Hacker News Reader

Comment by SwellJoe | Hacker News Reader

Trivial to simulate. Though I disagree that AI writing is impossible to differentiate from human prose, at least for now. It's still pretty obvious and still much worse than human prose (or, at least, much less interesting to read), though some are better than others, and I'm better able to spot writing from models I use regularly (Claude has a very distinctive style I can spot from a mile away, but that's true partly because I read its prose nearly every day when using it for coding).

View on HN

Depends on what’s being written, and who the audience is. Anything of any length would be hard to simulate in a way that would fool an author - writing has a certain flow to it. A cadence. The editing and restructuring, deleting of words, typos you don’t catch until some random reread, rephrasing of sections because you want to use the original phrasing later in the piece.

Could you simulate something be typed? Trivially. Could you simulate something be drafted? Honestly, even if you wanted to put in all that time and effort, I’m not even sure LLMs are sophisticated enough to send the logical drafts, loops and edits that would pass a writers sniff test

To add to this, one can’t ignore the relationship between signal and receiver. I’d imagine most people on HN have enough pre-LLM reading experience to have a decent sense of what was written by an LLM versus a human.

And as LLMs get better at producing human-like text, that same pre-LLM reading experience, which helps people tell the two apart, will become less and less common.

I think you could simulate something that passes a sniff test. A writer would probably spot implausibilities in the simulation if they paid attention, but then we're back to square one, because you can spot that something was written by an LLM after you've wasted your time reading it and realize that you've been led around in circles with superficial information and no coherent train of thought, but by then your time has already been wasted.

"It's still pretty obvious and still much worse than human prose"

You have no idea how many false positives and how many false negatives you have in your judgement. It is indeed impossible to differentiate between badly written human text and somewhat good written llm text.

> You have no idea how many false positives and how many false negatives you have in your judgement.

The default LLM style is pretty deterministic. "Not X, Not Y. Just Z", etc.

There are phrases and cadences which are very rare in human prose (<5%) but unusually common (+95% occurrences in LLM prose). It is not unreasonable to look at content which is 95% LLM tells and conclude that that an LLM authored it.

I have noted, IRL, that those people who read very little, and only read when they have to (work docs, etc) are literally unable to tell that a piece of prose sounds like an LLM even when it has about 12 occurrences of "Not X. Just Y" or "Not X, Not Y. Just Z" in as many paragraphs.

It's really similar to special effects in tv/movies: Whenever there's something new it's hard to tell, but as we get used to it and start seeing the patterns it becomes easier to tell. The older the special effects are the more obvious it is. Quite often you can just kind of tell if something was special effects or practical effects without being sure why. And there are always some really well done ones that slip past everyone, or odd lighting that makes it look fake (a thread a day or two ago on here about a legal case involving a photograph, some people in the comments thought it was a painting).

You've found the "holy grail" of LLM writing differentiation! Your opinion is in the "Goldilocks zone" between vacuous and paternalistic!

I wanted to be sure I wasn't lying to myself. So, I built this: https://prose-or-con.com/

I can still detect AI-written prose with pretty high confidence (I average about 85% accuracy after a bunch of rounds while testing, though I'm rushing because I'm testing), but it has gotten harder. Opus 4.8, GLM 5.2, GPT 5.5, and DeepSeek v4 Pro, are able to fool me sometimes. Mistral writes so weird it almost trips me up, but then I go, "oh, that's just Mistral". Qwen sometimes accidentally switches to Chinese (filtered out of the game, just a thing I found in testing). I'm doubling the corpus, currently, with a focus on the frontier models that have been able to fool me more often, so it'll probably get a little harder. GPT 4o is disgusting, just horrible saccharine nonsense (and, I guess that's the model that's caused the most psychosis with its sycophancy).

And, I have to concede that the best models prose has gotten better, especially when they're prompted to write in a specific style, like literary fiction or an encyclopedic article about a specific topic. Still flowery and verbose, though, and never really getting to the point (the AI doesn't want anything, and it shows).

Trivial to simulate basic keystrokes. But I don't think it's trivial to simulate the natural process of drafting something. There's no concrete heuristic or algorithm (yet) for judging these types of replays, but I'd be impressed if someone can actually make a program that reliably simulates a natural keystroke-by-keystroke thought process which appears human when replayed in this way.

Do you know how many things throughout history people claimed confidently they could definitely tell the difference when it was later scientifically proven that they could not? And in all cases they believed that they were actually special and they were in fact correct despite the fact that this keeps happening.

I don't think I'm special. I think many people can tell the difference.

Edit: Also, I'm surprised images have gotten to the point where I have a hard time detecting AI in some cases, and they got there more quickly than prose. I really thought prose would be the first to fall. Video is still detectable. Music still detectable (by someone that enjoys music and pays attention to it). But, AI prose still outs itself pretty quickly.

OK, this is a silly thing to do, but I wanted to be sure I wasn't imagining that I can tell when prose is written by AI. So I made a game. Turns out Claude Opus can fool me better than any other model, but even it can't fool me a majority of the time. I average about 85% accuracy on this (Claude prepared the corpus, I'm going in with very little foreknowledge). GLM is also very close to being able to convincingly write like a human. I'm not as good at detecting AI as I expected I would be, but I'm still pretty consistently able to detect AI prose.

https://prose-or-con.com

GPT 4o likes writing poetry with bees in it, for some reason. Qwen models are decidedly purple in their prose.

Or maybe 2 out of 3 guys you call "AI" actually wrote it themselves and you were too prejudiced to believe them.

Doubtful. I read a lot of forum posts, as I maintain a popular OSS project. And, I read a lot of books from the before times. And, I read a lot of LLM prose, because I use them every day for coding and various assistant type tasks. It's still pretty obvious when an LLM writes something.

I even did post-training on Gemma 4 (a small model which has very good prose for a model, especially a small model) to try to make it able to write more like me, purely as an experiment to learn how training LoRAs works, and using data that I know is ethically sourced. It still distinctly writes like an LLM, with a few of my annoying quirks baked in: Inappropriate use of ellipses, too many parentheticals, occasionally dismissive tone. But, it can't stop doing the LLM things, either, without becoming incoherent.

Reply on news.ycombinator.com