Text is simply not information dense enough to be able to decode some arbitrary signal of provenance from it. Sure you might be able to detect today's tells (particular sentence structures preferred by Claude, phrases, etc) to get you some arbitrary chance percentage it was machine generated, but it's a bad fiction to perpetuate that any of this is anything more than tarot card reading.
Images, absolutely, there are tell-tale artifacts from today's generators that simply aren't emitted by "natural" paths to create them, and you can "detect AI" with high confidence (for now). Words, no, the signal is far too sparse and we are well into undetectable sophistication with today's models, let alone tomorrow's.
Whether a text was written by a human or not is just a single bit of information. So you can't rule out its detectability a priori, since even the shortest text contains more information than that.
As long as LLMs are used to write texts humans wouldn't want to write if they could help it (that's why they're getting an LLM to do it, after all), they'll remain detectable. Even if the reasoning might end up equivalent to "This looks like spam; no human in their right mind would write this spam by hand if they could get an LLM to write it, therefore it's most likely written by an LLM."
There are two problems, false positives and changing the LLM's pattern.
It's really easy to have a false positive and false positives can be very harmful if the person using the detector isn't aware of that risk.
It's also very easy to change the pattern of LLM output. You can provide basic prompting that will significantly change the structure of the output. For example, having it utilize the Wikipedia article on signs of AI writing and avoid everything it describes. https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
Not really. The false positives for the SOTA detector are very very low.
"It's also very easy to change the pattern of LLM output."
Not in a way that can reliably avoid detection. The problem is the patterns are baked into the distribution itself. It's smoothed over, so it becomes difficult to prompt your way out of that.
"Text is simply not information dense enough to be able to decode some arbitrary signal of provenance from it...it's a bad fiction to perpetuate that any of this is anything more than tarot card reading."
Not true at all. Pangram is highly effective and has a very low false positive rate.
The post here is impressive for a small project, it looks like they independently thought of one of the core ideas Pangram uses of creating twins to compare.
Signal is easier to detect with more data to work with.
Largely AI generated books are a vastly different situation than a one paragraph homework assignment. But multiple rounds of homework assignments would change the accuracy.
> but it's a bad fiction to perpetuate that any of this is anything more than tarot card reading.
Hard disagree. LLMs (especially base ones, that only received pre-training) can produce output that is undistinguishable from human writing (because that's what they were trained to do).
But commercial chat models are specifically tuned in a way that maximizes user engagement.
It's that specific tuning that is very easy to spot when reading AI slop, and that's not surprising that it's easy to spot automatically either.
And I don't think that's going to change anytime soon, unless their incentives change.
(We can say exactly the same thing about man-made stuff optimized for a specific purpose, like stock photography, clickbait titles or industrial food: they aren't stereotypical because their creator lacks the skill to make them otherwise, they are like that because that's what works best).
> But commercial chat models are specifically tuned in a way that maximizes user engagement. It's that specific tuning that is very easy to spot when reading AI slop, and that's not surprising that it's easy to spot automatically either.
There are two problems with this.
The first is that it would still misclassify human-authored text written under the same incentive, and most people have various incentives to "maximize engagement".
And the second is that then people would just make other models that are tuned for defeating that sort of classifier, which would be used whenever the classifier is being used.
All of that may be true, but pangram currently has a false positive rate of about 1 in 10000, and this has been tested by feeding in thousands of texts written before 2020.
That may not last if AI companies start trying to build models that fool it, but for the time being at least, modern models do have strong tells.
>and this has been tested by feeding in thousands of texts written before 2020.
And these text didn't train the model in the first place? I just want to ensure clarity on that.
>pangram currently has a false positive rate of about 1 in 10000
Says Panagram.
The problem with just looking at old text is language is a living thing. Say for example I make up the world 'oklambroahaha' right today. Both humans and AI pick up that word and start using it. Now lets say the model says that anything that uses oklambroahaha is 100% AI, you can't just point and say, "well my detection AI is correct on things 20 years old, so it's right skibbidy toilet 6/7".
There is a ton of evidence that use of AI changes the way we speak and write, so it will just turn these AI detectors into bullshit generating classifiers.
Pangram won't know how much AI written text they fail to detect, though, and detectors is a great tool to adjust methods of generating less AI-sounding text.
You can get an arbitrarily low false positive rate by sacrificing against false negatives. It's trivial to make it zero, just classify everything as human-generated. Meanwhile a false negative rate of even 1% is a pretty big problem since someone can easily use LLMs to generate 100x the volume of text and then use whichever ones make it through the classifier.
And that's before anyone even tries to get the LLM to generate a different style of text. Or for that matter creates a "style model" that rephrases text.
You don't really need a style model - current models are very good at doing "style transfer" of a model text onto whatever it has written if you just have it do it chunk by chunk. It takes more to prevent it from being detectable by good detectors, but it does remove a lot of the worst tells.
The point being that you wouldn't need the developers of the most popular models to themselves be trying to fool classifiers because their output could be run through an independent special purpose one designed to remove the tells the classifier is looking for, and the special purpose one wouldn't need to be made by anyone with the resources to create a good general-purpose model since it only has to do that one thing.
> The first is that it would still misclassify human-authored text written under the same incentive, and most people have various incentives to "maximize engagement".
The thing is, humans are significantly worse at maximizing numerical goals than computers.
> And the second is that then people would just make other models that are tuned for defeating that sort of classifier, which would be used whenever the classifier is being used.
Anyone can already do that right now, just grab unsloth studio and fine-tune your local Gemma, but nobody cares. People posting slop content don't care if pangram or I flag their slop with certainty, they are using the easiest option, which is commercial chat models. And given this segment of user doesn't care, the provider have zero incentive to provide a dedicated stealth model for that purpose.
I mean, back when I was spam filtering setting up a simple Bayesian classifier was easy. Train it on your spam and ham and it worked damned good. "Mission Accomplished".... until it wasn't. Spam rates started climbing and it started getting harder than ever to filter them.
There is always an incentive to get spam to bypass filters, so as your filters increase in accuracy, those attempting to pass said filters adjust their behaviors.
Spammers/cheaters/whateverers will at least just use a second pass filter that uses one of these 'ai scoring' systems to beat said AI scoring systems. So while it's worthwhile to do it at this moment, this window will rapidly close.
I don't think it's a very good remark, as there's significantly less email spam than 20 years ago.
Another example is ad-blocker-blocker. There was a little bit of an arm race between ad blockers and advertisers in the middle of the 2010s, but it didn't last long. Advertisers mostly just decided not to care about ad-blockers.
>Advertisers mostly just decided not to care about ad-blockers.
Directly not to care because they lost in court.
And yet the biggest advertizer on Earth (Google) decided to change their browser to make adblocking far more difficult. That or they say "just use an app, oh and turn on notifications". I'm not exactly sure who you think won the arms race there, but it seems like we the user did not.
There is significantly more spam than 20 years ago, just less of it reaches your inbox. This is a very important distinction as the cost of spam filtering is just as high as ever. On top of that most people have given up on their own email servers and instead depend on Google/Microsoft to do it for them. This allows these companies to have an overwhelming influence on email on the internet, to the point they can send spam with near impunity, and where if your system does it will be nuked from orbit by their systems.
And much like now Google supplies both the email spam, and the solution to the spam, they'll gladly supply the LLMs spam and the LLM solution while applying their 'flavor' of what's allowed to the entire internet.
> The thing is, humans are significantly worse at maximizing numerical goals than computers.
I'm not sure this is even the right premise.
Existing LLMs try to maximize engagement, and they often write in a particular style that has tells, but these two things are not necessarily related. Over-using em-dash or whatever isn't the thing that maximizes engagement.
So the two problems really are, what happens to the actual humans whose writing style is a close match for what a given generation of LLMs output? And, what stops LLMs from using a different style when someone wants to fool the classifier?
> People posting slop content don't care if pangram or I flag their slop with certainty, they are using the easiest option, which is commercial chat models.
They don't care as long as the consequences of identifying it are immaterial, but in that case what's the point of classifying it? Whereas if they need to fool the classifier some threshold percentage of the time in order for enough of their spam to get through, they're going to care.
> Over-using em-dash or whatever isn't the thing that maximizes engagement.
It's the thing that minimizes the loss during the RLHF phase, and the RLHF phase is the one that's aimed at maximizing engagement (it's literally trained on that).
> what happens to the actual humans whose writing style is a close match for what a given generation of LLMs output?
If a human, for instance because its writing gets polluted by reading too much AI slop, matches the style of an LLM closer than a certain threshold, then his own writing is going to be flagged as well. Whether it's an actual problem or merely a theoretical one is an open question. (unlike OpenAI and Anthropic, humans writers do have an incentive to avoid being flagged as AI).
> And, what stops LLMs from using a different style when someone wants to fool the classifier?
In theory: nothing. In practice if you fine-tune your own model: nothing. In practice with commercial models: the interests of the model making company.
> And, what stops LLMs from using a different style when someone wants to fool the classifier?
Websites have pretty much stopped using ad-blocker-blockers, it seems that it's not a fight worth fighting for them. Does that mean that ad-blockers are useless?
Most people don't even care about ads, I don't think they care about slop either, that's why there's slop posts and obnoxious websites that are unreadable without an ad blocker. A slop blocker used by 10-20% of the internet users wouldn't change the calculation more than ad blockers did.
They're also designed to not offend anybody, so their output tends to be very bland even compared to the most milquetoast of human beings. I was only surprised once when ChatGPT responded with an enthusiastic "hell yes" seemingly organically, but 99.9% of the time these AI services clearly are instructed and trained to provide flavorless word vomit. I don't think there's a technical reason why an LLM couldn't produce totally convincing output, but internet grifters don't need to go through that trouble. It's like how most phone, email, and social media scams come off as completely transparent to most of us, but that's the whole point; we're not the target audience of the scams. Readers looking for substance, nuance, and real opinions aren't going to notice if something with written by an LLM – unless there are some cliche punctuation tells.
When DANmode bypasses were a common thing the LLMs would drift significantly far from corporate speak.
But that's the point of corporate speak, you tend not to say thing that may offend your clients and deprive the company of future revenue. Of course there are some companies that make their living being 'counter-culture' and saying what they want, but they are a small percentage of all revenue.
It does mean that this will have a drift problem if it's just trained on the idiosyncrasies of model fine tuning. That's fine! But it is something to be aware of.
i think one thing overlooked by this perspective is that many of a detectors adversaries are not that sophisticated. so despite this i think it is a useful thing to try to do. particularly when people are trying to do fraud which will often having to use abliterated models and generally trying to be as economical in their efforts
Sure it is; we do it all the time, and then we modify each other's etc, etc; english we speak today was spoke yesterday waspake the same in yesteryears; we have no trouble dating english or other languages to a time.
A better argument is people themselves are just too influenced by reading that they'll sound like LLMs in a couple of years.
I think figuring out if a text is AI-made is a losing battle. What could work is gauging how much effort went into writing the text, regardless of who the author might be. What's easy today is generating mountains of text that are extremely hard to read. What requires effort is knowing how to engage the reader, how to keep out extraneous information, and how to keep the text as short as possible without losing details. That needs effort, with or without AI.
The easiest way is to keep track of the text's edit history, keeping a block of edits over time and having them signed by a timestamp authority. The final edit history can then be inspected by some external authority, then signed if the edit history looks human. I have a blog post from 2023 on this topic: https://helbl.ing/Written-Proof-of-Work/
For Google Doc users, you can already inspect the edit history over time to verify that text is written by a human.
That human might have used AI. You can never know. Hand fixed AI output, human just polished the corners? Light rewording of a full text written by hand, because the author is not confident in their writing? Actual human text, but after researching with AI?
I am working on a browser extension to help with that. Basically it interposes on any text field and canvas and if user pastes a large amount of text (copied form example from a chat bot), the extension will "replay" that text at normal, human-editing pace, and introduce typos that are fixed through later edits.
…All I know is that sometimes I will read e.g. a HN submission, and it becomes pretty obvious partway through that the article was AI generated.
If I can do it, an algorithm should be able to do it. Maybe in the the models will get so good that it is literally impossible to differentiate human vs computer authorship, but that’s obviously not the case today.
I've noticed there seems to be a default style that is easier to detect. I've noticed it harder to detect when asking an LLM to use a different style (more conversational, avoid sounding like an AI, don't use emdashes, etc). I wonder if that's what you're picking up too - the instances where people make no effort to change the style of the output.
Sufficiently advanced AI use is probably fine. The slop everyone complains about has certain tells specifically due to some combination of the following:
- The author is conducting some kind of hustle.
- The author doesn't bother editing.
- The author lacks the taste and awareness enough to see it looks.
- The author thinks you, the reader, lack taste and awareness.
- The author is using it as a kind of smoke bomb to get rid of you.
In such cases, nothing is done about the LLM's distinctive "voice". It dominates the text and it's easy to detect. It stands as a signifier of the above, even if it's otherwise not intrinsically a problem to use AI.
The classifier does not seem so big, I wonder if something like it for English could be used in a browser extension to run against every single paragraph being displayed ?
If the internet is going to drown in LLM text it would be nice to have tools to detect that automatically just like we have adblockers today to avoid wasting time on ads.
The article mentions that AI texts are often caught by multiple models, so hopefully text from newer LLMs could still be caught without updating the model?
> Eventually, I faked my way through the thesis, and life moved on.
This is a very startling admission! I checked the Chinese (original?) version of the post, and saw the author uses the word "糊弄" (in the place of "faked"); I'm not a native speaker but I think this may come across more as a self-effacing comment on the low quality and/or effort behind their thesis, whereas the English version implies fraud. May be wise to change this!
I don't know if the Chinese text implies something different, but I think even in English it's pretty normal for people to claim they 'faked' their way through something without referring to fraud.
E.g. "I faked my way through the interview!" = "I did my best to respond to questions I did not feel fully prepared for, and managed to get through the interview"
Well cheated would definitely imply fraud. “Faking it” as in “fake it till you make it” is more like pretending you know about a topic until you learn enough on the job to participate competently.
Small encoder-only transformers are excellent at classifying LLM-Generated Text. I built an on-device iOS app using a custom small encoder that achieves an AUROC of 99.81 on RAID-bench.
I could be wrong, but I just don’t see how trying to “detect” LLM generated texts is ever going to work. The only thing that makes any sense if you truly want to have confidence a human wrote it is some type of “proof of work“ system. I think there’s a lot of interesting ways to approach the proof of work problem with different pros and cons, but that is where our energy should be focused if we seriously want to solve this problem.
I think the fundamental problem is that training current SOTA AI models is very expensive. If a simple "classical" model can detect them, presumably at much lower algorithmic cost, then why wouldn't the model trainers use these same tools to feed back into their models to improve them at low cost to make them better? It's an arms race. Any cheap pattern can and presumably will be used to retrain if it becomes and effective way to catch AI.
In part because model vendors specifically prefer when people think that lots of content is produced by their model. The more Claude-like writing appears on the internet, the more signal there is to investors that people are using Claude for a greater number tasks.
It’s simply not a priority. The labs can do many things. Making text non-LLM is not really that useful. Analogous to Facebook not picking up the obvious $20 bill in front of them. It’s because they’ve got $100 bills at their feet they’re picking up.
Not a priority currently. Selling services to spammers... I mean marketers is still big money and eventually someone will pick it up. If training costs ever drop, then it's one of the first things that will happen.
Neat. I will implement something like this for myself. I just need to reduce the spam a little. Imperfection is okay for a social network context like HN.
I had done the same for classifying and generating bookmarks of thousands of datasheets, along with a very naive yolo-based classificator (to detect pages made out of diagrams and pictures mostly).
Done with GLM-OCR, I had to watch text sloooowly crawl out of the llm and still have to live with hallucinations and the model not following the schema
The problems are simply too great if an LLM detector has any false positives at all. Imagine how soul-crushing writing an entire dissertation by hand and having it rejected because some “good enough” LLM detector decides you write too much like an AI.
As I recall, a few years ago (in the era of first generation LLMs), a professor in Texas used an anti-plagiarism tool that flagged more than one-third of the class using AI in an exam, and used that finding to give them a failing grade.
If memory serves, one student objected strenously and ran the professor's own work (published 10 years earlier) into the same tool and it flagged that work as AI-generated.
Exactly. The more corporate and proper you tend to speak, the more likely it's to classify you as an LLM. It's like the classifiers want us to talk like trash at their current rate. This seems to be really problematic for ESL speakers/typers that may have been trained on a smaller, more proper subset of the language.
Am I the only who largely enjoys the output of LLMs more than most stuff written by humans? I find myself coming back to old chats with ChatGPT frequently because the output is amazing.
I'm not late if people constantly put effort into finding LLM text, or every other comment on hacker news is either about something being LLM generator.
After seeing comments on hacker news attempt to call an article from 2015 as generated by an LLM, I have very little faith in commenters having any ability in actually detecting AI written text.
And that's just one particularly egregious case I remember. Posters that are technical writers or use English properly get called bots quite commonly when their post history shows a writing style going back over a decade.
But now that LLMs are causing a language drift in English users our filters of "that's an LLM" will become even more useless.
there is not much point in detecting LLM generated text, in that humans are useing info from LLM's, but obfusicting it's origin, with there own garble, along with purely human garble, and almost(but not quite) human LLM product meaning that the threshold for rejecting "data" must be lowered, which personaly means a very very low tollerance for wierdness, except where it can yield imediate possitive cash flow
for the rest I do my own research and verification thank you very much
Text is simply not information dense enough to be able to decode some arbitrary signal of provenance from it. Sure you might be able to detect today's tells (particular sentence structures preferred by Claude, phrases, etc) to get you some arbitrary chance percentage it was machine generated, but it's a bad fiction to perpetuate that any of this is anything more than tarot card reading.
Images, absolutely, there are tell-tale artifacts from today's generators that simply aren't emitted by "natural" paths to create them, and you can "detect AI" with high confidence (for now). Words, no, the signal is far too sparse and we are well into undetectable sophistication with today's models, let alone tomorrow's.
Whether a text was written by a human or not is just a single bit of information. So you can't rule out its detectability a priori, since even the shortest text contains more information than that.
As long as LLMs are used to write texts humans wouldn't want to write if they could help it (that's why they're getting an LLM to do it, after all), they'll remain detectable. Even if the reasoning might end up equivalent to "This looks like spam; no human in their right mind would write this spam by hand if they could get an LLM to write it, therefore it's most likely written by an LLM."
There are two problems, false positives and changing the LLM's pattern.
It's really easy to have a false positive and false positives can be very harmful if the person using the detector isn't aware of that risk.
It's also very easy to change the pattern of LLM output. You can provide basic prompting that will significantly change the structure of the output. For example, having it utilize the Wikipedia article on signs of AI writing and avoid everything it describes. https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
"It's really easy to have a false positive"
Not really. The false positives for the SOTA detector are very very low.
"It's also very easy to change the pattern of LLM output."
Not in a way that can reliably avoid detection. The problem is the patterns are baked into the distribution itself. It's smoothed over, so it becomes difficult to prompt your way out of that.
"Text is simply not information dense enough to be able to decode some arbitrary signal of provenance from it...it's a bad fiction to perpetuate that any of this is anything more than tarot card reading."
Not true at all. Pangram is highly effective and has a very low false positive rate.
The post here is impressive for a small project, it looks like they independently thought of one of the core ideas Pangram uses of creating twins to compare.
You can see how it works here: https://arxiv.org/pdf/2402.14873
This sounds like it was edited by an llm.
Signal is easier to detect with more data to work with.
Largely AI generated books are a vastly different situation than a one paragraph homework assignment. But multiple rounds of homework assignments would change the accuracy.
> but it's a bad fiction to perpetuate that any of this is anything more than tarot card reading.
Hard disagree. LLMs (especially base ones, that only received pre-training) can produce output that is undistinguishable from human writing (because that's what they were trained to do).
But commercial chat models are specifically tuned in a way that maximizes user engagement. It's that specific tuning that is very easy to spot when reading AI slop, and that's not surprising that it's easy to spot automatically either. And I don't think that's going to change anytime soon, unless their incentives change.
(We can say exactly the same thing about man-made stuff optimized for a specific purpose, like stock photography, clickbait titles or industrial food: they aren't stereotypical because their creator lacks the skill to make them otherwise, they are like that because that's what works best).
> especially base ones
Did you actually try them? I did.They generated even more "slopey" text than instruction-tuned ones.
> But commercial chat models are specifically tuned in a way that maximizes user engagement. It's that specific tuning that is very easy to spot when reading AI slop, and that's not surprising that it's easy to spot automatically either.
There are two problems with this.
The first is that it would still misclassify human-authored text written under the same incentive, and most people have various incentives to "maximize engagement".
And the second is that then people would just make other models that are tuned for defeating that sort of classifier, which would be used whenever the classifier is being used.
All of that may be true, but pangram currently has a false positive rate of about 1 in 10000, and this has been tested by feeding in thousands of texts written before 2020.
That may not last if AI companies start trying to build models that fool it, but for the time being at least, modern models do have strong tells.
>and this has been tested by feeding in thousands of texts written before 2020.
And these text didn't train the model in the first place? I just want to ensure clarity on that.
>pangram currently has a false positive rate of about 1 in 10000
Says Panagram.
The problem with just looking at old text is language is a living thing. Say for example I make up the world 'oklambroahaha' right today. Both humans and AI pick up that word and start using it. Now lets say the model says that anything that uses oklambroahaha is 100% AI, you can't just point and say, "well my detection AI is correct on things 20 years old, so it's right skibbidy toilet 6/7".
There is a ton of evidence that use of AI changes the way we speak and write, so it will just turn these AI detectors into bullshit generating classifiers.
Pangram won't know how much AI written text they fail to detect, though, and detectors is a great tool to adjust methods of generating less AI-sounding text.
You can get an arbitrarily low false positive rate by sacrificing against false negatives. It's trivial to make it zero, just classify everything as human-generated. Meanwhile a false negative rate of even 1% is a pretty big problem since someone can easily use LLMs to generate 100x the volume of text and then use whichever ones make it through the classifier.
And that's before anyone even tries to get the LLM to generate a different style of text. Or for that matter creates a "style model" that rephrases text.
You don't really need a style model - current models are very good at doing "style transfer" of a model text onto whatever it has written if you just have it do it chunk by chunk. It takes more to prevent it from being detectable by good detectors, but it does remove a lot of the worst tells.
The point being that you wouldn't need the developers of the most popular models to themselves be trying to fool classifiers because their output could be run through an independent special purpose one designed to remove the tells the classifier is looking for, and the special purpose one wouldn't need to be made by anyone with the resources to create a good general-purpose model since it only has to do that one thing.
> The first is that it would still misclassify human-authored text written under the same incentive, and most people have various incentives to "maximize engagement".
The thing is, humans are significantly worse at maximizing numerical goals than computers.
> And the second is that then people would just make other models that are tuned for defeating that sort of classifier, which would be used whenever the classifier is being used.
Anyone can already do that right now, just grab unsloth studio and fine-tune your local Gemma, but nobody cares. People posting slop content don't care if pangram or I flag their slop with certainty, they are using the easiest option, which is commercial chat models. And given this segment of user doesn't care, the provider have zero incentive to provide a dedicated stealth model for that purpose.
I mean, back when I was spam filtering setting up a simple Bayesian classifier was easy. Train it on your spam and ham and it worked damned good. "Mission Accomplished".... until it wasn't. Spam rates started climbing and it started getting harder than ever to filter them.
There is always an incentive to get spam to bypass filters, so as your filters increase in accuracy, those attempting to pass said filters adjust their behaviors.
Spammers/cheaters/whateverers will at least just use a second pass filter that uses one of these 'ai scoring' systems to beat said AI scoring systems. So while it's worthwhile to do it at this moment, this window will rapidly close.
I don't think it's a very good remark, as there's significantly less email spam than 20 years ago.
Another example is ad-blocker-blocker. There was a little bit of an arm race between ad blockers and advertisers in the middle of the 2010s, but it didn't last long. Advertisers mostly just decided not to care about ad-blockers.
>Advertisers mostly just decided not to care about ad-blockers.
Directly not to care because they lost in court.
And yet the biggest advertizer on Earth (Google) decided to change their browser to make adblocking far more difficult. That or they say "just use an app, oh and turn on notifications". I'm not exactly sure who you think won the arms race there, but it seems like we the user did not.
There is significantly more spam than 20 years ago, just less of it reaches your inbox. This is a very important distinction as the cost of spam filtering is just as high as ever. On top of that most people have given up on their own email servers and instead depend on Google/Microsoft to do it for them. This allows these companies to have an overwhelming influence on email on the internet, to the point they can send spam with near impunity, and where if your system does it will be nuked from orbit by their systems.
And much like now Google supplies both the email spam, and the solution to the spam, they'll gladly supply the LLMs spam and the LLM solution while applying their 'flavor' of what's allowed to the entire internet.
> The thing is, humans are significantly worse at maximizing numerical goals than computers.
I'm not sure this is even the right premise.
Existing LLMs try to maximize engagement, and they often write in a particular style that has tells, but these two things are not necessarily related. Over-using em-dash or whatever isn't the thing that maximizes engagement.
So the two problems really are, what happens to the actual humans whose writing style is a close match for what a given generation of LLMs output? And, what stops LLMs from using a different style when someone wants to fool the classifier?
> People posting slop content don't care if pangram or I flag their slop with certainty, they are using the easiest option, which is commercial chat models.
They don't care as long as the consequences of identifying it are immaterial, but in that case what's the point of classifying it? Whereas if they need to fool the classifier some threshold percentage of the time in order for enough of their spam to get through, they're going to care.
> Over-using em-dash or whatever isn't the thing that maximizes engagement.
It's the thing that minimizes the loss during the RLHF phase, and the RLHF phase is the one that's aimed at maximizing engagement (it's literally trained on that).
> what happens to the actual humans whose writing style is a close match for what a given generation of LLMs output?
If a human, for instance because its writing gets polluted by reading too much AI slop, matches the style of an LLM closer than a certain threshold, then his own writing is going to be flagged as well. Whether it's an actual problem or merely a theoretical one is an open question. (unlike OpenAI and Anthropic, humans writers do have an incentive to avoid being flagged as AI).
> And, what stops LLMs from using a different style when someone wants to fool the classifier?
In theory: nothing. In practice if you fine-tune your own model: nothing. In practice with commercial models: the interests of the model making company.
> And, what stops LLMs from using a different style when someone wants to fool the classifier?
Websites have pretty much stopped using ad-blocker-blockers, it seems that it's not a fight worth fighting for them. Does that mean that ad-blockers are useless?
Most people don't even care about ads, I don't think they care about slop either, that's why there's slop posts and obnoxious websites that are unreadable without an ad blocker. A slop blocker used by 10-20% of the internet users wouldn't change the calculation more than ad blockers did.
They're also designed to not offend anybody, so their output tends to be very bland even compared to the most milquetoast of human beings. I was only surprised once when ChatGPT responded with an enthusiastic "hell yes" seemingly organically, but 99.9% of the time these AI services clearly are instructed and trained to provide flavorless word vomit. I don't think there's a technical reason why an LLM couldn't produce totally convincing output, but internet grifters don't need to go through that trouble. It's like how most phone, email, and social media scams come off as completely transparent to most of us, but that's the whole point; we're not the target audience of the scams. Readers looking for substance, nuance, and real opinions aren't going to notice if something with written by an LLM – unless there are some cliche punctuation tells.
When DANmode bypasses were a common thing the LLMs would drift significantly far from corporate speak.
But that's the point of corporate speak, you tend not to say thing that may offend your clients and deprive the company of future revenue. Of course there are some companies that make their living being 'counter-culture' and saying what they want, but they are a small percentage of all revenue.
It does mean that this will have a drift problem if it's just trained on the idiosyncrasies of model fine tuning. That's fine! But it is something to be aware of.
i think one thing overlooked by this perspective is that many of a detectors adversaries are not that sophisticated. so despite this i think it is a useful thing to try to do. particularly when people are trying to do fraud which will often having to use abliterated models and generally trying to be as economical in their efforts
Sure it is; we do it all the time, and then we modify each other's etc, etc; english we speak today was spoke yesterday waspake the same in yesteryears; we have no trouble dating english or other languages to a time.
A better argument is people themselves are just too influenced by reading that they'll sound like LLMs in a couple of years.
It depends on how much text. For example, chardet often falls down on short strings, but 1K characters it nails it.
I think figuring out if a text is AI-made is a losing battle. What could work is gauging how much effort went into writing the text, regardless of who the author might be. What's easy today is generating mountains of text that are extremely hard to read. What requires effort is knowing how to engage the reader, how to keep out extraneous information, and how to keep the text as short as possible without losing details. That needs effort, with or without AI.
The easiest way is to keep track of the text's edit history, keeping a block of edits over time and having them signed by a timestamp authority. The final edit history can then be inspected by some external authority, then signed if the edit history looks human. I have a blog post from 2023 on this topic: https://helbl.ing/Written-Proof-of-Work/
For Google Doc users, you can already inspect the edit history over time to verify that text is written by a human.
That human might have used AI. You can never know. Hand fixed AI output, human just polished the corners? Light rewording of a full text written by hand, because the author is not confident in their writing? Actual human text, but after researching with AI?
Exactly. Detecting AI writing is an arms race that can only end with detection coming in second place.
I am working on a browser extension to help with that. Basically it interposes on any text field and canvas and if user pastes a large amount of text (copied form example from a chat bot), the extension will "replay" that text at normal, human-editing pace, and introduce typos that are fixed through later edits.
Any specific reason as to why you'd want to make that, outside of intentionally enabling fraud?
…All I know is that sometimes I will read e.g. a HN submission, and it becomes pretty obvious partway through that the article was AI generated.
If I can do it, an algorithm should be able to do it. Maybe in the the models will get so good that it is literally impossible to differentiate human vs computer authorship, but that’s obviously not the case today.
How do you estimate your false negative rate?
I've noticed there seems to be a default style that is easier to detect. I've noticed it harder to detect when asking an LLM to use a different style (more conversational, avoid sounding like an AI, don't use emdashes, etc). I wonder if that's what you're picking up too - the instances where people make no effort to change the style of the output.
Sufficiently advanced AI use is probably fine. The slop everyone complains about has certain tells specifically due to some combination of the following:
- The author is conducting some kind of hustle.
- The author doesn't bother editing.
- The author lacks the taste and awareness enough to see it looks.
- The author thinks you, the reader, lack taste and awareness.
- The author is using it as a kind of smoke bomb to get rid of you.
In such cases, nothing is done about the LLM's distinctive "voice". It dominates the text and it's easy to detect. It stands as a signifier of the above, even if it's otherwise not intrinsically a problem to use AI.
There cant be a way / except of course if you pay / mind to my syllables
The classifier does not seem so big, I wonder if something like it for English could be used in a browser extension to run against every single paragraph being displayed ?
If the internet is going to drown in LLM text it would be nice to have tools to detect that automatically just like we have adblockers today to avoid wasting time on ads.
(the article was a good read, thanks!)
I assume different models will have different distribution, so it has to be kept updated?
The article mentions that AI texts are often caught by multiple models, so hopefully text from newer LLMs could still be caught without updating the model?
You know what GAN is, right?
In training all you have to do is take their model as the adversary and then it's useless.
> Eventually, I faked my way through the thesis, and life moved on.
This is a very startling admission! I checked the Chinese (original?) version of the post, and saw the author uses the word "糊弄" (in the place of "faked"); I'm not a native speaker but I think this may come across more as a self-effacing comment on the low quality and/or effort behind their thesis, whereas the English version implies fraud. May be wise to change this!
I don't know if the Chinese text implies something different, but I think even in English it's pretty normal for people to claim they 'faked' their way through something without referring to fraud.
E.g. "I faked my way through the interview!" = "I did my best to respond to questions I did not feel fully prepared for, and managed to get through the interview"
Well cheated would definitely imply fraud. “Faking it” as in “fake it till you make it” is more like pretending you know about a topic until you learn enough on the job to participate competently.
Small encoder-only transformers are excellent at classifying LLM-Generated Text. I built an on-device iOS app using a custom small encoder that achieves an AUROC of 99.81 on RAID-bench.
I could be wrong, but I just don’t see how trying to “detect” LLM generated texts is ever going to work. The only thing that makes any sense if you truly want to have confidence a human wrote it is some type of “proof of work“ system. I think there’s a lot of interesting ways to approach the proof of work problem with different pros and cons, but that is where our energy should be focused if we seriously want to solve this problem.
> I just don’t see how trying to “detect” LLM generated texts is ever going to work
He literally demonstrated a working system in this post. Do you mean you'll never get to 100% accuracy? Clearly, but you don't need that.
I wonder about this technique vs simple SVM classifiers: https://x.com/rosmine/status/2056406399471558872?s=20
This article is about training a classifier to detect synthetic text.
The link you sent is for generating text which attempts to defeat those classifiers.
I think the fundamental problem is that training current SOTA AI models is very expensive. If a simple "classical" model can detect them, presumably at much lower algorithmic cost, then why wouldn't the model trainers use these same tools to feed back into their models to improve them at low cost to make them better? It's an arms race. Any cheap pattern can and presumably will be used to retrain if it becomes and effective way to catch AI.
In part because model vendors specifically prefer when people think that lots of content is produced by their model. The more Claude-like writing appears on the internet, the more signal there is to investors that people are using Claude for a greater number tasks.
It’s simply not a priority. The labs can do many things. Making text non-LLM is not really that useful. Analogous to Facebook not picking up the obvious $20 bill in front of them. It’s because they’ve got $100 bills at their feet they’re picking up.
Not a priority currently. Selling services to spammers... I mean marketers is still big money and eventually someone will pick it up. If training costs ever drop, then it's one of the first things that will happen.
It’s an arms race where the AI companies are at an extreme disadvantage due to relative training costs.
Neat. I will implement something like this for myself. I just need to reduce the spam a little. Imperfection is okay for a social network context like HN.
It will work for a bit, but as people start speaking more like LLMs and LLMs start training using said classifiers as a GAN, it will become useless.
I had done the same for classifying and generating bookmarks of thousands of datasheets, along with a very naive yolo-based classificator (to detect pages made out of diagrams and pictures mostly).
Done with GLM-OCR, I had to watch text sloooowly crawl out of the llm and still have to live with hallucinations and the model not following the schema
The problems are simply too great if an LLM detector has any false positives at all. Imagine how soul-crushing writing an entire dissertation by hand and having it rejected because some “good enough” LLM detector decides you write too much like an AI.
As I recall, a few years ago (in the era of first generation LLMs), a professor in Texas used an anti-plagiarism tool that flagged more than one-third of the class using AI in an exam, and used that finding to give them a failing grade.
If memory serves, one student objected strenously and ran the professor's own work (published 10 years earlier) into the same tool and it flagged that work as AI-generated.
EDIT: HN item from June 2023 https://news.ycombinator.com/item?id=36215823
Exactly. The more corporate and proper you tend to speak, the more likely it's to classify you as an LLM. It's like the classifiers want us to talk like trash at their current rate. This seems to be really problematic for ESL speakers/typers that may have been trained on a smaller, more proper subset of the language.
It depends on the application. Dissertation? Hell naw. Blog post? Absolutely, run it through that thing.
The problem is that ed-tech is absolutely ravenous for an LLM detector and would rather use snake oil than accept that it might not be possible.
Am I the only who largely enjoys the output of LLMs more than most stuff written by humans? I find myself coming back to old chats with ChatGPT frequently because the output is amazing.
I wouldn’t go that far… but it can be kinda like Wikipedia, clean and readable.
Anything too “clever” and “snappy” = instaLLM
This is also how I pretty much filter LLM generated text in my head.
today, sure.
Tomorrow, the LLMs will be training the humans thought patterns that will directly start skewing their natural writing.
Generation alpha is going to have a lot of trouble if we keep perpetuating the myth that you can really interpret text in an ongoing fashion.
I think you're about a year late for this revolation.
https://www.washingtonpost.com/opinions/2025/08/20/chatgpt-c...
I'm not late if people constantly put effort into finding LLM text, or every other comment on hacker news is either about something being LLM generator.
After seeing comments on hacker news attempt to call an article from 2015 as generated by an LLM, I have very little faith in commenters having any ability in actually detecting AI written text.
And that's just one particularly egregious case I remember. Posters that are technical writers or use English properly get called bots quite commonly when their post history shows a writing style going back over a decade.
But now that LLMs are causing a language drift in English users our filters of "that's an LLM" will become even more useless.
there is not much point in detecting LLM generated text, in that humans are useing info from LLM's, but obfusicting it's origin, with there own garble, along with purely human garble, and almost(but not quite) human LLM product meaning that the threshold for rejecting "data" must be lowered, which personaly means a very very low tollerance for wierdness, except where it can yield imediate possitive cash flow for the rest I do my own research and verification thank you very much