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

It's just text.

* 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.

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?

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.

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.

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.

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

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.

"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.)

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.)

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...

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.

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.

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.

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

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.

If we imagine a set of all human ideas that these models have access to, then the set of possible discoveries would be something like the superset of all possible combinations of those ideas. I think all LLM discoveries are bounded by that space.

Looking at the recent OpenAI math discoveries, that seems to be pretty much what happened. Existing ideas were used as building blocks, the model found a valuable combination, and the result was something new that had real value.

Pangram agrees: https://www.pangram.com/history/3de33376-94e3-404d-bbb0-751a...

at the least, this is problematic for peer-review because the absolute number of submissions outpaces the time availability of a finite number of expert reviewers. We cannot quickly generate expert human reviewers, and so the community might converge towards half-baked solutions (AI-generated reviews or rejection systems, vastly expanded referee pools, etc.) that tend to erode trust and and make scientific communities more adversarial.

Have you tried doing that or even read the article?

The article says that their detector flags 0.4% of pre-AI papers as AI-written.

If I paste the first page from this paper (https://www.fourmilab.ch/etexts/einstein/specrel/specrel.pdf) in https://unslop.run/app, I get a 0% chance that it was AI-written.

My problem therefore is: we are seeing more and more papers written with tools that are known to make up facts, citations, and even entire papers. And the number of papers has increased, too. I therefore see it less as "people are being more productive" and more "people are releasing bad science much faster than we can keep up with".

A possible conclusion for this could be: If the majority of CS papers is AI written, let's just accept this reality universally and stop worrying about it altogether.

In other words, it "sounds smart" without necessarily having anything to back it up. In even more critical terms, it's very good at bullshitting.

Unfortunately for us, the scientific community current relies on a certain amount of trust. (To do otherwise is very expensive! see: bitcoin). When you introduce a known-bullshitter to write your papers, every human in the loop effectively has to defend against an adversarial attack. Not just the readers at home, or the peer reviewers, but even the author needs to be wary that the facts and arguments coming out of the LLM are true and meaningful.

Personally, I've been a minor contributor to several high-profile papers. I don't know how every field does it, but in my experience, the corresponding author (generally the PI or other senior scientist), is responsible for the accuracy of the paper. They ultimately have to trust the people who did the work that the facts are true. Introducing LLMs into the mix make it more difficult for them to identify and review sections they are unsure of. (An honest person will typically write at a confidence level reflecting their certainty. LLMs do not do this in any reliable way.)

I've also found that LLMs frequently use metaphors that are unhelpful, or used out of context in a field that isn't familiar with them. This makes understanding the text more work, for no good reason. Introducing terms or definitions with low relevance reads as impressive at first glance, but avoiding the standard terminology in the field just adds confusion. (As an analogy, imagine if you were reading a CS paper that, for no particular reason, devoted a section to a new data structure called an "akimbo tree," which after much untangling, you realized was a reinvention of a randomized splay tree.)

If fake papers weren't already a big problem before AI and the fields had already been policing themselves adequately, if this was already a functioning high-trust domain, maybe we could ignore this a bit more, but the fields already manifestly had problems. People taking advantage of that are reasonably more likely to use AI. The pressures to publish or perish provide the voltage and the AIs are a rather convenient path-to-ground.

I agree in some sense that if a truth is published, it doesn't matter if the AI or a human published it. However there are perfectly reasonable reasons to be concerned that AI usage is correlated to not publishing truths, especially in a world where merely being human-generated was already not an adequate check against that.

Pre-LLMs, a paper with no spelling or grammar errors showed that somebody had put effort into writing and editing it. If they cared about the presentation, they probably also cared about the content. LLMs routinely produce nonsense that looks superficially like high-quality work.

There are far too many papers to read all of them. LLM slop is evidence that something is probably low quality. As the saying goes, "if you can't be bothered writing it, I can't be bothered reading." The rare outliers will get enough citations and recommendations to overcome this filter.

What about the papers that graduate to proper publication?

Arxiv is full of pre-prints that anyone can upload.

Even if they manage to avoid straight-up factual incorrectness, the writing is full of ambiguities and vagueness when you look closely. Meanwhile, space is wasted repeating the same claim in multiple ways, or explaining something simple.

They also seem unable to resist the hype/advertising tone, overselling the contribution while exaggerating the limitations of related work.

I would much rather read grammatically incorrect or awkward sentences.

While academic writing does have a few pointless historical conventions, the huge majority of "status quo writing style" is a logical consequence of 1) minimizing ambiguity, 2) organizing ideas coherently, 3) distinguishing opinion/interpretation from fact, and 4) providing enough detail to be reproducible.

> The article doesn't seem to mention consideration of AI for polishing human work.

Because it isn't a consideration. You are what they are looking for.