How broken the AI-content detection system is

[Snoocums]

My stories are routinely rejected, accused of being written (at least in part) with the help of AI.

At this point, you’re probably thinking, "So what? You’re not the first and you won't be the last. Get in line, bro!"

But let me give you the whole story, chapter and verse, so you can see just how inconsistent, ridiculous, and self-defeating this site's underlying system really is.

First off, you need to know that I’m an Italian author, which means I'm not a native English speaker. I publish erotica in Italian, and a while back, I figured I’d try my hand at English too. So, I did the most logical thing: I took my old Italian stories and translated them.
To be absolutely clear, the translation was done by yours truly: a human being with two arms, two legs, two eyes, a dick, two balls, and a brain made of neurons. Not a machine. I spent hours consulting dictionaries and thesauri to find synonyms, vary the vocabulary, refine the phrasing, and strive to give every character their own unique voice (which is hard enough in your mother tongue, let alone a foreign language).

Anyway, I submitted these human-translated stories to Literotica and... they were rejected. Flagged for AI.
I reviewed them, spending even more hours trying to figure out what went wrong, looking for any pattern that might give off AI vibes.
I resubmitted... rejected again. Flagged for AI.
I sent them back one more time, adding a note stating that this is 100% my own work and explicitly asking for a review by a human reader... rejected yet again. Flagged for AI.
This happened with the first story. Then the second story. And the third story.

By the fourth one, I decided to do something you’re really not supposed to do. But I was so desperate, and I wanted to see just how far this twisted system of false positives actually goes, so I did this: I took one of my Italian stories, dumped it into Google Translate, and ran an automated translation. What came out was a mix of rigid, textbook English, unidiomatic phrasing, and literal translations from Italian that you native speakers might manage to understand, but will definitely find clumsy. To give you an idea, here is a little taste of that Google Translated text:

This year, as we've been doing for some time now during the winter, Francesca and I treated ourselves to a relaxing and wellness weekend at a spa in Slovenia.
Even if it was just for the weekend, completely disconnecting from our routine, turning off our phones and abandoning ourselves to the warmth of the pools, saunas, and Turkish baths, recharged our batteries to last until spring.
The day to set off had arrived; we'd booked a new place a few weeks ago, so the countdown had only heightened our desire to leave.
During the trip, however, bad luck had it that my car suddenly stalled several times... I couldn't continue driving on the highway in these conditions; it could be dangerous, so I reluctantly decided to pull over and park at a lay-by.
I called a tow truck in complete despair; they arrived and took us to a mechanic who assessed the problem and told us the car would be ready in the late afternoon.
This means we'll be arriving at the spa very late and will have wasted one of our two relaxing days.
You can imagine our state of mind, but on the other hand, we couldn't do anything except hope the mechanic would perform a miracle.
Moral of the story: we leave at 6:30 PM and arrive at our destination at 8:30 PM... we're surprised to see the facility... on the internet it looked completely different, but in person it's small and even a bit dated... this certainly doesn't help our morale, on the contrary.

So, I took this text, and without even giving it a manual pass, I uploaded the whole thing to Literotica.

The result? The story got accepted.
(Of course, I pulled it back to my drafts before it actually got published, not only because it's definitely not a text I'd want native speakers to read, but also because publishing a Google-translated story just felt ethically wrong.)

Do you see the paradox now?
A story meticulously reviewed and polished by a human is flagged as AI. A story blatantly translated by Google Translate is accepted as human.
I have no idea (and I guess nobody does) what kind of AI detectors the site is using, but they are doing a truly phenomenal job if a text processed by Google Translate slips right through while stories meticulously written and polished by a human author get flagged.
They want to crack down on AI use, but the paradox is that if a foreign author actually wants to get published on their site, using AI is the absolute best shortcut!

 

[Esposito_Okami]

I'm an Italian native speaker too, although I write directly in English. I don't claim my English is at a native-speaker level, but it's certainly not written with the help of AI. Yet, without fail, Literotica rejects my stories, accusing me of using AI.

At this point, it's almost as if they're encouraging Italian authors to write in their own language, run the text through Google Translate, and still get it published...

 

[EmilyMiller]

My stories are routinely rejected, accused of being written (at least in part) with the help of AI.

At this point, you’re probably thinking, "So what? You’re not the first and you won't be the last. Get in line, bro!"

But let me give you the whole story, chapter and verse, so you can see just how inconsistent, ridiculous, and self-defeating this site's underlying system really is.

First off, you need to know that I’m an Italian author, which means I'm not a native English speaker. I publish erotica in Italian, and a while back, I figured I’d try my hand at English too. So, I did the most logical thing: I took my old Italian stories and translated them.
To be absolutely clear, the translation was done by yours truly: a human being with two arms, two legs, two eyes, a dick, two balls, and a brain made of neurons. Not a machine. I spent hours consulting dictionaries and thesauri to find synonyms, vary the vocabulary, refine the phrasing, and strive to give every character their own unique voice (which is hard enough in your mother tongue, let alone a foreign language).

Anyway, I submitted these human-translated stories to Literotica and... they were rejected. Flagged for AI.
I reviewed them, spending even more hours trying to figure out what went wrong, looking for any pattern that might give off AI vibes.
I resubmitted... rejected again. Flagged for AI.
I sent them back one more time, adding a note stating that this is 100% my own work and explicitly asking for a review by a human reader... rejected yet again. Flagged for AI.
This happened with the first story. Then the second story. And the third story.

By the fourth one, I decided to do something you’re really not supposed to do. But I was so desperate, and I wanted to see just how far this twisted system of false positives actually goes, so I did this: I took one of my Italian stories, dumped it into Google Translate, and ran an automated translation. What came out was a mix of rigid, textbook English, unidiomatic phrasing, and literal translations from Italian that you native speakers might manage to understand, but will definitely find clumsy. To give you an idea, here is a little taste of that Google Translated text:


So, I took this text, and without even giving it a manual pass, I uploaded the whole thing to Literotica.

The result? The story got accepted.
(Of course, I pulled it back to my drafts before it actually got published, not only because it's definitely not a text I'd want native speakers to read, but also because publishing a Google-translated story just felt ethically wrong.)

Do you see the paradox now?
A story meticulously reviewed and polished by a human is flagged as AI. A story blatantly translated by Google Translate is accepted as human.
I have no idea (and I guess nobody does) what kind of AI detectors the site is using, but they are doing a truly phenomenal job if a text processed by Google Translate slips right through while stories meticulously written and polished by a human author get flagged.
They want to crack down on AI use, but the paradox is that if a foreign author actually wants to get published on their site, using AI is the absolute best shortcut!

That’s wild, sorry for your experience. It seems like the current situation is not good for those with English as a second language.

 

[Louys90]

I knew the system the site uses to decide what is and isn’t AI-generated had some flaws, but this... wow... just wow.

 

[SmilingLez]

Even the top rated AI detection tool is flawed, and I doubt a top-rated AI detection tool is used to screen stories here. If you submitted the Gettysburg address, it would be rejected with the AI automatically generated text.

 

[qidniz]

Even the top rated AI detection tool is flawed, and I doubt a top-rated AI detection tool is used to screen stories here. If you submitted the Gettysburg address, it would be rejected with the AI automatically generated text.

I tried the Gettysburg Address on an online detector (aidetector.com), and it gave a 4% AI rating (not 0%!) to a text it recognized in advance. The analysis said

The provided text matches the Gettysburg Address verbatim, exhibiting 19th-century oratorical style with irregular sentence lengths, deliberate anaphora, and a single unbroken paragraph rather than the symmetrical multi-paragraph blocks common in LLM output. Its language contains period-specific diction and a measured but non-monotonic emotional arc focused on national purpose, without the predictable virtue-listing or inspirational framing that ends LLM paragraphs. Because the content is a well-documented historical artifact predating all modern language models by more than a century, the probability of AI authorship is negligible.

Why isn't 'negligible' 0%?

So I tried it on my own stuff, straight from my text editor. It came back with 15% AI, with this analysis

Strong human indicators include variable sentence lengths, abrupt emotional shifts from casual banter to panic, incomplete thoughts, slang like [redacted] and [redacted], and genuine uncertainty in the characters' voices. The story shows lived-in family dynamics and raw reactions rather than polished inspirational arcs or encyclopedic summaries. Minor repetitions and the abrupt cutoff further support organic authorship over mechanical generation. No consistent 4-6 sentence paragraphs or seamless virtue-based conclusions appear.

All that and still 15% AI? I suppose that if the G'Burg Address can't get a 0% AI rating, nothing can.

 

[SmilingLez]

I tried the Gettysburg Address on an online detector (aidetector.com), and it gave a 4% AI rating (not 0%!) to a text it recognized in advance. The analysis said


Why isn't 'negligible' 0%?

So I tried it on my own stuff, straight from my text editor. It came back with 15% AI, with this analysis


All that and still 15% AI? I suppose that if the G'Burg Address can't get a 0% AI rating, nothing can.

When I put several sections of the Gettysburg Address in ZeroGPT(free and online) it came back as 100% AI generated. Then I tried Martin Luther King's I've got a dream speech, several paragraphs once again, and it came back as 0% AI generated

 

[EmilyMiller]

I tried the Gettysburg Address on an online detector (aidetector.com), and it gave a 4% AI rating (not 0%!)

This evaluation below is from a reputable AI detector, Pangram. Pangram is a Machine Learning model trained on text produced by humans and by AI with the same prompts / on the same subjects. It’s not an LLM (LLMs are are a specific type of machine learning model with other functions wrapped around them). I wish we’d stop with the Shakespeare and Lincoln text being flagged as AI - by Mickey Mouse AI detectors maybe.

Note: Even Pangram is fallible - it has false positives and negatives. All AI detectors are fallible. Some - like the ones that some people cite to discredit all AI checking - are fucking awful. Others are just unreliable.

 

[qidniz]

Even Pangram is fallible - it has false positives and negatives. All AI detectors are fallible. Some - like the ones that some people cite to discredit all AI checking - are fucking awful. Others are just unreliable.

My interest in this subject is quite minimal, basically limited to the chance that a submission of mine runs afoul of whatever AI detection mechanism Lit uses. Maybe they use Pangram, or maybe they use something worse (if Pangram is the "best"). Without a statement from TPTB, there is no reason to assume that what they actually use is any good.

The fact that run of the mill AI checkers are not particularly good in general - I don't think this is a controversial statement - basically means that we are helpless against random false positives.

 

[SmilingLez]

Any AI company can skew data to make themselves look better.. Marketing hype with one purpose and that is to make their company money. And if you believe any AI is at least 90% accurate all of the time, I've got some beachfront property in Kentucky for sale

https://gptzero.me/news/gptzero-vs-pangram/

 

[EmilyMiller]

The fact that run of the mill AI checkers are not particularly good in general - I don't think this is a controversial statement

1. The free ones you can use on-line
2. Asking ChatGPT or another LLM to opine
3. Pangram
4. Commercial AI checkers

From convos, I believe Lit uses option 4. But I have no idea what exactly it is. Nor does anyone else. It’s certainly not Biff’s free AI check-o-matic, whose owner wants to sell you their de-AI-ification tool.

 

[SkyBubble]

When I put several sections of the Gettysburg Address in ZeroGPT(free and online) it came back as 100% AI generated. Then I tried Martin Luther King's I've got a dream speech, several paragraphs once again, and it came back as 0% AI generated

You think that's bad, that Will Shakespeare guy was a sneaky AI user, according to the AI detectors.

 

[EmilyMiller]

if Pangram is the "best"

I’ve run segments of my own work through Pangram. Just out of curiosity, I don’t use either genAI or grammar checkers, just human alpha and beta readers.

It has never flagged my work as AI. What normally happens is I either get 100% human + high confidence. Or some large percent human + high confidence and some low percent human + medium confidence.

When I look at the sections which have only medium confidence, it’s normally snippets of dialog, which could be veering into formulaic as there is only a finite number of ways of expressing things. Also, if the piece has more a more elevated / literary tone, it can sometimes lead to a small percent of the text being flagged as human but with only medium confidence.

YMMV.

—

My autocorrect keeps changing percent to pervert - should I be worried?

 

[EmilyMiller]

You think that's bad, that Will Shakespeare guy was a sneaky AI user, according to the AI detectors.

BS

 

[SkyBubble]

BS

View attachment 2647532

View attachment 2647533

I've run his work through multiple detectors. It almost always comes back as AI.

 

[SkyBubble]

1. The free ones you can use on-line
2. Asking ChatGPT or another LLM to opine
3. Pangram
4. Commercial AI checkers

From convos, I believe Lit uses option 4. But I have no idea what exactly it is. Nor does anyone else. It’s certainly not Biff’s free AI check-o-matic, whose owner wants to sell you their de-AI-ification tool.

Rumors (and just rumors, nothing official) indicate they may be using Copyleaks and/or Originality.

 

[EmilyMiller]

I've run his work through multiple detectors. It almost always comes back as AI.

See above. You need to find a better AI detector.

 

[EmilyMiller]

I've run his work through multiple detectors. It almost always comes back as AI.

I wish we’d stop with the Shakespeare and Lincoln text being flagged as AI - by Mickey Mouse AI detectors maybe.

 

[ElectricBlue]

The fact that run of the mill AI checkers are not particularly good in general - I don't think this is a controversial statement - basically means that we are helpless against random false positives.

Not completely helpless. Don't write bland content that reads like AI junk is probably a better place to start, rather than agonising over the tools you do or don't use, or whatever detection tools the site might use (and how fallible those tools might be).

If a writer's content keeps getting rejected for suspected AI, and it's not because of artefacts left hidden in the text (which I suspect, is what most rejections are all about), then the only thing left is the text itself. Something is triggering the checkers, and if it's not technical gunk, what's left?

 

[RejectReality]

All well and good to tell newcomers go "git gud", but if they can never get published to get any feedback they're sure at a disadvantage in that department.

 

[ElectricBlue]

All well and good to tell newcomers go "git gud", but if they can never get published to get any feedback they're sure at a disadvantage in that department.

They are, but the fact remains - if it's something in the content it's something in the content, and that is under the writer's control. They might not like the message, but all that remains once all the technical guff is taken out of the equation are the words.

No-one in any of these threads has ever suggested any other reason for repeated AI rejections. It's always, "It's not me, it's the system that's busted." Explain to me then, how the other hundreds of other stories sail through.

 

[Louys90]

I’ve run segments of my own work through Pangram. Just out of curiosity, I don’t use either genAI or grammar checkers, just human alpha and beta readers.

It has never flagged my work as AI. What normally happens is I either get 100% human + high confidence. Or some large percent human + high confidence and some low percent human + medium confidence.

When I look at the sections which have only medium confidence, it’s normally snippets of dialog, which could be veering into formulaic as there is only a finite number of ways of expressing things. Also, if the piece has more a more elevated / literary tone, it can sometimes lead to a small percent of the text being flagged as human but with only medium confidence.

YMMV.

—

My autocorrect keeps changing percent to pervert - should I be worried?

I've run several of my texts through Pangram, and every time they came back as 100% human-written. But when I upload them to Literotica, they're rejected as AI-generated.
I know no AI detector is perfect, and false positives are inevitable. What I find hard to understand is how a text that one detector rates as 100% human-written can be rejected by another, whatever AI detector Literotica is using.

 

[XerXesXu]

Here's an interesting arxiv paper (subject to peer-review) which proposes narrative 'tells' as opposed to stylistic 'tells'. Primarily because getting AI to remove the stylistic tells from your 5000-word AI story reduces flagging as AI from 90% to 3%. It's much harder to remove narrative tells. Table 2 is instructive. The scattergram for Human is greater than for any of the AIs used and there is some overlap in the horizontal and vertical planes due to the wide dispersion of human narrative choices.

StoryScope

 

[EmilyMiller]

I've run several of my texts through Pangram, and every time they came back as 100% human-written. But when I upload them to Literotica, they're rejected as AI-generated.
I know no AI detector is perfect, and false positives are inevitable. What I find hard to understand is how a text that one detector rates as 100% human-written can be rejected by another, whatever AI detector Literotica is using.

They all work in different ways and come to different conclusions.

 

[SkyBubble]

Not completely helpless. Don't write bland content that reads like AI junk is probably a better place to start, rather than agonising over the tools you do or don't use, or whatever detection tools the site might use (and how fallible those tools might be).

If a writer's content keeps getting rejected for suspected AI, and it's not because of artefacts left hidden in the text (which I suspect, is what most rejections are all about), then the only thing left is the text itself. Something is triggering the checkers, and if it's not technical gunk, what's left?

AI is adapting, though. Some things that used to be considered signs of AI writing are now considered signs of human writing and vice versa. As AI trains on more and more material, it evolves.