This topic was highlighted in the It’s FOSS newsletter!
Anyway; I can answer the first half, sort of! How do they know it’s AI? Well, they don’t. Not 100%. But they have a very good idea when it is.
For context, I’m also very interested in writing, literary magazines, etc. I’ve been published in quite a number over the years and have been on the selection committee for several as well. Some of these litmags have seen submission rates increased x100 over the last few years, as LLMs became more publicly accessible, which can be a massive strain on our resources.
There is no single, foolproof detection-method. However, there’s a number of ways I’d identify AI prose both just semantically and with tools.
- You can scan content with AI detectors. GPTZero, Copyleaks, Pangram, etc. There’s loads of them. If you search them up, you’ll see a lot of chatter about how they’re notoriously unreliable… But that’s mainly false negatives. When GPTZero, for example, flags something as AI-written, it’s unlikely to be wrong. Then if you’ve tried it across multiple detectors that use different algorithms, it’s a bit like a false positive across multiple pregnancy tests - not impossible, but increasingly implausible.
- You can request drafts, notes, or even a manuscript with Track Changes; if someone uses AI, it’s typically unlikely (though again, not impossible) they’re going to have the detailed notes associated with human-written content. Track Changes is an interesting one as it shows someone the various sentences, backspaces, line edits. AI written prose is generally pasted in in a few chunks, written very Beginning-Middle-End with no drafting or jumping around, and so on…
But to be honest, the first sift is mainly just through authorial cadence and AI tells in prose.
When you read a lot of it, you see a lot of the same ticks over and over and over across submissions. AI isn’t ‘writing’ in the sense you or I would write? It’s a probabilistic pattern engine and it selects the next most likely word based on the prompt and the training data. Yeah, okay, the training corpus is huge, but because of that you actually get much more literary convergence. It regresses to the mean. You can’t really tell it ‘use a literary style, not AI prose’ because it’s blending together all these millions of training samples and selecting from that smooth, frequent middle of moderate sentence length, a particular rhythm, specific character archetypes. Factor in frequency bias, the way models favour statistically strong signals, and optimisation for safe, clear, ‘helpful’ prose… and you end up with this quite specifically blah and bland prose style common across Claude output, ChatGPT output, etc.
Humans do a lot of this too, to an extent. Seeing three or four of these really common AI ticks isn’t enough to say “well, this must be AI’“. But if I see repeated and condensed AI patterns across the prose, well, it’s going in the ‘probably AI’ pile for further investigation.
Shy Girl is a novel that got a lot of attention for this recently, and is a good example of this flavour of “AI Prose”!
Aside from the prose itself, you’ve also got the way it pushes characters towards archetypes with little nuance. Dialogue can be repetitive or not make logical sense (replying to things that haven’t happened yet, characters spout dialogue that sounds right, but is actually a total non sequitar when you think about it), they have knowledge that they couldn’t possibly have in context, make wildly accurate assumptions, everyone has that same Marvel Dialogue quippiness… I could go on. Again, not a sign in and of itself, but when you’ve got fifteen of 4o’s favourite lines in four paragraphs and then your characters are saying things like: “You’re a menace” four times in a row… yeah, it’s going in the pile.
It’s an interesting dilemma, though frustrating when you’re doing the sift. Competitions or mags that pay out cash prizes are often totally inundated with unedited, AI slop.
