Forum: Author Hangout

Thank you, Lazeez!

I get that, but it would make sense to give authors a wider selection of AI tags; just like you do for "sex": none; little; some etc. So, something like: AI Generated; AI Assisted; AI Polished; AI Edited; No AI Used.

if I find a cheap way to do the AI generation detection and make it automatic, I may implement an AI generation grading flag with multiple levels.

Your cheapest resource is your members - allow members (possibly premier- or paid-only to exclude bit-stuffers) to submit an AI score in the same way as the normal story scores. You may then want an additional account preference something like 'Treat non-AI scored as XX%' to take care of filtering newer stories, or preload the AI score with something derived from what you're already doing.

We'll see though, if I find a cheap way to do the AI generation detection and make it automatic, I may implement an AI generation grading flag with multiple levels.

Good luck. My prediction is that it won't get cheaper.

1) Current LLMs are on the verge of being able to consistently pass as human, when prompted carefully. I have four chapters of a story written that each score 100% human on GPTZero. I wasn't necessarily trying to breach any AI checkers, but rather improve the quality of writing that the LLM produces for me.

2) Many really good writers use style that we commonly associate with AI. The AIs were trained on their writing. It will become increasingly difficult to figure out which work is written by a human, and which by a human driving an AI.

Personally I've stopped caring. I'm focused on the plotlines and characters -- their journies, their emotional states, hot sex scenes etc. I hope people like the stories, but if people don't care to read them because they don't like the tools I use, that's ok too.

I hope we're going to get some really clever writers trying new things just to be different from the crowd. A modern analog to James Joyce, perhaps.

Exactly. And when digital photography revolutionized the work of darkroom artists like Ansel Adams, it brought a renewed emphasis on composition amongst other things.

It's past 3am so I won't go hunting for the link but last week on Twitter I saw a claim a study proven mathematically fair AI detection is basically impossible.

Or rather it's allegedly impossible to reduce false positives (100% human detected as whole or part AI) below 20% or so, or else, the actual AI detection rate rapidly collapse catastrophically, like, to reduce false positives to 10% you would have accept more than 60% false negatives (actual AI passing as human) or some such.

Interestingly, if remember it right, the false negative rate could actually be brought quite low, but of course at the expense of increased false positives, but not growing as rapidly.

So there's perhaps a sweet spot possible that actualy do detect over 95% AI as AI and "only" accuses one out of four humans as AI falsely... and of course, that's unusable the way institutions like education want to use AI detection, but what they (and us here) would want is objectively impossible as no false accusations mean no AI detection at all either.

All AI models reach a point where they are as functional as the best humans. Image recognition reached that point in the mid 2010's, voice just before 2020 and so on. It used to be that oncologists would spend hours poring over cell micrographs looking for cancerous cells, but around 2018 (if memory serves) AI systems became so proficient that they were producing much more accurate results in minutes. The role of the oncologist changed moving more heavily to a clinician role rather than diagnostic.

Generating output is a problem that has lagged recognition (mostly because researchers didn't start to seriously tackle it until a decade after recognition) but it's on the same curve. We should expect the recognition of AI generated output to become much more difficult as the quality of the generation tools grow. The only solution, IMO, is some kind of watermarking convention where the tools embed an undetectable (to humans) signature in the output. Something like stochastic watermarks in images, but applied to text.

The only solution, IMO, is some kind of watermarking convention where the tools embed an undetectable (to humans) signature in the output. Something like stochastic watermarks in images, but applied to text.

There is nothing of the sort that can be applied to text files. Sure, you can embed hidden tags in a Word or ODT document, but that's easy enough to eliminate by converting to plain text.

Embed the signature in the words themselves, perhaps. Word orders or specific word combinations?

Embed the signature in the words themselves, perhaps. Word orders or specific word combinations?

In a plain text file, anything like that would be easily edited out.

It could be done I think. Worth investigating, in any case. Here's a concept.

The approach would be to treat this as a multi-generator linguistic watermarking system. Each participating AI generator would have a unique identifier and watermark key, and would encode that identifier through ordinary, meaning-preserving language choices such as synonym selection, punctuation, contractions, or sentence structure. Because the number of generators is limited, the identifier itself can be compact and then protected with error-correction and a cryptographic authentication code.

The watermark would be distributed redundantly across many locations in the text rather than tied to fixed character positions. That makes it more likely to survive routine edits, reformatting, small insertions or deletions, and partial copying. A detector would test the text against the known watermark schemes for the participating generators, recover the encoded pattern, correct errors, and report the most likely generator together with a confidence score and estimated alteration rate.

This would not detect all AI-generated text. It would only identify text produced by generators that deliberately implement the watermark, and substantial paraphrasing, translation, or adversarial rewriting could remove it. The value is therefore in reliable attribution among a known set of cooperating generators, not universal AI-text detection.

Anyone who knows how the "watermarking" worked would be able to edit it out of plain text files with minimal effort.

And you couldn't keep how the watermarking works secret,
because AI detectors have to know how to look for it.

I envision many redundant signatures in a piece and a checking model to inhibit their removal. It wouldn't prevent someone determined to remove the signatures. What it would do is diminish the value of the flood of slop we currently see. The goal is to make detection more reliable than it currently is by reducing false negatives.

I envision many redundant signatures in a piece and a checking model to inhibit their removal. It wouldn't prevent someone determined to remove the signatures. What it would do is diminish the value of the flood of slop we currently see. The goal is to make detection more reliable than it currently is by reducing false negatives.

I understand what you are envisioning. It simply won't/can't work the way you envision with a plain text file.

A good challenge. I might spend some time on this.

I've posted a short story called Steganography Sample. I posted it as plain text rather than the HTML I would normally use.

It uses a technique called linguistic steganography to encode a watermark. The watermark algorithm made 29 modifications as part of its encoding process.

It uses 3 encoding channels a) a synonym library b) function word replacement ("that/which", etc) and c) adverb / modifier placement changes. It also uses a shared 256bit encryption key to prevent spoofing of the watermark. I've been able to successfully recover the watermark despite minor edits with a reasonable degree of confidence after the edits. The version I posted was not modified by me, and has two poor substitutions, which you may notice. However, after correcting them, the watermark was still recoverable with the same level of confidence as previous, proving that that the outputs can be lightly manipulated without impact to the watermark. It won't protect against an adversary determined to defeat it, but it will identify AI generated text for most common use cases.

To make this work broadly vs the prototype I built you would have to implement a public / private key system so that detection could be open and broad but only a smaller group of encoders could write watermarks. You'd also have to build much more robust substitution dictionaries and semantic models to ensure meaning wasn't changed when the subs occurred. But I think the idea is not just plausible but also implementable.

FYI, I didn't know about linguistic steganography when I started playing with this, but it turns out that there had been substantial work done in this area academically AND there is current work being done in using LLMs to implement.

The issue was never could you put the "watermark" in a plain text file.
The issue is how do you prevent anyone with access to the file as plain text from editing it out?
And the answer is you can't.

Please - download the story, identify the watermarks and edit them out. I think you might be able to degrade it, but unless you do a wholesale rewrite I am skeptical you can remove it entirely.

Now, one scenario where they could be removed would be to ask an LLM to rewrite it. But if you did that, and the LLM was a subscriber to the watermarking convention, then it would simply write a new watermark.

I believe what you're doing is different from bulk text generation, which is what this potentially might impact.

And let's face it, while there are millions of models in hugging face, it's a research hub. Text generation is a subset of what's there, and most of those models are not going to be useful to the ordinary person doing bulk text generation unless they're willing to set up a sophisticated local environment (which it sounds like you have).

This is targeted at commercially available tool sets.

That argument says there is no need to flag. Good stories will get recognized as good stories. Right?

I've said before that I have AI generated stories with high-ish scores. Clearly some people enjoyed them, however they were written

That argument says there is no need to flag.

That argument just shows that the people using AI have still not understood where the problem with AI-writing lies.

The need to flag is for people who don't want the slop to appear in their feeds. Without the tag, they can't apply any exclusion rules.

And I don't doubt you have AI-generated stories with high scores. ...even though I have still not seen an AI story on this site with a high score AND a vote-count that should reflect an "awesome" tale. I mean, for Christ's sake, just take Megumi Kashuahara as an example. She just won the Clitorites for "Best new Author of 2025" (even though she's been publishing here since 2022 somehow), has ~500 followers... and most of her stories STILL only have between 20-50 votes. Even her own followers can't be bothered to leave a vote on all the AI-Crap she posted, and instead just close the site to look for something else, apparently.
But there are also people liking the ever-same AI-generated tales on YouTube disguised as reddit stories. Not everyone has a problem with AI-slop all sounding the same, and doing nothing but rehashing the ever-same plots with different character names.

But MOST people do.

*EDIT*

Smashwords, D2D, KDP,... EVERY publisher rejects AI-written stories and put measures to detect them into place. Why? Because ever since the flood of AI-slop books on Amazon made the news, book sales across the board went down by 12.7% to 21%. Because people don't want to pay money for the crap! And if they already paid money for a book, just to find out it is dogshit, they feel cheated and will look for a different store (if they try to buy any more at all).

Are you sure it's because they are discriminating enough to not want to "pay for that crap"? I would argue that what you see happening in the music world is a better analogy. Something like 70% of new submissions to Spotify were AI generated, which depressed royalties for all artists dramatically. Isn't it more likely that the volume of generated content is drowning out the hand written authors works? After all, very few written works rise to the level of truly memorable. Most are little more than 1960's era pulp fiction. And people, by and large, are content to read it. All art forms experience this. Why aren't there more art-house films? Hollywood wants moneymakers. Ingmar Bergman's films don't bring in the kind of money that the latest Marvel blockbuster can.

When craft becomes a business, mass production homogenizes the product.

If you're going to generate material with AI you have to pay extra attention to all the non-mechanical issues you just described because plot continuity, dialog continuity, florid prose, and transitioning errors can all occur in generated text. Paragraphing can also be an issue, as can predictable rhythms. But punctuation and spelling it excels at…

Are you sure it's because they are discriminating enough to not want to "pay for that crap"?

Yes. Because it's not an opinion. It's what customers have told online platforms about why they stopped paying.

Spotify is being overrun with AI generated music. Like you said. The part you didn't mention is how the subscribers are furious about it, and so many canceled their subscriptions that Spotify had to mark AI-Musik as such, so people can avoid it.
Same for YouTube Music, by the way, if you want to remain in the Music industry.
Or you could check out Steam, to see how AI in games is faring, where over half the users said they won't buy AI-influenced games, and the rest simply trusted the refund-policy if the games are shit.
And I don't think I need to remind anyone here about the 50% drop in sales on KDP last year, right?

We have customers avoiding AI-Products in the music industry, the games industry, and the book-publishing business.

So...yeah. The "authors" who use AI can bitch and moan as much as they want. It doesn't change the fact that more and more customers are avoiding any product that may is associated with AI, because every piece of AI-generated art they come into contact with is the same rehashed, repetitive, dull, buggy, slop.

I think you proved my point. Spotify, despite a small group of vocal complaints relative to their overall subscriber base, hasn't seen a significant drop in subscriptions. Their response has been to deny royalties to AI generated content - what that means is that they are keeping the revenues for themselves. KDP has responded by enforcing labeling but not banning. More content means more gross sales for themselves, even if individual author revenues have suffered. Notably, neither have outright banned the content. They've responded in a way that allows them to say they are doing something without having to forgo the revenue they must be earning from these products. These are businesses that depend on small margins and large volumes. They are not traditional publishers that incur the expenses that traditional publishing houses incur. More material to sell means more profits. And most importantly (for human authors) they are decidedly NOT taking the author's side. The only business that I'm aware of that has taken a more decisive position is Tidal, which recently announced a ban on all AI generated content.

Spotify, despite a small group of vocal complaints relative to their overall subscriber base, hasn't seen a significant drop in subscriptions.

...they lost 25% of their user base in the quarter after the flood started, until they implemented the option to filter AI-crap out. Roughly 150 million users. The reason they didn't lose any Premium members is because they are paid for the whole month, so Spotify had time to implement their measures before shit hit the fan completely.The ones who had paid upfront naturally stayed. The ones who hadn't... well, they're gone!

KDP has responded by enforcing labeling but not banning.

Yes. As I wrote in the post you're responding to, there ARE people enjoying AI-Slop. And there's no reason to take that away from them. But all the people who don't enjoy it will appreciate the tag, so they can filter it out of their feeds and recommendations.

Kinda like here on SOL.

The presentation you provided showed MAU increased by a 10 million, not a loss of 150m. The loss you described would have been the biggest in their history, and catastrophic. Instead they describe healthy growth in both users and revenue.

Where are you getting this data?

Where are you getting this data?

From the sheets Spotify published and I linked in my post. Yes, the amount of registered users rose by ten percent. But the users didn't actually use the platform and just went away (without deleting their accounts), leading to a 25% drop.

That presentation is an SEC investor deck. There no place in it that says what you have said, or even implies it. For starters, they do not discuss churn of any level, never mind 25%. Now Spotify's definition of MAU is very forgiving (if you login in once in a month and play more than 1 millisecond of anything you are counted as active), but the presentation contains no metrics that could be used to support your statement. If they had published Q/Q of any of these, that would be useful, but they didn't.

* total listening hours,
* average listening time per user,
* daily active users (DAU),
* DAU/MAU ("stickiness"),
* retention,
* skips per session,
* search vs. recommendation usage,
* discovery rates,
* premium churn.

What you provided as evidence was an upbeat investor report.

I'm tempted to believe that you're using an AI to find your proof points and it's hallucinating… but I know how much you hate the tech.

You know, maybe because it's only doing spelling and grammar checking, continuity checking, and building a story bible, not actually writing prose.

There's a HUGE difference.

Dude, for crying out loud... can you maybe stop preaching how awesome your AI tools are for long enough to actually read my post? What I said was that, if you create a good story, nobody gives a shit how you did it, because there's a difference between AI-Slop and stories.

Or, maybe, think about why you have such a burning need to defend yourself for using AI all the time?
For your sake, I hope it's just your attempt to justify the price tag for a computer that runs eight local models, and the $300/month Grok subscription, all to "help spot plotholes".

Your responses indicate you're freaked out about AI and appear to miss no opportunity to denigrate it, which is strong evidence of someone who is intimidated by it and thus afraid of it.

I give up.
I wrote twice now that nobody would give a shit about AI-assistance, as long as you don't let it create slop. As long as your story is good, nobody will avoid it just because you used AI for some part of it.
And your takeaway is that I'm "freaked out about AI" and am "denigrating it"? Yeah... maybe use Grok to interpret my posts. You seem to be troubled with reading comprehension.

Are you addressing that comment to me?

I said that I accepted the forced AI tag on most of my stories because a significant part of the "scaffold" text was AI generated. I was complaining about the forced AI tag on my latest story because ALL of it was indeed my writing (without the quotation marks!). I did use AI for a final "polish" but all that did was some surface-level tweaks on maybe 10% of the text.

Have you read the story in question? Does it read like "AI generated slop" to you?

Are you addressing that comment to me?

No, the board software automatically makes it an answer to the topic starter, if you hit the "Reply to topic" button on the bottom of the thread.

the board software automatically makes it an answer to the topic starter, if you hit the "Reply to topic" button on the bottom of the thread.

...which is why you should always use the bent arrow 'reply to post' for post-specific replies