From pending purgatory to AI hell

[CleverGenericName]

I don’t usually post on the forum (I never thought I’d be writing to you, but…). I had a story stuck in pending purgatory for months, and it was just officially rejected for suspected use of AI. I have found other people’s posts on this very frustrating issue helpful, so I thought I would share my experience and what I learned (or suspect) from it, to pay it forward.

Is there a “Fast Track” for successful authors?

I am not sure what the threshold is for successful authors, but I have just under 1,500 followers and over 750,000 story views on 20 stories, most of which are in the 4.8s, and I still had a story sent to pending Purgatory. So, if there is a “Fast Track,” it is only for the most rarified of authors.

Is there a “Pending” bug?

I would guess not. I would hypothesize that the pending bug is actually stories being flagged as AI-generated and moved to a much longer queue for human review. I think the reason that multiple resubmissions sometimes work to get out of pending is that AI-detectors are powered by LLMs (i.e., they are themselves a form of AI) and are notoriously fickle and inconsistent. So, each resubmission creates a new opportunity for the AI-filter to say, “No AI here, move along.”

I haven’t tested this theory myself, but I suspect that if you just deleted and resubmitted a story every time it was pending for more than 3 days, it would eventually get lucky with the AI filter and pass through (assuming the issue was a false-positive flag for AI use).

How did I finally get out of “pending” purgatory?

The second time I deleted the story after it had been pending for more than a month, I realized it was holiday-themed, so I resubmitted it as part of the winter holiday contest. That seemed to accelerate the rejection process.

Did I use AI to help write my story?

As someone whose neuro-spiciness includes dyslexia and a starter pack of other related challenges, I use Grammarly to catch grammatical and spelling issues that I am literally unable to perceive otherwise. This hasn't been an issue with 20 of my 21 submitted stories to date. My wife also reads through my stories before publication and usually catches additional issues, and I catch more when I use the “Read Aloud” function in Word.

Do you know what you did that tripped the Literotica AI Detector?

No. I ran my story through five different AI testing sites (including one paid site), and they all came back with a 95%+ chance of the story being entirely human-created. After the second stint in pending purgatory, I went back to an early draft of the story and completely rewrote it, and that number rose to a 99%+ chance that the story was entirely human-created.

I ran the story through a couple of general-purpose LLMs to see what they thought, and the consensus seemed to be that the most “AI”-like elements of my story were: lack of spelling mistakes, and grammatical consistency over a very long story (28K+ words). So maybe long story + clean edits = AI flagged by Literotica?

I’m honestly kind of stumped as to what else I can do to prove that my story is not AI-assisted. It is impossible to prove a negative (actually, that is not strictly true from a mathematical perspective; it’s more like “non-existence claims in open systems cannot be conclusively proven”), and without actually knowing which AI filter is being applied to my stories, it’s tough even to mount a defence.

I could try rewriting it again, but that is a depressing prospect, without any clear indication of what the perceived issue is in the first place. Or I could find another site to publish it on, I guess, but I feel a strong sense of loyalty to the folks who have taken the time to write and comment on my works on this site, and asked me to keep sharing my work here.

What are you going to do now?

I honestly don’t know. I mainly write for my own enjoyment and to manage my anxiety on days when my wife’s chronic illnesses are particularly bad, and this experience has been highly demotivating. At the same time, I’ve gotten a number of really kind and encouraging comments on my stories from people who seem to appreciate my efforts, which makes me want to try and push through my frustration.

I had two stories published for the Winter Holiday contest, both of which I spent significantly less time polishing and editing than the one that was stuck in purgatory. Maybe in the future I will just post stories when they are “good enough” and not worry so much about polishing and editing them. Or I might switch my focus to something entirely different, like writing a book or screenplay.

Anyway, I hope this is helpful for some of you facing a similar challenge.

 

[AwkwardMD]

Grammarly continues to be an all-too-common thread in these rejections.

I'm sorry this happened to you. Do you have a version of the document from before you let Grammarly at it, or were you following its suggestions the whole time?

 

[CleverGenericName]

I try to use Grammarly as sparingly as possible (I ignore its suggestions on wording/phrasing), basically just as a mechanical check on dyslexia and related issues (homophones are the bane of my existence), so unfortunately, I don't have pre-/post- Grammarly versions. But I would definitely buy your observation/hypothesis that there is something about Grammarly's grammatical recommendations that is getting flagged by whichever AI filter is being applied to submissions.

 

[AwkwardMD]

I try to use Grammarly as sparingly as possible (I ignore its suggestions on wording/phrasing), basically just as a mechanical check on dyslexia and related issues (homophones are the bane of my existence), so unfortunately, I don't have pre-/post- Grammarly versions. But I would definitely buy your observation/hypothesis that there is something about Grammarly's grammatical recommendations that is getting flagged by whichever AI filter is being applied to submissions.

The more time has gone on, the higher the likelihood that anyone coming here to talk about a rejection also used Grammarly. Used to be 1 in 3, now it's more like 7 or 8 out of 10. It makes me think Grammarly is doing more under the hood than it should. The commonality is too great to be ignored, and wildly disappointing for a product that used to be a net good with a positive impact on allowing everyone to express themselves with confidence.

For the record, though, Lit's AI detector is not powered by an LLM.

 

[CleverGenericName]

If you don't mind me asking (and without burning any secret Lit sources), how certain are you that Lit's AI detector is not powered by an LLM? Because if that's the case (and it is certainly consistent with what I have seen of its output), a known weakness of many of those tools is that the more polished and statistically consistent the prose in a story, the more likely it is to be flagged as AI-generated. So, ironically, because I really love(d) the story that was rejected and spent way longer polishing and editing it than I have for most of my other work, it passed the statistical threshold into AI-ness. I can't wait to tell my wife that her editing skills are statistically indistinguishable from ChatGPT

I guess if I want to publish this story on Lit, I need to go in and de-edit it... what a pain. That being said, I really appreciate your thoughtful feedback. You are a scholar and a gentleperson.

 

[PennyThompson]

The more time has gone on, the higher the likelihood that anyone coming here to talk about a rejection also used Grammarly. Used to be 1 in 3, now it's more like 7 or 8 out of 10. It makes me think Grammarly is doing more under the hood than it should. The commonality is too great to be ignored, and wildly disappointing for a product that used to be a net good with a positive impact on allowing everyone to express themselves with confidence.

For the record, though, Lit's AI detector is not powered by an LLM.

Grammarly widely introduced new LLM features just about one year ago, and they've been slowly turning up the dial on how pushy the AI features are.

All of the people saying they've used grammarly for years without a problem are absolutely correct... Because it used to be an advanced spell checker, nothing more.

Now it recommends changes for things like tense and tone and vocabulary, which really means it's rewriting your sentences for you.

So the people saying "I haven't changed my workflow," aren't lying, it's just that the tools they rely on are now mechanically and structurally the same as ChatGPT, and are "helpfully" interacting with their writing process far more than most grammarly users realize.

 

[AwkwardMD]

If you don't mind me asking (and without burning any secret Lit sources), how certain are you that Lit's AI detector is not powered by an LLM?

I am about as certain as a person can be.

 

[AwkwardMD]

So, ironically, because I really love(d) the story that was rejected and spent way longer polishing and editing it than I have for most of my other work, it passed the statistical threshold into AI-ness. I can't wait to tell my wife that her editing skills are statistically indistinguishable from ChatGPT .

This is not *quite* accurate, but it is close enough that you and your wife are allowed two extra feathers in your cap on a temporary basis. If you want to wear them for a period of longer than three months, you'll have to submit a form.

 

[AG31]

I don’t usually post on the forum (I never thought I’d be writing to you, but…). I had a story stuck in pending purgatory for months, and it was just officially rejected for suspected use of AI. I have found other people’s posts on this very frustrating issue helpful, so I thought I would share my experience and what I learned (or suspect) from it, to pay it forward.

Is there a “Fast Track” for successful authors?

I am not sure what the threshold is for successful authors, but I have just under 1,500 followers and over 750,000 story views on 20 stories, most of which are in the 4.8s, and I still had a story sent to pending Purgatory. So, if there is a “Fast Track,” it is only for the most rarified of authors.

Is there a “Pending” bug?

I would guess not. I would hypothesize that the pending bug is actually stories being flagged as AI-generated and moved to a much longer queue for human review. I think the reason that multiple resubmissions sometimes work to get out of pending is that AI-detectors are powered by LLMs (i.e., they are themselves a form of AI) and are notoriously fickle and inconsistent. So, each resubmission creates a new opportunity for the AI-filter to say, “No AI here, move along.”

I haven’t tested this theory myself, but I suspect that if you just deleted and resubmitted a story every time it was pending for more than 3 days, it would eventually get lucky with the AI filter and pass through (assuming the issue was a false-positive flag for AI use).

How did I finally get out of “pending” purgatory?

The second time I deleted the story after it had been pending for more than a month, I realized it was holiday-themed, so I resubmitted it as part of the winter holiday contest. That seemed to accelerate the rejection process.

Did I use AI to help write my story?

As someone whose neuro-spiciness includes dyslexia and a starter pack of other related challenges, I use Grammarly to catch grammatical and spelling issues that I am literally unable to perceive otherwise. This hasn't been an issue with 20 of my 21 submitted stories to date. My wife also reads through my stories before publication and usually catches additional issues, and I catch more when I use the “Read Aloud” function in Word.

Do you know what you did that tripped the Literotica AI Detector?

No. I ran my story through five different AI testing sites (including one paid site), and they all came back with a 95%+ chance of the story being entirely human-created. After the second stint in pending purgatory, I went back to an early draft of the story and completely rewrote it, and that number rose to a 99%+ chance that the story was entirely human-created.

I ran the story through a couple of general-purpose LLMs to see what they thought, and the consensus seemed to be that the most “AI”-like elements of my story were: lack of spelling mistakes, and grammatical consistency over a very long story (28K+ words). So maybe long story + clean edits = AI flagged by Literotica?

I’m honestly kind of stumped as to what else I can do to prove that my story is not AI-assisted. It is impossible to prove a negative (actually, that is not strictly true from a mathematical perspective; it’s more like “non-existence claims in open systems cannot be conclusively proven”), and without actually knowing which AI filter is being applied to my stories, it’s tough even to mount a defence.

I could try rewriting it again, but that is a depressing prospect, without any clear indication of what the perceived issue is in the first place. Or I could find another site to publish it on, I guess, but I feel a strong sense of loyalty to the folks who have taken the time to write and comment on my works on this site, and asked me to keep sharing my work here.

What are you going to do now?

I honestly don’t know. I mainly write for my own enjoyment and to manage my anxiety on days when my wife’s chronic illnesses are particularly bad, and this experience has been highly demotivating. At the same time, I’ve gotten a number of really kind and encouraging comments on my stories from people who seem to appreciate my efforts, which makes me want to try and push through my frustration.

I had two stories published for the Winter Holiday contest, both of which I spent significantly less time polishing and editing than the one that was stuck in purgatory. Maybe in the future I will just post stories when they are “good enough” and not worry so much about polishing and editing them. Or I might switch my focus to something entirely different, like writing a book or screenplay.

Anyway, I hope this is helpful for some of you facing a similar challenge.

Thanks for taking the time to lay all this out.

 

[millamomud]

I write my stories in word, do not let anyone or anything touch my documents and still get flagged. I tried commenting and providing a word doc outlining the writing progress, averaging around 400 words per writing session on a span of about six weeks in which the story was written, to no avail.

I'm basically about to give up publishing on here. It's apparent that lits detection does not work properly for German texts. Fortunately we Germans do have some alternatives.

 

[Britva415]

Is there a “Pending” bug?

I would guess not. I would hypothesize that the pending bug is actually stories being flagged as AI-generated and moved to a much longer queue for human review

This doesn't explain why the delete-and-resubmit workaround would work. Wouldn't that just put the story at the back of the extra-long "queue" again?

I think the reason your story finally got handled after months without employing the workaround is: "Permanently"-pending stories sometimes do eventually get discovered, probably through manual searches for them.

Maybe you benefited from the increasing volume of other people's resubmits with notes about the bug, and @Literotica responded to awareness of the situation with a one-time search for stuck stories.

The site's advice to authors is to not touch a pending story for 15 days. They don't come out and say what you should do after that time, but they're definitely aware that people are using this workaround and they're wordlessly supporting it by approving resubmitted stories.

That's why I don't believe there's any queue which they're deliberately allowing to take longer than 15 days.

 

[IWroteThis]

I’m honestly kind of stumped as to what else I can do to prove that my story is not AI-assisted. It is impossible to prove a negative (actually, that is not strictly true from a mathematical perspective; it’s more like “non-existence claims in open systems cannot be conclusively proven”), and without actually knowing which AI filter is being applied to my stories, it’s tough even to mount a defence.

I use writing software (yWriter) that backs up individual scenes as I'm writing. With those showing the process, I can prove provenance, which is good enough to settle copyright or AI claims.

This doesn't explain why the delete-and-resubmit workaround would work. Wouldn't that just put the story at the back of the extra-long "queue" again?

The most likely explanation for that is the inconsistency of the AI detection tools. No matter how many times it gets flagged and dumped back at the end of the line, you only need it to pass once to skip that queue and get published.

 

[AwkwardMD]

The most likely explanation for that is the inconsistency of the AI detection tools. No matter how many times it gets flagged and dumped back at the end of the line, you only need it to pass once to skip that queue and get published.

This logic doesn't hold up. If it was inconsistent, AH members would be seeing rejections en masse just like we experience other bugs. The authors here, who represent a wide spread across demographics, English-first proficiency, writing platforms, kinks, and categories, are largely unaffected by this.

Lit's AI detector is not inconsistent. The pending bug is an unrelated problem that other people conflate with AI rejections incorrectly.

EDIT: It’s not because we're special. Laurel doesn't particularly care about us.

 

[MillieDynamite]

I've had one AI rejection the entire time I've been here. It was the last story I tried to publish, and it had been up for sale for a year. I wrote it last year, don't remember the month, it was edited by my publisher, not using Grammarly, and I never take Grammarly suggestions. For some reason, I don't know, it was rejected for possible AI use. I mean, I don't know what part of the writing caused the text to be tagged. I haven't had time to go through it and look for what might be considered AI in the text. I'll try again at some point when I have time to go through it. But for the record, I don't use AI, don't accept Grammarly's changes, and don't let it rearrange anything the way it suggests.

 

[MillieDynamite]

And while an automated AI detector might not use LLM, it's still an AI-based system. I don't believe that Laurel or Manu can read all the stories to determine whether AI is used; they must be using an automated text reader, and any automated text reader, even ones looking for banned content, uses AI to do the work.

 

[Britva415]

while an automated AI detector might not use LLM, it's still an AI-based system

That might be a bit like saying that an iron grille over a propane burner is a food based system. It's not "food based" even though food is what you use it for.

The Lit AI detector might not necessarily use AI to do its job. I could be wrong but I feel like @AwkwardMD has opined that it doesn't.

 

[IWroteThis]

This logic doesn't hold up. If it was inconsistent, AH members would be seeing rejections en masse just like we experience other bugs. The authors here, who represent a wide spread across demographics, English-first proficiency, writing platforms, kinks, and categories, are largely unaffected by this.

Lit's AI detector is not inconsistent. The pending bug is an unrelated problem that other people conflate with AI rejections incorrectly.

EDIT: It’s not because we're special. Laurel doesn't particularly care about us.

Yes, the logic does hold up.

So, to generalize, every submission is given a percentage likelihood of being AI generated. It doesn't have to be 100% likelihood to be flagged. How the cutoff is determined may vary, but they all have one. Anything that scores below that line is not flagged as AI. Anything that reaches or exceeds that score is flagged as AI. Simple, right?

Well, not really.

To make up numbers, let's say that the cutoff is 75%. To fudge a bit, let's pretend we can know the true score (ts) of a story. So, in a perfect world, any story where ts >= 75% is flagged as AI while any story where ts < 75% is not. Now, there's two issues that come into play. First, AI detection tools are not good enough to always calculate the ts, so let's call it the calculated score (cs). Second, each time the story is reviewed, it can get a different cs. If you run a big enough set of stories through it enough times, you can determine the range of precision (rp).

Now, the fun part…

If ts+rp<75%, that story will never be flagged as AI. If ts-rp>=75%, that story will always be flagged as AI. However, where ts is within 75%±rp, that story may or may not be flagged as AI, depending on where within that range of precision it's scored.

So, to apply the logic to Literotica, a lot of stories are never going to be flagged as AI. This is especially true for all of the unedited first drafts that are submitted and seem to be approved in bulk on a daily basis. You're also going to have a pool of stories that will always be flagged, whether because they are AI generated or simply check too many boxes. Then, you have those that could go either way. When they score low, they get published. When they score high, they go in the queue for review. Those stories, if resubmitted, may pull a low score and get published, or they may pull a high score and get stuck in Pending again. The lower the ts, the more likely for it to work. The higher the ts, the more times it will probably take to catch an outlier on the low side, but they still can.

So, once again, the logic fits the current situation quite well.

 

[AwkwardMD]

And while an automated AI detector might not use LLM, it's still an AI-based system. I don't believe that Laurel or Manu can read all the stories to determine whether AI is used; they must be using an automated text reader, and any automated text reader, even ones looking for banned content, uses AI to do the work.

Optical character recognition technology for computers dates back to 1974. AI is not required for this task.

 

[MillieDynamite]

Changing an image (OCR) of characters into digital characters isn't the same as comprehending what's being read. Understanding a word, or even identifying banded words, requires more than ORC tech. And ORC doesn't scan digital files; it scans printed sheets of paper.

Optical character recognition technology for computers dates back to 1974. AI is not required for this task.

 

[MillieDynamite]

ORC can recognize your face to unlock a door or your phone, but it can't tell if it's looking at a picture or you.

 

[AwkwardMD]

ORC can recognize your face to unlock a door or your phone, but it can't tell if it's looking at a picture or you.

I'm not arguing that OCR is how it works. I'm telling you that scanning a story is trivial even for a 486.

 

[MillieDynamite]

But making a judgment on AI influence isn't a simple matter, no matter how trivial the matter is to scan the file. Determining what words mean and how they are used for either forbidden content or AI useage requrires at least a rudimentary AI-like technology. There were AI programs back in the mid-90s of the last century.

I'm not arguing that OCR is how it works. I'm telling you that scanning a story is trivial even for a 486.

 

[Britva415]

each time the story is reviewed, it can get a different cs

This is the part which seems suspect.

Partly because we don't know Lit's automated AI detection uses AI. If it doesn't, then there is every reason to believe that the outcome is deterministic and repeatable.

And partly because it's also not clear whether it would be non-deterministic (unpredictable) even if AI were used to detect suspected AI. An AI which can't produce repeatable results with the same input seems like a pretty poor AI.

 

[MillieDynamite]

Do a simple search for "A non AI software tool that detects AI writing" and this is what you get.

Followed by many AI software packages to detect AI.

 

[AwkwardMD]

But making a judgment on AI influence isn't a simple matter, no matter how trivial the matter is to scan the file. Determining what words mean and how they are used for either forbidden content or AI useage requrires at least a rudimentary AI-like technology. There were AI programs back in the mid-90s of the last century.

I will not discuss how Lit's AI detector works. Your only choices are to trust me when I say that I side with Lit, or not. Either is fine, but I push back on misinformation.