I Saw Goody Proctor Talking to Claude: When AI "Detection" Backlashes

"This is obviously AI."

I've seen or heard variations of that phrase a lot in the last couple years. They span media from writing and still photographs to audio and video content; people have invoked them for ostensible war footage, ads for local RenFaires, social media comments, and reviews. Where AI is concerned, many of us are playing a giant game of Clue, picking up enough cards to convince us that we can accuse Mr. Grok in the datacenter with the six-fingered hand.

Those accusations have gotten harder as AI has gotten more advanced: you're probably not going to spot a nightmarish but telltale mouth in the middle of a forehead these days. The lines blur, but many experts and "experts" remain confident in their ability to spot the LLM influence.

Writing is no exception, and that's where I've seen the most allegations. So far, I haven't faced any myself (I can assure my readers that I produce only the finest, hand-crafted artisanal books of vaguely historical debauchery, if you're curious), but they come up a lot in the publishing community and specifically in the romance genre. They're also common when it comes to visual novels, one of my many nerdy hobbies.

The real deal

Let's get one thing out of the way first: some authors have absolutely used AI in their work. Some admit to it freely, like "Coral Hart," an author with dozens of pseudonyms and a school for AI romance. Her argument: to beat other authors, she needs volume, AI can speed up her writing, and, well, insert vaguely libertarian greed-is-good, just-obeying-the-market sentiments here.

Lena McDonald and K.C. Crowne assumed defensive postures after readers found and circulated LLM prompts they'd actually left in their published work. (A self-interested bit of advice: hire an editor next time, hon.) They say they use AI for edits (heh), refining their work, or developing ideas.

Hachette canceled Shy Girl after people started analyzing the prose for signs of AI, and in this case, the author admitted that the editor she hired for the self-published version used LLM assistance.

This is also true on the visual novel side, where most companies have used AI to enhance or polish art for a while. While none of the major studios have admitted to putting AI-generated text in the final product (aside from those specializing in "virtual partners," where it's the whole gimmick), the fact that AI-centered developer Series Entertainment acquired Pixelberry, the company behind US visual novel titan Choices, has not exactly boded well for those who want to avoid the technology and its products.

I absolutely understand why people are wary. Unlabeled AI writing is a problem, and there are even solid arguments against the labeled version, especially once major studios and publishers get involved. Corporations will take any excuse to cut corners; if that means replacing employees with LLMs that don't ask for overtime or pension plans, it's naïve to think they'd take the high road or that they'd inform consumers when they did start switching in AI.

That said

As I and my illustrious fellow author Nick Rogers have pointed out a few times in this very journal, AI draws from human writing. It uses em-dashes? Sure, so does Emily Dickinson. Triads? Take it up with the Bible, sport. Scratch any telltale "sign of LLM" and you'll find a construct that people had been using for quite a while before ChatGPT ever entered the scene.

After an…extensive…breakdown of "clues" that a new Romance Club visual novel came from AI (a scandal based in part on the writer not having much of a social media presence, OH THE HORROR), another user made an equally lengthy post pointing out how all of the "telltale signs" are classics like syntactic fragmentation, abstract metaphor, and parallel phrasing. (Oh no, I've just used a triad—my cover is blown! BEEP BOOP.)

Pangram reported that 60% of H.M. Wolfe's new novel, Daggermouth, comes from AI. Wolfe and publisher Simon & Schuster both vigorously deny it. Pangram's accuracy rate is pretty good, but it's hit false positives before.

Then there's the question that keeps coming up: what counts as AI writing?

The line

Let's note right off that some people don't want any AI model involved in any stage of the writing process, or indeed any part of human life. It's not my place to argue for or against that stance here, but it bears acknowledging as one place to draw the line: if you call up an LLM in relation to your work, it's AI-generated, no ifs, ands, or buts.

Beyond that stance, though, cases like Crowne's and McDonald's challenge us to define how much AI use is unacceptable. Can authors use AI for self-edits? How about as a "rubber duck" to bounce plot elements off of? Along the same lines, what about as a way to determine the basic plausibility of scenarios: if my character has to drive a truck down a country road after a killer fractures her skull, can I ask AI what factors would make her success more or less believable?

Those last two are, for me, particularly sticky. I sympathize with the argument that publication for pay isn't a human right: if you want other people to buy your work, you should hire and pay a human, join a critique group, or trade favors with a volunteer beta reader. (Archive of Our Own is down for at least two weeks as of the time of writing—you can probably find many talented people bouncing off the walls.) To the best of my knowledge, though, nobody makes their living listening to plot ideas or giving advice on very specific scenarios.

(Caveat: getting that sort of advice from many LLMs involves benefiting from information the original creators might not have agreed to put in a model and at least some amount of environmental impact. Choose wisely.)

The Shy Girl case brings up a different area of uncertainty: subcontractors. To the best of my knowledge, author Mia Ballard hasn't said whether or not she knew that her original editor was using AI and whether or not she gave said editor any guidance around it.

First of all, my opinion is that any editor who's charging for their services shouldn't go near an LLM with a project unless they have explicit permission, in writing, from the author. (If you're billing anything close to Editorial Freelancers Association rates, don't even think about it.) When I wrote above that you should find a human editor, I was assuming a modicum of professionalism.

I'll admit, though, that I'm going by general discussion here: the official rules haven't really been written as far as I'm aware. While various authorial organizations have started to establish labeling standards, the topic of how much the people you hire need to disclose—and how responsible you are for checking on them—is a wrinkle that will need to be ironed out.

Why here? Why them?

In the introduction, I mentioned that most of the AI writing allegations I've seen have related to romance novels or visual novel apps. Some of that's selection bias: I spend my time in those communities more than I do, say, forums devoted to thrillers or inspirational fiction. To correct that, I did a couple searches around AI allegations and novels, and I found a frankly interesting split.

On the one hand, there's romance or (in the case of Shy Girl) romance-adjacent horror or thriller: written for a popular audience, often for and by women. The other genre where AI suspicion has made the news has been award-winning literary fiction: Thélyson Orélien's novel It Was That or Die, for instance. The novel was removed from the Prix Goncourt longlist.

I can't say for sure what's going on here, but I have a couple theories.

One is pressure. The romance genre thrives on tight turnaround dates and lengthy series. While I spend less (by which I mean no) time in literary circles, my secondhand impression is that they're highly competitive, and contests often have strict deadlines. Both of those factors could drive an author to use LLM assistance; both could also inspire speculation when writers do keep up with the schedules.

Another is genre-specific linguistic tropes. Orélien says his work uses the "imagery, repetition and rhythm" of Caribbean and Haitian languages, which an AI detector trained on standard French might not be set up to deal with. While I'm by no means comparing genre romance prose to this standard in any other way, the genre does include a fair number of standard phrases, and has done so since long before AI. (There are only so many ways to describe, let's say, love and other indoor sports.)

Then there's prejudice. "There seems to be a troubling assumption that when a Black writer produces work that attracts significant attention, the work could not possibly be their own. Meanwhile, there are popular white writers who have publicly mentioned using AI for their writing processes and they're still doing fine," said Jerry Falade, author of Call Me, I'll Hide the Body, which publishers and agents dropped after unsubstantiated accusations. Orélien's initial accuser was a right-wing "anti-woke" columnist posting anonymously on Elon Musk's X.

As far as romance is concerned, well, it's notable that both Ballard and Wolfe are Black, while Hart and McDonald are white. (Crowne may be a syndicate, and for that I have no judgment: Carolyn Keene is a hero of our time.) The former two are facing widespread controversy, and Ballard's book has been canceled, while Hart and McDonald seem to have shrugged off public opinion.

Romance itself also gets a lot of judgment as a genre. (I'm not saying that this is all because it's largely by and for people other than cis het men, or because a lot of the aforementioned men seem to panic that it might give women higher standards, but I'm not not saying that.) Once you've dismissed a category of fiction as "trashy," "formulaic," and so forth, it's less of a leap to decide that a number of its authors must not be "really" writing their own books—even if you yourself read them. Maybe even more so if you feel the need to try and make the genre respectable by calling out bad examples.  

Which witch hunt is which?

Despite certain politicians' attempts to claim the phrase "witch hunt" for their own purposes in the last decade or two, it's got a long and honorable use, and I'm not letting them have it without a fight. When reasonable caution becomes a moral panic and accusations of bad behavior are weaponized, the term seems highly applicable.

"This just feels like AI" absolutely fits. I've seen it thrown around without any more reason than vibes, or used against authors whose work stepped on a particularly inflamed and fungal set of fandom toes. Suspicion of AI also opens the door to really hurtful behavior, as when National Eisteddfod judges mocked an entrant's work on live TV: after all, if you convince yourself the author used AI, then they're a bad person and you can be as mean as you want, right?

There's another way AI writing accusations fit the witch hunt label: they ask the subject to prove a negative. Without the use of some very creepy tracking software, there's no way to say for sure that any book written after 2022 or so didn't involve at least some AI. Absence of evidence is not evidence of absence, and so "this gives me the AI ick" is up there with "well, he kind of seemed like a Communist" in terms of statements anyone can make to try and incite the worst impulses in their fellows.

We're still hashing out where to draw the line with LLMs, how to label them, and what professional standards govern detection. Meanwhile, maybe the rest of us should hold fire until we know the facts. As one Reddit user says: "name one situation ever where witchhunters made the situation better."