives). Their accuracy is not great and many a student complain about their teacher incorrectly accusing them of not having written an assignment themselves.
As an aside: Wikipedia is offering a Guide on detecting AI writings and yes, it addresses the em-dash.
The question this poses is what this all means next, not just for writing, but for authenticity online and in print. If a tiny piece of punctuation can create this much noise, just imagine what’s coming when the lines between human and machine generated text blur even more or disappear altogether, just like the outcome of a sum calculated by a human can be indistinguishable from one computed by a calculator.