Top
Best
New

Posted by dopamine_daddy 14 hours ago

How we measured AI writing across arXiv, and where the measurement breaks(unslop.run)
209 points | 152 comments
pbui 12 hours ago|
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 :|

NitpickLawyer 12 hours ago||
When "detectors" first started popping up all over the place, all of them rated the declaration of independence as 100% AI written, so... Yeah, these things just don't work. And what's even more dangerous is that people that don't understand how any of it works use these tools, and accuse people of using AI, sometimes with grave consequences. Students have been through this, at all levels of education.
thesmtsolver2 7 hours ago|||
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.

ycombinete 2 hours ago|||
It's the irl version of being accused of hacking in Counterstrike.
roysting 5 hours ago|||
I’ve been accused of being a bot myself. I’ve started leaving some incorrect autocorrects in, hopping it would stave off people for whom it seems making accusations of being a bot is soothing or a coping mechanism when they’re confronted with information that does not fit into their world view or perception of self.
jszymborski 5 hours ago||||
In fairness, the original crop of detectors were based on perplexity scores and were entirely useless (famously, the declaration of independence often came back as 100% AI generated).

I'm not convinced we'll ever have full-proof detectors, and certainly the false-positive rate will make them irresponsible for accusations of intellectual/academic fraud, I do think that LLMs are easy for folks to sniff out on average so I imagine it's possible to detect many instances.

Pangram's detector is anecdotally very accurate in my tests. This detector appears to be fine-tuned on a very small dataset (200 papers per subject), and suspect the problem might be in part that.

yorwba 8 hours ago|||
They use a "threshold calibrated so pre-ChatGPT papers flag at 0.4%", so these things do work most of the time. It also means that there are known false positives, so for any given paper, scoring above the threshold isn't irrefutable proof of AI usage. But for things like estimating the overall proportion of AI writing, you only need to be correct on average, so individual false positives don't matter much.
jasonfarnon 7 hours ago||
It would be interesting to see how that .4% varies across fields on arXiv. I imagine some areas influenced LLM writing much more than others and are more susceptible to false positives. Anyway it's a little strange to me that these detectors do so poorly (at least by what I hear on this site). There is as much labeled data as you would need to train on pre-LLM human vs LLM text. If their out of sample errors are as good as has been claimed, the only methodological flaw I can think of is that post-LLM human text and LLM text have been converging.
yorwba 1 hour ago||
There's a table in the article breaking down the rate by field.
miohtama 1 hour ago|||
It means machine text and human text cannot be distinguished from each other.

It's just text.

sean_pedersen 6 hours ago|||
"Any attempt to build AI generated content (deep fake) detection systems is flawed, since the outputs of such a system may be used to train an even better fake data generator. This leads to an equilibrium state of digital uncertainty: nothing in the digital realm can be deemed as real anymore - only as digital. I do not care if a digital artifact is human or AI made - I only care if it is useful to me. Useful content is on point, factual and at best surprising (teaches something new)." - https://seanpedersen.github.io/posts/digital-uncertainty/
dgellow 11 hours ago|||
Could it be that your papers are literally in the training set?
cansofgrease 12 hours ago||
It's been trained on a work and then distilled, the false positive noise has to be absurd.
dopamine_daddy 14 hours ago||
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 the peak was at 65%. Mathematics barely moved away from 0.7%, though the proof heavy math texts might just not get picked up by the detector properly.

All this is a detector estimate of a statistical signal and not a proof any given author used AI. Machine written can also mean heavy AI-assisted editing.

drewcrawford 1 hour ago||
Couple of methodological notes

* pre-chatGPT is not an effective control because language evolves. In the arxiv corpus in particular there are "fashions" in research depending on what gets funded lately, not to mention many new words and topics not invented before a given paper.

* In general, detecting AI from content seems difficult as humans write like they read. To the extent there are unique factors recognizable as AI and to the extent humans read them, they will eventually incorporate them into their writing style. Accordingly, you'd need to model a rolling window of "AI tells" that decay at some rate.

phreeza 13 hours ago|||
This is a pretty stunning result. The time series looks really convincing. Is the way the detector itself is trained orthogonal to this or could there be some "leakage" in that the pre-chatgpt text is in the (positive) training data?
dopamine_daddy 13 hours ago||
I tried my best to avoid leakage. If you're curious about how I trained the detector I have a writeup on it: https://unslop.run/blog/how-our-ai-text-detector-works

FYI this is all relatively new so there might be lots of issues and iterations coming.

Paracompact 11 hours ago||
My honest first impressions, since I think the project is well intended: This writeup is itself AI, and I would venture to call it slop. The Calibration section is borderline uninterpretable, and I challenge any non-author who claims to understand it to answer some basic peer review questions about it.
simonreiff 9 hours ago||
I have to concur, unfortunately. This isn't reliable, reproducible, or interpretable.
paxys 12 hours ago|||
What detector are you using? How can you be sure of its accuracy given that every commercial AI detector has been debunked?
JamesBarney 12 hours ago||
Almost every. Pangram is pretty accurate on longer texts. I haven't seen any glaring examples of false positives.
malshe 3 hours ago|||
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.
adamgordonbell 5 hours ago|||
This! Pangram is very good. They claim 1 in 10,000 fp rate.

I had to change my mind on AI detectors after playing around with it.

It would be interesting to hear how this detector compares. It also seems to be aiming for low fp rate.

cgio 4 hours ago|||
A signal, I put an AI generated of mine (with heavy human guidance on aesthetics mostly and some minor human editing) and got 5% only.
0x000xca0xfe 11 hours ago|||
Could you share some pre LLM false positives? Would be interesting to see what is tripping the detection. Did a tiny fraction of authors write like LLMs, before LLMs?
jmcqk6 12 hours ago|||
Extend it to scan papers from pre-2020. That should give you a better baseline accuracy for your detection system.
andycasey 13 hours ago|||
this is neat! is the model available somewhere for local execution? or even a lookup table with your results for all arXiv pre-print codes? I want to run it on lots of pre-prints and I don't want to kill your server
dopamine_daddy 13 hours ago|||
Thank you, I plan to release the arxiv preprint codes. Also don't worry about killing my server, let me know if you succeed :D
exe34 12 hours ago|||
Wouldn't that provide an excellent signal to tune models to be less like AI? I suppose that's a good thing ultimately.
economistbob 13 hours ago||
Thank you for this work.
ryandvm 12 hours ago||
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 as leadership could ever tell, with their flawed metrics, that is all that matters. It will be years before we figure out if this is a good idea and worth the cognitive atrophy.

Anyone not using LLMs all day is just not going to be as prolific. I can't imagine that the same factors aren't at play in the scientific research community where it's all about how much you can publish.

pcrh 8 hours ago||
>I can't imagine that the same factors aren't at play in the scientific research community

I can't speak for the entire scientific research community, but I can say that for basic biomedical research (not clinical research), LLMs are mostly ignored. They simply don't have the ability to parse most raw wet lab data.

This obviously has not stopped the creation of a legion of startups, new Asst Profs, etc, claiming that they are using "AI" to crack new frontiers. In practice, the most effective of these use machine learning, rather than LLMs.

addaon 11 hours ago|||
> It's hard to say if this code is structurally better or worse than before, but it's certainly voluminous and as far as leadership could ever tell, with their flawed metrics, that is all that matters.

Hah. I've been working with ChatGPT-5.6 Sol and a lot of guidance to port a particular geometry from 2800 lines of Python-scripted Rhino3D (basically a custom parametric CAD kernel in there, plus use of it) to Python-scripted OCCT/FreeCAD (an existing CAD kernel) and it's up to 36,000 lines and only about half way there. And I've been setting goals and targets for duplication minimization, code size reduction, etc. The results are fine for my purposes, but if I put a positive value on "voluminous" I'd be super impressed; with my actual negative value assignment on lines of code, this is more "absolute crap but still useful to me."

I think "absolute crap but still useful to me" is a pretty high value and not worth neglecting, but I'm definitely scared by leadership who sees the toilet overflowng and assumes that means that value has been maximized.

William_BB 10 hours ago||
> I think "absolute crap but still useful to me" is a pretty high value and not worth neglecting

If I extrapolate this example to my professional life, this code now manages millions of dollars, a single mistake can wipe it all out, it has to be maintained by 5 other engineers and understood by 5 other domain experts.

addaon 9 hours ago||
That's a real risk, yes. There's definitely a difference between cases where the code is the artifact being built, and cases where the code is being built to generate an artifact that is independently valuable and assessable. I'm happily in the latter case -- but I definitely share your fear of those who think current LLMs are valuable in the former case.
ffsm8 11 hours ago|||
> It's hard to say if this code is structurally better or worse than before,

Huh?

But that's easy to tell?

It's better then incompetents code, worse then a motivated average dev... But good enough hence the real question is value aka time& money invested/quality.

That's much harder to tell, and I currently think opus/fable generated code is decent enough to be safely in favor - at least on subscription

William_BB 10 hours ago||
> It's better then incompetents code, worse then a motivated average dev...

That's well put.

> But good enough hence the real question is value aka time& money invested/quality.

There's time invested SO FAR and time that will have to be invested to maintain it. In my experience, even with Fable, it's not there yet. It's the reason why it's easy to vibe code an app from scratch, but at some point when complexity significantly increases, the codebase becomes a mess.

MetaWhirledPeas 12 hours ago|||
> Leadership encourages this because from what they can tell, there is no downside.

This was my big fear before we saw price increases. Now I'm pinning all my hopes on AI being too expensive to justify further big corporate pushes. (Sigh.) I love having new tools, but I hate being pushed to use ______ tool to meet some managerial metric.

akomtu 11 hours ago|||
IMO, cognitive athrophy is the point. If humans become dependent on AI, those corporations will get enough leverage to change laws in their favor. For example, people don't want a huge data center in the middle of their city, but the corporations will say that the same people can't do anything without AI powered by this data center. That's why I think that making humans deficient and dependent on AI is the entire point.
Jtarii 10 hours ago||
This is just conspiracy nonsense.
jknoepfler 11 hours ago||
> "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 into GenAI, and I'm organizing efforts on my teams to mitigate those issues and actively reduce GenAI adoption.

If you mean my c-suite, you're mostly correct although they are already rumbling about seeing zero or negative ROI on GenAI investments.

> Anyone not using LLMs all day is just not going to be as prolific

Agreed, but prolific != productive.

sandeepkd 11 hours ago|||
> leadership

Not sure if its just me, lately I have started feeling pretty offensive about the increased usage of the word. Its management not leadership by any means.

> "leadership encourages this because from what they can tell, there is no downside"

For most people in management its easier to pick the current set of slangs/abbreviation's, general trend and go with it. Understanding the details would take time, raise questions and no one in management has time or political capital to spend on it.

mewpmewp2 7 hours ago||||
What size company are you talking about?
OleksandrC 11 hours ago|||
The code written by agents generally seems to reflect what you (the human driver) asked for. If you discuss the approach and architecture first, asking the right questions in the process, and then let it implement - the result is quite on point with the frontier models. Might need some minor touch-ups if agent missed some common conventions or guessed the expectations wrong (but again, salvageable with follow-up prompts). And when it comes to line-by-line logic within functions, I would argue that today's models are LESS likely to make a mistake in there than humans - especially if you cross-review with another model (e.g. "write with Claude, review with GPT").

Humans write slop too, you know. Just saying.

pkage 13 hours ago||
As with all text-only AI detection schemes, I am concerned about the accuracy of the detection. I'm skeptical of the methodology, specifically the final join of the three detector scores---how can you be sure that that final step does not introduce any biases? There's no source available, so it's difficult to tell exactly how this works or reproduce the research.

I've also uploaded text samples from my own (unreleased) research from pre-LLM era, and it's seemingly scoring pretty high on the LLM-detection scores. On other papers, nearly every sentence is highlighted as red "machine-leaning," but that does not impact the score? Additionally, there are dramatic differences between the scores for identical text with and without LaTeX formatting, despite the fact that it should not matter.

The takeaway from this should be "it is difficult to detect generated text and we should be careful about accepting results simply because they confirm a hypothesis."

--

Relatedly, the text above scores as highly machine-written, despite the fact that I just wrote it with my human hands, I promise :)

make3 12 hours ago|
The most concerning is the use of these tools to detect cheaters in schools.
Eextra953 10 hours ago||
I've looked at this same problem in an academic environment and have come to conclude that there is no way to reliably detect AI writing using only text. The reason for this is that no detector can take two identical inputs and classify one as synthetic (LLM) and one as organic (Human) and this situation can easily happen at the sentence or even paragraph level. Posed as a question, if a human writes a paragraph that happens to be exactly the same as a paragraph written by an LLM how do you classify that paragraph? There are a finite number of words and a finite way of combining them within an academic setting/field which means that we can't build a perfect classifier with just text. In practice, there are obvious tells and LLM-isms but these also change with time and each model has different tells so that even if a paper is full of LLM like writing we have no way to disambiguate between organic text that appears synthetic or synthetic text that appears organic.

A more interesting question, to me, is looking at a corpus of essays and analyzing how writing has changed with the introduction of LLMs. We can look at changes in vocabulary, linguistic features, style embeddings, regular embeddings, typos, errors, and references over time. When looked at in this way it is clear that academic writing has changed at the population level but what has led the change is harder to track down.

Calavar 10 hours ago|
> There are a finite number of words and a finite way of combining them within an academic setting/field which means that we can't build a perfect classifier with just text.

Maybe for very short phrases, but otherwise I disagree. Phrasing very quickly runs into a combinatorial explosion. In the words of Noam Chomsky, "Virtually every sentence that a person utters or understands is a brand-new combination of words, appearing for the first time in the history of the universe."

In my opinion, the difficulty in LLM/human text discrimination isn't that a person might coincidentally write exactly the same text as an LLM would, but rather that 1) LLMs aren't hard locked to a single phrasing (so this is a tougher problem than matching to a single static document, e.g. plagiarism detection) and 2) text has relatively low information density, so you need quite a bit of it to gather enough data to run a statistical test with a reasonably narrow confidence interval.

willquack 13 hours ago||
> If a tool marks 40% of new papers as machine-written but also marks 20% of papers written before ChatGPT existed, the real story is the 20% nobody mentioned.

When 65% of the papers you read have the characteristics of being AI written, whether or not you use AI to write, your writing will be influenced by the AI style. I imagine this must be particularly the case for newbie researchers who are still developing their writing style

aionwikipedia 9 hours ago||
even if your writing is influenced by "the AI style," it would likely only show up in patches, a word or rhetorical flourish maybe, but AI-generated text doesn't "sound like AI" in patches, it does so consistently across the entirety of the output. it's very hard for a person to maintain the syntactical patterns of AI writing over an extended period of time unless they know exactly what those are -- probably more knowledge than any person currently has, really -- and are applying a level of detail to each clause comparable to forging a painting.

there's also the fact that "the AI style" has changed over time. for instance, the word "delve" is notorious as an "AI sign," which it was up until mid-2024, at which point it dropped off sharply and has now basically disappeared from LLM output. so if someone happened to pick that up due to reading it everywhere, their writing is now less characteristic of AI, not more.

serial_dev 13 hours ago||
I might be missing something but what’s the real story of the 20%.

To me it sounds like 1. Either your tool is just not that good and reliable as you thought, 2. AI is trained on human written articles, so some of that human written content informed the now established “AI slop”.

There are people who shipped “slop” before AI.

cisophrene 12 hours ago||
> There are people who shipped “slop” before AI.

The funny thing is that "slop" was defined by the writing habits of AI model, which we have learned to pick upon and recognize.

The "It's not X, it's Y", the rhetorical questions and other patterns would have been the tools of a skilled writer, and those people writing "like AI" before AI most likely would have been recognized as such.

lingeringpine 13 hours ago||
I am not a native English speaker. This is not surprising to me. I think most of the papers we write would be flagged by AI detectors.

It is not because we ask LLM to write us a paper about X. It is because we are bad at writing in a scientific style, and american editors expect us to do it. With LLMs, we can write in basic sentences and tell the LLM the idea and it converts that to nice writing.

If you write each paragraph and have an LLM make that paragraph more scientific, it is entirely your paper, but it is flagged as LLM generated. If you have an LLM write the paper but speak good enough English, you can make it look human even though it is not human.

robotsliketea 4 hours ago||
It seems that LLMs often insert new ideas even when asked only to polish writing without changing it: https://bsky.app/profile/tomerullman.bsky.social/post/3mq33c...

I don't think this is surprising. Good technical writing is very precise. If you're starting from non-technical writing, I suspect that in most cases you can't make it "sound scientific" without adding new claims or changing the meaning. (Maybe you are being more careful, but this is something that worries me in general.)

aionwikipedia 2 hours ago|||
most papers like this are flagged by AI detectors because, by your own admission, AI was used to generate the final output. what exactly did you expect here?
dopamine_daddy 13 hours ago|||
Yes that is exactly what I suspected. And I see absolutely nothing wrong with this.
vidarh 12 hours ago|||
My English is fine, and I did hand-write my MSc thesis in English, but scientific style writing is very tedious. Personally I don't think I'd even consider writing a paper "by hand" if I were to write another one.

There really is no point, as long as you verify the content matches your intent and edit out anything poorly written.

Frankly, I've read plenty of papers by native English-speakers over the years that'd strongly benefit from being rewritten by an LLM too...

bananaflag 13 hours ago|||
> It is because we are bad at writing in a scientific style, and american editors expect us to do it. With LLMs, we can write in basic sentences and tell the LLM the idea and it converts that to nice writing.

This is a sin.

JadeNB 12 hours ago|||
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 behind the scientific content of their papers, what do I care how the exposition was produced, especially if the role of LLMs is properly disclosed?
never_inline 2 hours ago|||
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.

slopinthebag 12 hours ago|||
[flagged]
exe34 12 hours ago|||
I feel like they should release the original before the llm chewed on it.
zfnmxt 13 hours ago|||
> If you write each paragraph and have an LLM make that paragraph more scientific, it is entirely your paper

"More scientific" is not some merely stylistic thing that faithfully preserves the original meaning of what you wrote. The precise details of each paragraph matters a lot in terms of what and how it communicates. The fact that these details do matter means that, according to my accounting, it is not entirely your paper.

Also, I find striving for "scientific" to be a pretty undesirable thing. Why should papers read like that? What is the benefit? The best papers (in terms of their writing and communication) are unpretentious and conversational. I'm pretty sure I'd prefer your "basic sentences", especially if they were wholly yours. (I understand that there are also external forces at play here as you mentioned.)

simonreiff 8 hours ago|||
Being charitable, "more scientific" means more rigorous, reproducible, reliable based on the consensus of that scientific community (if any consensus exists, which it usually does in math and hard sciences), transparent in methodology, and candid about authorities cited and limitations to the experimental approach. Science is not a style guide; the purpose of a research publication is to impart a result to experts in the field and make claims that can be rigorously scrutinized. The best way to make a claim credibly is to demonstrate full fluency with the tools and techniques of the academic community in applying them thoughtfully to the chosen problem. But the result of doing that often looks downright impenetrable to the general public; see, e.g., Niven 1947 (1-page proof that pi is irrational). It's not the point of science to make it accessible to everyone. The point is to convince experts in the field of the correctness of the claim. The audiences are often totally different, and the tone and qualities of the writings likewise must vary.
smcg 12 hours ago|||
You might prefer that, but journal editors might not.
zfnmxt 12 hours ago||
I've never had any trouble publishing in a more conversational style and have read plenty of papers that do the same. Of course, this varies by field and venue and format (journal vs. conference) and even reviewer and it's difficult to understand what the true constraints are and what the state of the publishing system is in this regard.

But I suspect that a lot of academic's feelings about it are informed by what others have told them and how they've been trained, rather than by what's actually permissible in the publishing system.

bjourne 13 hours ago||
There is a fine line between using an LLM to clean up grammar and spelling and using it for text generation.
efitz 44 minutes ago||
> Here is the method, the results, and an honest account of the limitations.

This paper was written using AI, to be honest

bilsbie 13 hours ago||
Just to play devils advocate. These kind of papers are very verbose and boilerplate. I can imagine using AI to write 90% but then the actual novel content and explaining what’s important could be handwritten.

Perhaps that’s what’s happening.

breezybottom 12 hours ago||
If 90% of it is boilerplate, then you really need to question if you have something worth publishing as an academic paper.
IncreasePosts 11 hours ago||
It's generally always better for an academic's career to publish something, rather than publish nothing.
NoImmatureAdHom 12 hours ago||
Yes, this is the thing. In the scientific enterprise, the writing is mostly wasted time. Why would you write it up if machines can do it competently?

Sure, some people are artists - but most aren't.

tstactplsignore 12 hours ago|
One challenge with this approach is: could it be possible that the detector is simply learning to recognize words and jargon used more in the literature post 2022 as 'AI'?

For example, LLMs love to talk about LLMs (and the people who write with LLMs love to write about LLMs). Could "large language model" itself therefore be flagged as an AI-like phrase by this approach? It didn't exist much in the literature before 2022, does now, and certainly does more in AI-generated text: but, it is not actually a great way to distinguish modern AI generated text from human written text.

A helpful control would be to show that on some cohort of papers that can be declared reasonably clean of LLM generated text post 2023 there are very low rates compared to the arxiv.

For example, while papers in the journals Nature and Science are unlikely to be entirely LLM free at this point, if those were tested through 2026, we should see a line significantly lower than the arxiv's growth.

dopamine_daddy 12 hours ago|
You raise a valid point. Just by intuition I'd say if this were true, it would probably just be a small fraction of the actual flagged articles. I will still look into how I can mitigate this when I update the detector.

The difficulty with this is then: How do you get a clean post 2023 dataset? I have no straightforward idea for this. You can't use other AI detectors to build it because then you'd never outperform them.

More comments...