I know this is being discussed everywhere, and I try to avoid bandwagons and the news of the day, but damn if this isn’t too worrying to let slide. Alexandra Alter writes for the NY Times (archived):
For months, speculation has been building online that a buzzy horror novel, “Shy Girl,” was written with the help of A.I. The novel, about a desperate young woman who is held hostage by a man she met online and forced to live as his pet, was self-published in February 2025. The book quickly found an audience among horror fans, and Hachette published it in the United Kingdom last fall and planned to release it in the United States this spring, billing it as “an unapologetic, visceral revenge horror novel.”
Earlier this year, Max Spero, the founder and chief executive of Pangram, an A.I. detection program, heard of the claims about “Shy Girl” and decided to run a test of the full text. Its results indicated that the book was 78 percent A.I. generated. “I’m very confident that this is largely A.I. generated, or very heavily A.I. assisted,” said Spero, who posted his research on X in January. […]
In response to questions from The New York Times about the A.I. allegations against “Shy Girl,” Hachette told The Times that its imprint Orbit has canceled plans to release the novel in the United States and that Hachette will discontinue its U.K. edition.
You can get more details at the link; they don’t really matter, and the facts about this particular novel don’t really matter — it’s clear that so-called “AI” (large language models) will soon be producing work that nobody will be able to prove did not come from a human mind. I realize some will say “Therefore AI is intelligent!” and others will say “Who cares? Let a thousand flowers bloom!,” but as an old-fashioned humanist I feel the foundations crumbling. Will we have to go back to telling tales by the campfire (while making sure the tale-teller isn’t plugged in)?
I found this story interesting enough to dig into, and while I couldn’t
stealfind an actual text of the opus, it seems pretty clear that it is to a great extent female-victim BDSM aimed predominantly at female readers, of the kind that is currently so trendy.I suspect that the potential readers tend not to be looking for the creative attention to evocative non-boilerplate detail that a good human author might bring to bear.
Having rather too readily (and in hindsight, quite unfairly) dismissed “romance” as a readily-automated genre in another thread, I would nevertheless say that masochistic horror porn probably does lend itself rather well to composition without soul.
I also agree with a comment about this that I stumbled across: the really horrifying aspect of this is not the AI-composition, but what this story reveals about Hachette, and modern publishers’, decision-making on what they take up. Never mind the quality, look at the sales …
The commenter (who was not hostile to the actual genre itself) felt that it was unlikely that anyone at Hachette had actually read the work at all. (This was based on their assessment of the actual contents, not simple free-floating cynicism.) The decision to publish was evidently made on quite other grounds. That‘s the kind of “automation” to be afraid of. And to condemn.
[I should perhaps explain for new readers that the book had been already self-published on Amazon, and sold very well. This adoption of successful self-published works and ingestion of their authors into the publishing machine, with suppression of the original, is apparently pretty common.]
While the sort of Grub Street hacks and drudges who traditionally produced stuff like this were of course human beings created in the image and likeness of God and thus endowed with inherent dignity regardless of how hacklike or drudgesque the texts they emitted, I’m skeptical that this particular subgenre of literature is the optimal one to insist on the ineffable primary of the Human against the Machine.
I am struck by an interesting legal question. At least in the US, the law is currently fairly clear (maybe surprisingly so) that only compositions by human authors are entitled to copyright protection, so an AI-chatbot-generated text is unprotectable and may thus be freely bootlegged or appropriated or republished by anyone without getting permission or offering to pay royalties. OTOH, a text that is 78% robot-generated but 22% human-authored may still be protectable as to the 22%, and since a potential copier may have difficulty knowing exactly which that is, interesting complexities naturally arise. (I can think of some aggressive arguments for the position that 100% should be free for the taking under those circumstances, but anyone who proceeds on that basis before the law is more settled will be taking a risk, even if they offer to pay me to defend them in any ensuing lawsuit.)
I’m skeptical that this particular subgenre of literature is the optimal one to insist on the ineffable primary of the Human against the Machine.
Well, if this is as far as it will go, that’s OK then.
I have no such confidence! We are currently at the point where AI chatbots can produce summaries of legal research in a particular area that are not *always* obviously worse than such summaries prepared at much greater expense by extremely intelligent young human beings who have recently graduated from prestigious law schools. They can’t yet produce the more complex and even (in their way …) quasi-literary texts that someone like me then generates using those summaries as part of the raw material, and I hope to be retired before they can … I do worry for my firstborn child who is currently a law student or at least I fret that the daily routine and tasks of younger law school graduates are in the process of being dramatically transformed with no one knowing for sure whether the resultant opportunities for the humans will be better or worse than they have been in the recent past or for that matter in the distant past of the 1990’s when I was where my daughter will soon be, in a career-track sense. At least she followed in my footsteps and learned ancient Greek as an undergrad so she’ll have that over the robots until someone thinks it cost-effective to change that.
I think that the time may have come for a Campaign for Real Pornography.
David E.’s CAMRA-snowclone references are only comprehensible to those of us Americans familiar with the wilder shores of 1990’s Scottish music. (I thought this fellow was Glaswegian, but wiki sez he primarily grew up in Dundee.) https://www.youtube.com/watch?v=TYZ-J2PubpU
If I have understood how LLM AI software is “trained”, then it seems that it will become ever harder to detect AI generated content. Those AI programs ingest mind boggling amounts of internet material.
Newer iterations or further training will require ingesting ever more newer internet content, and that newer content will include an ever expanding percentage of AI generated material, including what is appropriately called slop. Ponder the inherent self-cannibalism. More and more excrement ingested as a percent of total consumption will result in an endlessly greater proportion of purified—that is to say free of human generated content—excrement coming out the back end. Slop in, slop out, says Mr. Miyagi.
The training issue is well documented. There is a shortage of original data already, so the AI trains on itself, producing a loss of entropy, and the result is a loss of diversity in the output. In other words, the output becomes worse, not better in the future.
Specifically, these novels will look more and more like they are written in the house style, we get a collapse in narrative complexity, and the voice, arguably the most important part of a literary narrative, loses its rhythms and idiosyncrasies.
So what we see is a short-term boost for publishers which will disappear because AI eats its own seed corn.
Also worthy of mention is that none of this is actually making money. Open AI has something like 1.4 trillion dollars of debt instruments against about 15 billion in revenue last year. So that novel you generated is highly subsidized, which probably won’t continue forever.
The LLM industry is running out of genuine content to plagiarise; as their investment model is crucially dependent on convincing investors that enormous returns on their money will certainly-for-sure arrive at some point in the future, they literally cannot afford the idea to take root that the performance of APEs has effectively hit a limit. This accounts for both the push to enormous amounts of hardware, which will (they assert) produce AGI out of LLMs by Just Trying Harder, and the interminable apocalyptic hype about how LLMs will either “solve physics” or doom us all (both narratives serve the actual underlying purpose equally well.)
Meantime, we should not be distracted by the marketing fantasies, but keep our eyes firmly on what the “AI”-pushers are doing right now. Polluting the noosphere, destroying internet search, and making real information sources even more unviable by denying them web traffic by substituting “AI summaries” are not unfortunate side-effects. They’re part of the political project.
(Look at who the “AI”-pushers fund, and who enables them politically. This is no mystery.)
Google has just patented a technique for replacing actual websites by “personalised” “AI” reconstructions, so that even someone clicking through to what they suppose to be a company’s own site will actually be presented with a Google-generated replacement. The idea seems to be that companies will need to facilitate their own violation by presenting Google with the right kind of data for it to do its thing. Upping the ante on the
https://en.wikipedia.org/wiki/Search_engine_optimization
which has already done so much to hamstring any attempts to search for genuine reliable commercial information.
As if we might all benefit from the opinion of one reputed to have handled, if not read, a book towards the end of the last century…
Of course the robots will have been trained on AI generated content, so our li’l darlin’s will be taught right.
@DE, logically, what we need is good (artistic) porn. Which, as you say, would be inimitable.
Who said that powerful emotions are an obstacle and not an opportunity for the artist?
masochistic horror porn probably does lend itself rather well to composition without soul.
i always hate to disagree with DE, but i will again (at length, apparently). certainly there’s plenty of badly-written porn of all kinds, and always has been (asstr.org will give you a fine assortment of by-now-vintage material if you want it). but people who like their porn on the page are impressively discriminating readers – enough so that current mainstream popular fiction is marketed with reference to very detailed taxonomies that come directly out of the non-commercial literary porn world: “hurt/comfort”, “enemies to lovers”, etc.
while the taxonomic categories that migrate across in that way are about plot and scenario (because that’s what’s legible as usable to marketing departments), the reading communities they come from (usually glossed these days as just “fic”, because there’s so much original writing that “fanfic” no longer works as an umbrella) are probably the primary current cultural space where debates over literary style and quality are a pervasive part of mass-scale conversations. the number of people active in the comments on the main fic hubs (which are community-maintained as public* institutions) is easily in the millions – one of those hubs has over 10,000,000 registered users; another shows over 4,000 people currently on the site as i type (about 1/3 registered and 2/3 guests). they also have entire ecosystems of “beta-readers” whose purpose is specifically to help make writers’ work better. and the taxonomies active within them include descriptors for all manner of stylistic aspects of writing – a few from the first lexicon my search pulled up are “crack”, “fluff”, and “idfic” – because readers want to know what they’re getting into, whether the work in question is a 100-word holmes/watson “drabble” or the million-words-so-far original serial i’ve been reading since 2022.
now, not all of that is porn, and not all the porn involves either masochism or horror – but that entire ecosystem of non-commerical writing comes very specifically out of pornographic women’s fanzine writing: “slash” fiction**. here‘s a classic interview with joanna russ on the subject.
and – to get back to the original subject at last – my experience being occasionally in those worlds, and much more consistently adjacent to them, is that they are far more intensely hostile to LLM-generated writing than any other sphere i move in. that’s in part because it is an element of the parasitic relationship between “legitimate” commercial publishers and the world of fic, which is (correctly) understood as creating opportunities for bad writers to cash in individually, with none of the benefits accruing to the reading communities they come from. e.l. james (Fifty Shades of Grey, originally a Twilight fanfic) is the classic example. and it’s partly because porn is so personal, and especially for readers into sex that’s considered “perverse”, the felt presence of the writer in the work is indispensible. russ talks about some aspects of that in slash in that interview and elsewhere; i hope a friend of mine will someday find a place to publish her thoughts on the pedagogical function of john preston’s pornographic work (Looking For Mr. Benson, especially) in the consolidating gay leather scene of the 1970s.
which is perhaps all to say that what Hachette tried to do would never have flown with the core audience for “masochistic horror porn” – precisely because composition with soul is far more actively demanded by its readers than they could imagine. which isn’t to say that it couldn’t’ve made them (in liz phair’s phrase) shitloads of money, just that the source of that money would have been in large part readers unaware of how much high-quality writing is out there that caters to them, available for free (with or without a tip jar) in explicitly non-commercial spaces.
.
* “public” in its actual sense, “by, of, and for the people involved”, rather than as a euphemism for “state-controlled”.
** named for the “/” in “k/s”, which abbreviated “kirk/spock”, the original focus of writing.
This is stunningly stupid, will result in gigantic lawsuits, and might seriously damage Alphabet.
AI will revolutionise education. Of course.
Education in the form “one teacher communicating with 30 children by means of lectures (and also grades and assignments)” is easy to revolutionise.
Too many things children need that such a teacher can’t or won’t give them. And (strangely) also some things chudren don’t need that she both gives and forces them to take.
And not only children. A friend of mine is struggling with mathematical analysis (in her 30s and without much experience with math) and needs someone to talk about it. My ex-wife and I are willing to do that but we can’t do it face-to face, and she, i think, is too shy to ask 30 silly questions a day by phone. Other her friends either can’t or don’t want to talk about it.