Having a paper automatically rejected because some automated system ranks your writing as too AI-like sounds like a nightmare, I'm glad to not have this problem in my work.
I like the point how the language of reasoning not only makes llm output sound eloquent and convincing but that this language makes them work altogether (or, at least makes them work better). This technique works well for us humans too, which is why techniques like five whys analysis work.
On the other hand, I still think it's important to call out lazy slop. And this can be done without focussing on structure or stylistic devices. In my experience I often come from the other direction, I approach a text in good faith and if after a couple of paragraphs I'm having trouble grasping the author's point, I start looking for the telltale signs. And more often than not I will find them.
Interesting article, however in practice I'd make a difference between text used for reasoning (to think about something), and finished text written to communicate (typically after the reasoning was done).
In his example, you might say "it wasn't Thursday, it was Wednesday" while thinking about it, but you'll just write "it was on Thursday" when texting someone else.
So in academics or in the workplace, the actual deliverable (a report, an email, etc) wouldn't use language used to reason about the topic, and shouldn't be LLM-looking if written correctly. Maybe the draft or personal notes would, but not what is sent.
Even though I am a strong critic of generative AI, I see its uses in academic writing insofar that non-native English speakers can polish their writing. However, this requires a well-structured and almost complete draft, not just bullet points, because you otherwise end up with hallucinations and 'wooden' formulations.
Given the large amount of low-quality papers from China or India, just to name two countries, I have reviewed in the last few years, I noticed that I have become a bit biased against the typical non-native English idioms used by citizens of these countries. This is a shame, as some of the best papers I have ever reviewed also come from these countries.
So in a way, as English has become a standard language of academia, LLMs can help standardise a high language level, as well, and reduce bias in the review process.
I see its uses in academic writing insofar that non-native English speakers can polish their writing
I absolutely reject this reasoning. Papers should be precise and exactly capture the intent of their authors. Many institutions with employees that do not speak English as a first language have "writing advisors" (or similar) to ensure (1) good writing, but more importantly (2) that what is written matches their intent. When this is deferred to an LLM, this subtly influences meaning and will lead to worse outcomes for these authors (e.g. by factual incorrectness or poorly expressed claims). Institutions without language advisors are worsening outcomes for their employees. Independent authors often have worse outcomes anyway for other reasons.
I noticed that I have become a bit biased against the typical non-native English idioms used by citizens of these countries
Some amount of unconscious bias is unavoidable, but my experience with reviewing papers has generally led me to give considerable benefit of the doubt when there are clear language issues, and to issue revision instructions with what I thought was unclear due to these language issues. Managing this requires self-checking. While poor use of language can obfuscate the ideas, methodology, or results somewhat, LLMs can (and, in papers I have reviewed, do) outright mangle these to something factually wrong, needing rebuttal and revision anyway or just getting the paper rejected instead. There are better mechanisms to manage this problem; we should not recommend their use.
That's not even touching the open question of whether this is plagiarism or not. This is a much larger debate that folks never seem to agree on the base facts of. I mention it because some venues forbid LLM assistance on these grounds, and that should also be respected.
The bigger issue is that we're basically forcing people to avoid natural writing patterns now cause of the whole hysteria over whether something is written by an LLM or not. And, ironically, companies wasted no time monetizing the hysteria by using LLMs to decide whether something is legitimately human or not, cashing in on the whole thing and becoming the arbiters of what acceptable writing is.
The author correctly points out that if people stopped being lazy, and started engaging with the content instead of style then the whole problem would go away.
I think the problem is the lopsided equation that also exists for PRs: to review a PR or a paper costs a lot of time and energy. To find out you just wasted an hour (which you will never get back) on LLM slop someone took only minutes to prompt is deeply unfair. Hence the demand for services that purport to detect AI writing style (because they can't look at the content), and people attempting to do the same by hand. "It's not just being lazy, it's being smart about your time investment" (SCNR)
To find out you just wasted an hour (which you will never get back) on LLM slop someone took only minutes to prompt is deeply unfair.
Not only is it deeply unfair to expect people to waste an hour reviewing something that is shaped like a valid thing and only took minutes to prompt, it's absolutely unsustainable. Even at the rate humans produce slop without an accelerant like this, it was hard to have enough reviewers to do a good job. Getting halfway through a review and realizing you're dealing with the output of a stochastic parrot is demoralizing and will chase people away from this under- (or un-) paid job that they were doing for the good of their sector.
"It's not just being lazy, it's being smart about your time investment" (SCNR)
You're not just being glib, you're right, too. Asking people to engage with a firehose of slop is awful.
I understand the motivation, but it's pretty clear that the process creates a lot of harm in practice, and it doesn't actually address the problem either. We need different metrics other than style to discern whether writing has merit.
It is easy to ask an LLM to write in a certain tone, but having it produce reproducible results, honest error analysis, and clean ablations without actually doing the work is a much harder trick to pull off. Style detection punishes the form while these metrics would focus on the lack of substance. For example, arxiv started banning people for a year if they have hallucinated references in their papers. That's the kind of thing that makes sense to focus on.
tjammer | 10 hours ago
Having a paper automatically rejected because some automated system ranks your writing as too AI-like sounds like a nightmare, I'm glad to not have this problem in my work.
I like the point how the language of reasoning not only makes llm output sound eloquent and convincing but that this language makes them work altogether (or, at least makes them work better). This technique works well for us humans too, which is why techniques like five whys analysis work.
On the other hand, I still think it's important to call out lazy slop. And this can be done without focussing on structure or stylistic devices. In my experience I often come from the other direction, I approach a text in good faith and if after a couple of paragraphs I'm having trouble grasping the author's point, I start looking for the telltale signs. And more often than not I will find them.
brice | 7 hours ago
Interesting article, however in practice I'd make a difference between text used for reasoning (to think about something), and finished text written to communicate (typically after the reasoning was done).
In his example, you might say "it wasn't Thursday, it was Wednesday" while thinking about it, but you'll just write "it was on Thursday" when texting someone else.
So in academics or in the workplace, the actual deliverable (a report, an email, etc) wouldn't use language used to reason about the topic, and shouldn't be LLM-looking if written correctly. Maybe the draft or personal notes would, but not what is sent.
FRIGN | 9 hours ago
Even though I am a strong critic of generative AI, I see its uses in academic writing insofar that non-native English speakers can polish their writing. However, this requires a well-structured and almost complete draft, not just bullet points, because you otherwise end up with hallucinations and 'wooden' formulations.
Given the large amount of low-quality papers from China or India, just to name two countries, I have reviewed in the last few years, I noticed that I have become a bit biased against the typical non-native English idioms used by citizens of these countries. This is a shame, as some of the best papers I have ever reviewed also come from these countries.
So in a way, as English has become a standard language of academia, LLMs can help standardise a high language level, as well, and reduce bias in the review process.
addison | 4 hours ago
I absolutely reject this reasoning. Papers should be precise and exactly capture the intent of their authors. Many institutions with employees that do not speak English as a first language have "writing advisors" (or similar) to ensure (1) good writing, but more importantly (2) that what is written matches their intent. When this is deferred to an LLM, this subtly influences meaning and will lead to worse outcomes for these authors (e.g. by factual incorrectness or poorly expressed claims). Institutions without language advisors are worsening outcomes for their employees. Independent authors often have worse outcomes anyway for other reasons.
Some amount of unconscious bias is unavoidable, but my experience with reviewing papers has generally led me to give considerable benefit of the doubt when there are clear language issues, and to issue revision instructions with what I thought was unclear due to these language issues. Managing this requires self-checking. While poor use of language can obfuscate the ideas, methodology, or results somewhat, LLMs can (and, in papers I have reviewed, do) outright mangle these to something factually wrong, needing rebuttal and revision anyway or just getting the paper rejected instead. There are better mechanisms to manage this problem; we should not recommend their use.
That's not even touching the open question of whether this is plagiarism or not. This is a much larger debate that folks never seem to agree on the base facts of. I mention it because some venues forbid LLM assistance on these grounds, and that should also be respected.
[OP] Yogthos | 5 hours ago
The bigger issue is that we're basically forcing people to avoid natural writing patterns now cause of the whole hysteria over whether something is written by an LLM or not. And, ironically, companies wasted no time monetizing the hysteria by using LLMs to decide whether something is legitimately human or not, cashing in on the whole thing and becoming the arbiters of what acceptable writing is.
The author correctly points out that if people stopped being lazy, and started engaging with the content instead of style then the whole problem would go away.
sjamaan | 4 hours ago
I think the problem is the lopsided equation that also exists for PRs: to review a PR or a paper costs a lot of time and energy. To find out you just wasted an hour (which you will never get back) on LLM slop someone took only minutes to prompt is deeply unfair. Hence the demand for services that purport to detect AI writing style (because they can't look at the content), and people attempting to do the same by hand. "It's not just being lazy, it's being smart about your time investment" (SCNR)
hoistbypetard | 4 hours ago
Not only is it deeply unfair to expect people to waste an hour reviewing something that is shaped like a valid thing and only took minutes to prompt, it's absolutely unsustainable. Even at the rate humans produce slop without an accelerant like this, it was hard to have enough reviewers to do a good job. Getting halfway through a review and realizing you're dealing with the output of a stochastic parrot is demoralizing and will chase people away from this under- (or un-) paid job that they were doing for the good of their sector.
You're not just being glib, you're right, too. Asking people to engage with a firehose of slop is awful.
[OP] Yogthos | 19 minutes ago
I understand the motivation, but it's pretty clear that the process creates a lot of harm in practice, and it doesn't actually address the problem either. We need different metrics other than style to discern whether writing has merit.
It is easy to ask an LLM to write in a certain tone, but having it produce reproducible results, honest error analysis, and clean ablations without actually doing the work is a much harder trick to pull off. Style detection punishes the form while these metrics would focus on the lack of substance. For example, arxiv started banning people for a year if they have hallucinated references in their papers. That's the kind of thing that makes sense to focus on.
travisgriggs | 3 hours ago
It’s like Seuss’s Sneetches. You pay to get stars on you belly, while others are paying to have them removed. Meanwhile, the McBeans get richer.
[OP] Yogthos | 41 minutes ago
exactly