On em-dashes and “AI detectors”
and the many reasons why I don’t and won’t use LLMs
Over the last couple months, I’ve seen an increasing number of writers make statements declaring that they do not use any AI in their writing. Neither do I, for a number of reasons. There is of course the devastating environmental impact this technology is likely to have in the coming years if the big AI companies have their way. I’m also deeply troubled by the quasi-religious beliefs that are driving some of the Silicon Valley billionaires who are pushing AI into everything, consequences be damned (btw, there is a key at the bottom of that essay explaining who holds which views).
Then there’s the fact that LLMs (Large Language Models, such as ChatGPT, Claude, and Gemini) were all trained on the work of writers without our consent or fair compensation. In fact, I’m part of the Anthropic class action law suit, as the company used pirated copies of my first two books, Whipping Girl and Excluded, to train Claude.
In addition to those ethical concerns, LLMs are essentially “word probability” machines that make statistical guesses as to what the correct response or output might be. This is why they are so prone to hallucinations, wherein they invent “facts” and “citations” out of whole cloth. This too has personally affected me: For a while, Google’s AI Overview incorrectly cited me as the person who coined the word “femmephobia,” which led The Atlantic’s Helen Lewis to use that AI hallucination in one of her anti-trans articles. (Neither she nor The Atlantic has published a correction, even though I contacted the magazine about it.)
Anyway, all of this sucks. But things have gotten even suckier for authors in the last month or so.
A few days ago, an author had their mega book deal revoked over questions of whether AI was used in the writing of the book (a charge the author denies). About a month earlier, a similar controversy occurred with the winner of the Commonwealth short story prize, when “social media users reported running his story through AI-detection software.” According to The Guardian’s reporting: “The story includes multiple ‘not x, but y’ constructions and lists of three, which some consider to be signs of AI use.” (The author denies the charge, but admits to using AI for research.)
The AI-detection software that was used to flag the story is Pangram, which Substack has recently added to its app to allow readers to scan for AI-written articles. This Bluesky thread (which is where I first learned about the Substack-Pangram partnership) includes comments from several authors who ran their own hand-crafted writings through Pangram and got flagged for being supposedly AI-generated. As one commenter said: “Pangram *is* AI.”
I admittedly cannot vouch for any of the aforementioned writers, as I’ve never heard of them before. But over the weekend as I was writing this, Kat Tenbarge—a writer I am familiar with and respect—released this video in response to her being falsely accused of using AI by [checks notes] Ozzy Osbourne’s son. Apparently, Osbourne thought an article Tenbarge wrote about the Tate brothers was written by AI because it included (in his words) “whatever they’re called, like the big hyphens” (aka, em-dashes). Tenbarge’s video includes proof (via Google Docs of drafts of her article saved along the way) that she wrote it organically. And she cites OpenAI itself who said, “AI writing detectors don’t work.”
So why does Jack “whatever they’re called, like the big hyphens” Osbourne care if some Rolling Stone article was written by or with AI? Well, I’d wager he doesn’t. Especially, since he doesn’t even know what em-dashes are! The most likely scenario is that he raised the issue because he is either 1) pro-Tate brothers and wanted to smear a journalist who wrote about the case against them, or 2) he knew a video smearing said journalist would be clickbait for the rabidly misogynistic Tate fanbase. (Notice how this is not a list of three, but two.)
In other words, we have entered the age where people who disagree with what authors have to say—the facts we raise and the arguments we make—will be inclined to simply claim that our articles and books were “written by AI,” regardless of whether they are or not.
For those of us who are legitimately concerned with the rise of AI-writing and how it is polluting our information ecosystem, I completely understand the desire to find a “tell”—like the proverbial “extra fingers” in AI art—but I doubt that will ever happen. Virtually every professional writer, at one point or another, has used “not x, but y” constructions, or made “lists of three.” Most writers use em-dashes—that link will take you to a passage from Whipping Girl, which was published in 2007, long before any of this LLM bullshit infiltrated our world. As I’ve argued before, AI writing uses em-dashes (and lists of three, and so on) because LLMs were trained on authors’ writings and we have long used all of these things!
Furthermore, even if a “tell” was ever found, AI companies could (and would) simply remove them via software updates. According to this article, this has already happened with em-dashes, which are greatly reduced in the latest versions of most LLMs.
Anyway, since we have seemingly entered a new era where randos who don’t know anything about writing will inevitably use em-dashes and “AI detectors” to smear articles and books they dislike as being “written by AI,” and given that many “gender critical”/anti-trans activists hate me and have tried to smear me many times before, it is inevitable that someone will eventually accuse me of this. So let me state for the record: I have never used LLMs in my writing, nor will I ever in the future.
And for those who might attempt to accuse me of such things, I will add that (like Kat) I save every draft for everything I write, so I can prove you wrong. Furthermore, on my writings webpage, I link to twenty-plus years of my online writings (which AI companies have no doubt scraped for their LLM training). If you examine them, you’ll find that while the content of my writing (the topics I cover) sometimes shift over time, my writing style (with all of its quirks) has remained fairly consistent.
If you are looking for a “tell” that I *don’t* use AI in my writing, there’s the fact that I often go a month (or two, or more) between publishing new essays, as they typically take me quite a while to produce. Many of them include citations (all of which are real and relevant) because I like to show my work and provide evidence to support my arguments. On top of that, as I say on my Patreon about page, “I prefer sinking my teeth into ‘201’ and ‘301’-level topics,” especially with regards to gender, sexuality, science, and social justice. Rather than repeating what everyone else has already said, I strive to say something new, inasmuch as possible. And LLMs are incapable of saying anything new, because they are “word probability” machines that spit out variations of things real-life authors have already said!
Finally, on top of all the aforementioned ethical concerns, the main reason why I would never use LLMs in my writing is that writing is thinking!
Like many writers, I often put together an outline before I start working on an essay—I’d imagine this is akin to a prompt that non-writers might use to guide an LLM to write the “perfect essay” for them. But the thing is, once I start writing, the outline almost always goes completely out the window. It will become obvious as I’m working through a piece that I actually need to make point C before point B. And I’ll realize that point D is superfluous or too much of a digression to include.
And more frequently than non-writers could possibly imagine, over the course of writing, it will strike me that point A—the main point I wanted to make going into the essay—is flawed, or incorrect, or not the true crux of the story. The essay often instead becomes about “point E,” which wasn’t even in my original outline. It’s only through countless hours of struggling to write and think through a complex topic that I am able to reach such a conclusion.
Like most people, I harbor lots of opinions that “seem correct” to me, despite the fact that I haven’t rigorously examined them. I almost never write about those, because why would I? But the beliefs that I am *most sure about* are the ones that I’ve spent many hours contemplating and dissecting while writing about them.
That is why I would never use LLMs in my writing. Because writing is thinking. And I don’t want LLMs to do my thinking for me.
This essay was made possible by my Patreon supporters — if you appreciate it, please consider supporting me there! You can also sign up for a paid or free subscription here: