Lots of interesting stuff in this post from Jill Walker Rettberg, analysing the text of novels from popular Norwegian thriller author and confessed user of LLMs Jørn Lier Horst. (Tip of the hat to Uncle Warren for this one.)
For instance, there are certain tics that apparently serve as tells of LLM-generated text which I recognise in my own (stubbornly un-assisted) fiction writing, such as what Rettberg identifies as “binary oppositions that cancel each other out at the level of language”: think of a character description that juxtaposes their seeming to have a positive trait, but secretly harbouring another one that negates it.
I have tended to think of and describe people in this way for as long as I can remember, though the reliability of my memory could of course be brought into question; perhaps taking myself as an initial template, the assumption of contradiction and/or paradox as central to character goes way back. Rettberg is as close to a authority on this stuff as there is at this point, as I understand it, so I’m willing to believe it’s a feature of LLM output; but the models must have absorbed that tendency from the training data, as I must also have done (albeit by different means).
There’s another example in Rettberg’s post that is perhaps easier to hypothesise around, if only on the basis of what I will openly confess to be my prejudices. She notes that the more recent works by Horst have a higher frequency of nouns relating to parts of the body, as does the now-notorious horror novel Shy Girl, and this made me think that, when it comes to specifically fiction-oriented training data, online fan-fiction archives have surely been very important to the companies building these models; as we know, they weren’t really that bothered about the copyright issues with legitimately published works, but fanfic’s limbo of legitimacy (as well as its ease of access) must have made it a very obvious target for scraping.
I don’t think it particularly controversial to suggest that, considered as a whole, the corpus of fan-fiction is going to have a greater focus on the human body; a very great deal of it is, after all, written to the end of imagining familiar characters getting it on. Have LLMs, when prompted to draw upon specifically fictional precedents, thus picked up a statistical preponderance of body bits? I’ve no idea, but it seems plausible.
(The use of fanfic—along with hell knows how much other less formally constrained writing from all over the internet—coaxed into various levels of smut and filth, despite all the attempts to guardrail that stuff. A model of a human language will inevitably be, to some extent, a model of the culture whose language it is. When that language is USian English… well, I think I can leave you to infer your own ending to that train of thought.)
The title of this post is one of the structural tics of LLM-assisted non-fiction that I personally find particularly grating1: a short declarative, usually summing up the argument of the preceding paragraph or serving as a section header.
“The something somethings the something else.”
This form is not unique to LLMs, of course—and if you were to ask me to speculate on where they picked it up from, I’d suggest marketing copy and consultancy reports (assuming you recognise a useful distinction between those two categories).
I know very well that, when I see it, I can’t be sure that it was produced by an LLM. But the more I see it—particularly in the writing of people I’ve been reading for a long time, and therefore have stylistic sense of—the harder it becomes to read with good faith. It might only be a case of people who write for themselves absorbing an increasingly common tic from the broader body of everyday writing; that’s how language works, after all, and it’s not an entirely conscious process. But it’s horrible nonetheless—like the “how do you do, fellow kids” meme, but in reverse.
Which surely means, as some folk have delighted in suggesting to me, that trying to keep my own writing free of the taint is a fool’s errand. Quite possibly so! But choosing to write at all is a fool’s errand in the first place—and thus, at this point in my life, a sort of psychic sunk-cost fallacy means that I’ll be taking my pathological attachment to words to the grave with me, thankyouverymuch.
- Coming up in close second on my shit-list: “What [x]d was the [y]”, where [x] is a past-tense verb and [y] is a noun. Something about this one reminds me of health-and-safety manuals, or the little moral explainers at the end of a certain sort of Saturday morning cartoon. ↩︎
Leave a Reply