Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

How we measured AI writing across arXiv, and where the measurement breaks

unslop.run

Post reply on HN
Print view View original

171–180 of 185 posts

Re: How we measured AI writing across arXiv, and where the measurement breaks

#171

Post by JamesBarney »

malshe wrote: post #154
JamesBarney wrote: post #75

Earlier quoted context omitted.

Almost every. Pangram is pretty accurate on longer texts. I haven't seen any glaring examples of false positives.

I’ve heard good things about Pangram but it’s trivially easy to fool it. I generated text using Opus 4.8 and then humanized it using Grammarly. Pangram determined it 100% human-written. Tried it numerous times with the same result. GPTZero claimed with 80% confidence that it was human-written but AI polished. Nothing in that text was human-written.

How long was the text? And mind sharing?

Re: How we measured AI writing across arXiv, and where the measurement breaks

#172

Post by ryandvm »

jknoepfler wrote: post #97
ryandvm wrote: post #85

There some real game theory mechanics at play when it comes to LLM usage in corporations. Devs are cranking out superficially superior code and documentation by just aiming the Claude Code fire hose at everything they can. Leadership encourages this because from what they can tell, there is no downside. It's hard to say if this code is structurally better or worse than before, but it's certainly voluminous and as far…

> "superficially superior code" Can you unpack that a bit? It produces measurably, meaningfully inferior code everywhere I see it in use. > "leadership encourages this because from what they can tell, there is no downside" As said 'leadership' I find this a bit puzzling. I'm seeing strong, quantifiable evidence of increasing churn, increasing incident count, and length of downtime from the date of our biggest push in…

Sure, by "superficially superior code" I mean that LLMs, by design, produce code that has syntax and idioms that are perfectly in line with industry norms due to the vast training corpus that it comes from. It looks great at first glance.

Normally, this is a signal that the developer is at least informed about best practices in the language and is usually a pretty good indicator that they're thinking clearly about the work they're doing. That, along with reputation, takes a lot of cognitive burden off the reviewer.

However, with LLMs that is all it means. You can be sure the code is norm-compliant, but in my experience that is all you can be sure of. There's not really a guarantee that it has thought deeply about the problem space.

It's like a bridge that looks great.

Re: How we measured AI writing across arXiv, and where the measurement breaks

#173

Post by chungusamongus »

jrm4 wrote: post #169
mghackerlady wrote: post #64

Earlier quoted context omitted.

That's not the point though, the point being made ist that an LLM written paper very likely doesn't actually find something despite looking like it does on the surface

I would like some sort of proof of this claim. I genuinely don't know, but I would suspect that the gap between "things found" in LLM v. Human papers is not as great as we would like it to be.

They don't have any proof, because the claim is wrong.

Re: How we measured AI writing across arXiv, and where the measurement breaks

#174

Post by chungusamongus »

grey-area wrote: post #100
chungusamongus wrote: post #18

Earlier quoted context omitted.

This is fallacious. The fact that someone used LLM to write their paper does not negate its findings.

It certainly casts suspicion on the ‘findings’ which may well be hallucinated or subtly wrong.

Because humans totally don't get things wrong or fabricate lies. Even the term "hallucination" is a clumsy way to describe what is happening when an LLM spits out an unsound statement. What you are doing is more akin to a hallucination than what the LLM is doing. Even in the process of criticizing this technology, you cannot help but to anthropomorphize it.

Re: How we measured AI writing across arXiv, and where the measurement breaks

#176

Post by speedstyle »

hereme888 wrote: post #98
SoftTalker wrote: post #87

Earlier quoted context omitted.

AI doesn't polish human work. It is more like an extruder that squeezes anything you put into it into a generic, formulaic shape, indistinguishable from writing that was lazily prompted because the writer couldn't be bothered to put forth any more effort than that.

"AI does not polish human work. It acts more like an extruder, forcing anything fed into it into the same generic, formulaic shape—indistinguishable from writing produced by a lazy prompt from someone unwilling to put in any more effort." There. AI-polished sentence.

Yes, this is noticeably worse than the original. The first sentence is rhythmically stilted and carries less impact. "is" to "acts" is not stylistic, we're no longer discussing the nature of the object but its behaviour. 'squeezing into a generic shape' is a transformation, each piece of writing stamped into conformity, whereas 'forcing into the same generic shape' must be pulled from a grammatical reading where all writing is being packed into one box. I think "writing produced by a lazy prompt from someone" is a wash, depends if they want to further detach the writer's agency from the result (and it still isn't produced by the prompt); I do prefer "unwilling" for the original tone; but it's certainly blander writing overall. If I came across it I'd even discount some of the metaphors and choices that are still present – since it's clearly written by AI, or a marketing copywriter, I'd guess they were "polish" rather than attempting to convey any particular connotation or nuance.

Re: How we measured AI writing across arXiv, and where the measurement breaks

#177

Post by LearnYouALisp »

ycombinete wrote: post #155
thesmtsolver2 wrote: post #143

Earlier quoted context omitted.

I replied to a comment with a structured informal proof on HN. A well established user here was adamant that I used AI because apparently humans never ever wrote proofs. This was a while ago. Any well crafted human output is now being dismissively cast as AI if the reader is challenged by the output intellectually/politically.

It's the irl version of being accused of hacking in Counterstrike.

Yeah but because there is tons of it, and not all the accusations are on the same grounds (and level of competence)

Re: How we measured AI writing across arXiv, and where the measurement breaks

#178

Post by LearnYouALisp »

cansofgrease wrote: post #71
pbui wrote: post #65

I'm not sure what to think... I uploaded a PyHPC workshop paper I wrote in 2011 and it said 27% machine. I also uploaded my PhD dissertation from 2012 and got back 40% machine, which is just barely below the 42% threshold. I don't publish anymore... but does this mean I wrote like a LLM or did LLMs learn from me? :p Update: I also uploaded a IEEE CLUSTERS paper I wrote in 2015 and it came back 74% machine written :|

It's been trained on a work and then distilled, the false positive noise has to be absurd.

"You made this? "... "I made this."

Re: How we measured AI writing across arXiv, and where the measurement breaks

#179

Post by RugnirViking »

dopamine_daddy wrote: post #2

I scored the full text of 12,750 arXiv papers from 2021 through 2026 to find out how many of these get flagged as machine written and how much it increased since the release of chatGPT. I purposely tuned the detector to avoid false positives. My detection rate pre chatGPT is around .4% for that reason. The biggest results: in Jan of 2026 about 39% of papers got flagged as AI written. In computer science speicifcally…

Have you considered trying the same sort of analysis between two time periods pre-ai?

What I mean is, are there statistical differences in how people write that would give similar looking results between say 1990-2000 and 2000-2008? Are we just detecting the natural progression of language here?

Re: How we measured AI writing across arXiv, and where the measurement breaks

#180

Post by JadeNB »

never_inline wrote: post #156
JadeNB wrote: post #59

Earlier quoted context omitted.

Which part is a sin? Using LLMs to deal with a lack of English-language fluency? I am a scientist (actually a mathematician, if it matters), and, if that's the way to deal with the practical hegemony of English in the scientific literature, then I have no problem with it. Rather that than people with important ideas can't get them before the scientific community. As long as the authors personally check and stand behi…

https://lists.isocpp.org/std-proposals/att-0486/Reply_to_Zer... Here's a paper by a non native English speaker. The lack of the formal style doesn't cause any issue. Preciseness is what matters.

It seems to me like we're saying the same thing. I'm saying that there is a practical bias, as in, it's there in the real world whether we like it or not, that can prevent some papers from appearing just because of limitations on English-language fluency, not because of any lack of scientific content. One solution is to get rid of it, but that's a long-term project. If we have an immediate remedy that works, if the authors still take responsibility for the scientific content, and if the use of the tool that provides that remedy is responsibly disclosed, then it seems like a win all around, except perhaps that it makes the need for the longer-term solution less obvious.
Post reply on HN