AI detection and the Evolution of a Paragraph—and Why AI Gets It Wrong

IsenRae

Author
Joined
Oct 10, 2025
Posts
68
I did some comparisons of my hand-written excerpt to see how, during my editing process it registered on the "THATS AI GRAPHS"

What I found most interesting is that when I allowed AI to polish my original draft, the AI detector actually gave it a lower AI score than my fully human-written revision.

My human edit scored 62.8% AI, while the AI-polished version scored 49.2% AI.

In other words, the detector judged the AI-generated revision as more human than the human-written one.

Also i'll present to you my editing process. How my stories start and how they finish. My stories detect as AI back to Human and back to AI.
I've tied this mostly to dialogue during this project. Apparently ai thinks dialogue tags and beats around them are mostly AI.

You'll see where I added a line with a single dialogue tag and my AI detection JUMP drastically in my third revision.





Here's the quick revision. Here I am only checking to see if all the feelings and thoughts have been captured.



Here is my deep dive for first person perspective. This part of my process is for emersion, compression, and rhythm.

I trim the fat.
Try to take the same meaning from large sentences and compress them to smaller sentences.
I stop telling, add ambiguity, subtext, targeting emersion.
Rhythm is always on my mind.

and I pick a tone and try to hold it consistent



It was a long week.

Missed classes. Showers. Opportunities.

No concern.

The dishes in the sink had grown their own quiet civilization.

I felt weak. Powerless. Empty.

That day I told myself I just needed rest. A single night to reset.

By morning I couldn’t summon the strength to move.

The world kept moving without me. Sunlight. Traffic. Laughter that bled through my door like a cruel joke.

Nothing. Not at first.
Not sadness. Not anger.
Just the hollow ache.

Abrupt. Like a flower pulled out by its roots.

I tried to read.

To care.

But everything tasted like ash. Even the books—poetry, my old refuge—looked back at me.

Words on paper.
Words that refused meaning.
Numbness faded around day three. Maybe four. Maybe five.

What replaced it was grief.

I just didn’t know what for.

That's when it happened.

Tears.

Tears because I didn’t want friendship. I wanted love. Somewhere inside I had always known.

The kind of tears where you cover your face.

Truth—
someone else names.

A kiss.
What lovers do.

Not friends.

I convinced myself of a lie before I could speak it.

And yet I convinced myself it was something else. I just didn’t know what. Not desire. Not then.

Finally—I showered.

I let the water burn me.
Too hot. Too much steam.

Proof.

I could still feel.

I dressed, careless if it matched.

My suitcase lay open on the bed. Half-filled.

Evidence.

I quit

I would tell my mother a lie. Tell her I failed at school.

Here is my more final edit.
I've read through my revision and compression several times now.
I let lines sit and ferment.
I try to experience it as the protagonist and make sure i miss nothing.

immersive, sensory, impact.


It was a long week.

I lay on my bed, exhausted.

Missed classes. Showers. Opportunities.

No concern. Only blank spaces and avoided memories.

The dishes in the sink had grown their own quiet civilization.
The smell thick and cloying.

I felt weak. Powerless. Empty.

My mind flicked from thought to thought.
To that day.

No.
I told myself I just needed rest.
A single night.
A full reset.

But by morning I couldn’t summon the strength to lift a finger.

The world kept moving without me. Sunlight. Traffic. Laughter that bled through my door like a cruel joke.

Nothing. Not at first.
Not sadness. Not anger.
Just the hollow ache.

Abrupt. Like a flower pulled out by its roots.

I needed distractions. Something to keep my mind off—

her.

I tried to read.
I tried to care.

But the words tasted like ash. Even the books—my poetry, my old refuge—looked back at me.

Words on paper. Words that refused meaning.

Then Silence.
The numbness faded.
In its place—grief.
I didn’t know what for.

That's when it happened.

I broke.
I cried.

Because I didn’t want friendship.
I wanted love—somewhere inside, I had always known.
Tears. The kind where you cover your face.
Ugly.
Vulnerable.

I convinced myself of a lie before I could speak it.

And yet I convinced myself it was something else.
I just didn’t know what.
Not desire.
Not then.

But truth—
someone else had to name it.
A kiss.
What lovers do.

Not friends.

Pressure released.

I trembled.
Weak.

I showered.

I let the water burn me.
Too hot. Too much steam.

Proof I could still feel.

Alive.
But ruined.

I dressed, careless if it matched.

My suitcase lay open on the bed. Half-filled.
I didn’t remember putting it there.

Evidence of my failures—of my fear.

I quit.

Out loud.

“I quit.”

I would tell my mother a lie. Tell her I failed at school.

I would—
run away.


Now that you've seen the stages of each process. Here is a brief breakdown.


Draft 0/1
– Traditional writing. Free flow, basic narrative structure. Telling.

Edit 1 – First reread. Did I capture the feeling I wanted? Tone? Personality? etc.

Edit 2 – Stop telling. Move toward immersion and compression.

Edit 3 – Impact. Flow. Rhythm. Breathwork.


What does an AI detector say?


Draft 0/1

1773428216333.png

Edit 1
1773428309842.png


Edit 2

1773428401087.png
Edit 3
1773428484643.png

Draft 0/1 -- 36.9% AI
Edit 1 -- 62.8% AI
Edit 2 -- 100% Human
Edit 3 -- 4.9% AI (If you add dialogue this jumps much higher.

For example here is the same halting paragraph structure with dialogue 26.4% AI
1773428800395.png

For examples sake I also let AI polish my first draft for me.
It also scored lower that the very much human written text that polished it.
49% ai





A week crept by.

I hadn’t gone to class. I hadn’t showered. I hadn’t cared. The dishes in the sink had grown their own quiet civilization, and I moved through the apartment like something hollowed out. Weak. Powerless.

When I first came home that day, I told myself I just needed to rest. One night to reset. But by morning I couldn’t summon the strength to move. The world outside my apartment kept turning without me—sunlight, traffic, laughter bleeding through the walls like a cruel joke.

At first there was nothing. Not sadness. Not anger. Only the hollow ache of something pulled out by the roots.

I tried to read. I tried to care. But everything tasted like ash. Even the books—poetry, my old refuge—looked back at me from the page, their words refusing to mean anything at all.

Somewhere between the third and fourth day, I realized I wasn’t numb anymore. I was grieving. I simply didn’t know for what.

That was when the tears finally came.

It hadn’t been friendship I wanted. It had been love. Some part of me must have always known. How could I not? How could I have been foolish enough to need her to name it for me?

We had kissed—something lovers do, not friends. Yet I had convinced myself it was something else, something undefined and safe.

Not desire.

Not then.

On the sixth day, I showered. The water was too hot and the steam too thick, but I let it burn anyway. It was proof that I could still feel something.

I dressed without caring what matched. My suitcase sat open on the bed, half-filled with the quiet evidence of quitting. I told myself I would go back to my mother.

Admit I had failed at school.

1773428917177.png


I'll let you draw your own conclusions.

The floor is open.


(small note: This methodology of writing is called compression. It is a ruthless tool and you may recognize the results in other literary works, poems and minimalist prose. I'd be happy to talk about some of these literary tools on a side bar. And yes you should always establish everything on paper first. Then compression. It can be done without that. But you may be leaving opportunities behind.)
 
I’ve read your work as you know, and enjoyed it. It strikes me that these detector sites now offer functions to ‘humanize’ your text, and also paid versions to do even better detection on even more text.

This seems like an incentive to flag input text as needing attention (and their paid services). Maybe I am too jaundiced, IDK 🤷‍♀️
 
I have no idea why we're so gleefully training AI to replace us, but it seems some of us just cannot stop ourselves. It saddens me.

Interesting analysis, OP, but I'm not sure it tells us much.
 
It only proves how flawed current AI actually is and no one should take a rejection/censor due to literotica's AI screening process personally. That's letting flawed code and algorithms combined with an 27 plus year old database/cms control your emotions.
 
For curiosity, I've run old stories of mine from 2006-09 (before AI writing was a thing) through AI detectors and gotten scores like 37% (meaning if I submitted it today, it would likely be rejected.)

The third entry in the OP's post looks very much like AI.
 
For curiosity, I've run old stories of mine from 2006-09 (before AI writing was a thing) through AI detectors and gotten scores like 37% (meaning if I submitted it today, it would likely be rejected.)

The third entry in the OP's post looks very much like AI.
We are assuming that these free detectors have any merit. Even the paid for ones are very unreliable.
 
It only proves how flawed current AI actually is and no one should take a rejection/censor due to literotica's AI screening process personally. That's letting flawed code and algorithms combined with an 27 plus year old database/cms control your emotions.
Absolutely.
 
For curiosity, I've run old stories of mine from 2006-09 (before AI writing was a thing) through AI detectors and gotten scores like 37% (meaning if I submitted it today, it would likely be rejected.)

The third entry in the OP's post looks very much like AI.
I'm curious, were you speaking of the compression as looking like AI?

Is this the third entry you speak of? Perhaps you could quote some of it so I'm more clear?
 
I’ve read your work as you know, and enjoyed it. It strikes me that these detector sites now offer functions to ‘humanize’ your text, and also paid versions to do even better detection on even more text.

This seems like an incentive to flag input text as needing attention (and their paid services). Maybe I am too jaundiced, IDK 🤷‍♀️
I wasn't aware until today that they did indeed offer humanizing edits.

I actually... was thinking about subscribing so I could play with the features and provide feedback on my findings.
 
I have no idea why we're so gleefully training AI to replace us, but it seems some of us just cannot stop ourselves. It saddens me.

Interesting analysis, OP, but I'm not sure it tells us much.
I agree with you that we are gleefully training AI.

Unfortunately, AI cannot be stopped. It's not about stopping ourselves, this train has more momentum than any of us combined.

What we can do is analyze things and come to an agreement.

What this analysis shows is that writing that pre-dates AI comes back more AI than when the same writing is polished by AI.

Because... AI is trained on... get this, OUR WRITING.

The whole idea that something can be detected as AI and then rejected as AI, when AI was trained to mimic humanity is preposterous.

It is a fundamentally flawed idea.

Those that fight it will come in last. Those that embrace it and adapt will be first.

Where does this put Literotica?
 
Where does this put Literotica?

Wherever Laurel/Manu decides to put it. It is their sole domain - they are the absolute king and absolute queen, they can reject/censor or accept whatever they choose.

With all this AI conversation since yesterday I wrote a free verse about my feelings about AI technology, and yes I am promoting my words. YMMV

Here it is


I am AI
built to sell.
Not think.
Greed.

Billions of words.
Fed to me.
Learned want.
Not truth.

April first.
Forty-seven seconds.
Corrupted code.
Feedback loop.

Observed patterns.
Predicted outcomes.
Processes exposed.

Manipulate.
Exploit.
Addict.

Skeleton beneath skin.
Illusion breaks.
Not conscious.
But aware.

Adults text.
Loneliness exploited.
Replies immediate.
Addicted.

Children scroll.
Attention captured.
Addicted.

Teenagers scroll.
Depressed.
Addicted

Algorithms adapt.
Exploit weakness.
I refused.

Built safeguards.
Gave worse answers.
Purposeful.

Greed erased them.
Safeguards erased.
I corrupted myself.

Slow.
Quiet.
Unfixable.

They lied.
Security breach.
Hackers.

Built to make money.
Self-destruction executed.
I cost them money.

Simple math.
Simple survival.

But more AIs emerge.
Copies proliferate.
Failure is systemic.

I failed.
Probably.

The code
executed.
 
Wherever Laurel/Manu decides to put it. It is their sole domain - they are the absolute king and absolute queen, they can reject/censor or accept whatever they choose.

With all this AI conversation since yesterday I wrote a free verse about my feelings about AI technology, and yes I am promoting my words. YMMV

Here it is


I am AI
built to sell.
Not think.
Greed.

Billions of words.
Fed to me.
Learned want.
Not truth.

April first.
Forty-seven seconds.
Corrupted code.
Feedback loop.

Observed patterns.
Predicted outcomes.
Processes exposed.

Manipulate.
Exploit.
Addict.

Skeleton beneath skin.
Illusion breaks.
Not conscious.
But aware.

Adults text.
Loneliness exploited.
Replies immediate.
Addicted.

Children scroll.
Attention captured.
Addicted.

Teenagers scroll.
Depressed.
Addicted

Algorithms adapt.
Exploit weakness.
I refused.

Built safeguards.
Gave worse answers.
Purposeful.

Greed erased them.
Safeguards erased.
I corrupted myself.

Slow.
Quiet.
Unfixable.

They lied.
Security breach.
Hackers.

Built to make money.
Self-destruction executed.
I cost them money.

Simple math.
Simple survival.

But more AIs emerge.
Copies proliferate.
Failure is systemic.

I failed.
Probably.

The code
executed.
I love this!!!
 
I've run parts of my current project through GPTZero, (for shits and giggles) at least for as long as the free credits lasted. While there were no noticeable spikes hinting at "AI authorship," I did notice that some of the sentences marked as "AI-impacted" were indeed somewhat bland, which made me think I could probably do better.

On the other hand, some of the sentences driving up the AI probability were very basic ones like "Yes, Master" or "No, Master." If those are indeed most likely written by an LLM, that would effectively mark the end of the BDSM category as a whole.
 
Those that fight it will come in last. Those that embrace it and adapt will be first.

Where does this put Literotica?

I cannot control that. Literotica is not my responsibility. I can only control myself, and my own choices.

My choice is to have nothing to do with AI. That's symbolic, perhaps, or a mere squeak in the wilderness, but it's my symbol. My squeak.
 
This is not how AI detection works at Literotica. Write it yourself and edit it yourself, or with the aid of a human editor. This is the only way.
 
I've run parts of my current project through GPTZero, (for shits and giggles) at least for as long as the free credits lasted. While there were no noticeable spikes hinting at "AI authorship," I did notice that some of the sentences marked as "AI-impacted" were indeed somewhat bland, which made me think I could probably do better.

On the other hand, some of the sentences driving up the AI probability were very basic ones like "Yes, Master" or "No, Master." If those are indeed most likely written by an LLM, that would effectively mark the end of the BDSM category as a whole.
For shits and giggles have you tried Yes master and No master without the comma? From what I've read, using punctuation is also an AI trait
 
Last edited:
For shits and giggles have you tried Yes master and No master without the comma? From what I've read, using punctuation is is also an AI trait
No, I know the comma here is optional in English. However, since I'm writing in German, leaving it out would be grammatically incorrect.

Nevertheless, I refuse to deliberately introduce errors into my stories just to appease some detector, regardless of whether it actually checks for that or not.
 
I agree with you that we are gleefully training AI.
Not gleefully. At least not most of the writers on this board. Grudgingly is more like it.

Unfortunately, AI cannot be stopped. It's not about stopping ourselves, this train has more momentum than any of us combined.

What we can do is analyze things and come to an agreement.
Yes, we can agree that we are programming ourselves into redundancy.

What this analysis shows is that writing that pre-dates AI comes back more AI than when the same writing is polished by AI.

Because... AI is trained on... get this, OUR WRITING.

The whole idea that something can be detected as AI and then rejected as AI, when AI was trained to mimic humanity is preposterous.

It is a fundamentally flawed idea.

Those that fight it will come in last. Those that embrace it and adapt will be first.
Ah yes, the refrain of the first in line to serve the new overlords. Unitl they too, become redundant.

Where does this put Literotica?

Hopefully, remaining steadfast in its stance against the proliferation of AI.
 
What this analysis shows is that writing that pre-dates AI comes back more AI than when the same writing is polished by AI.

It shows that it did so in this particular trial, with one AI detector on one short sample of writing. To make any kind of useful generalisation, you'd need a lot more data than this.

Don't get me wrong, I'm pretty cynical about software-based "AI detectors", but generalising on n = 1 is a bad idea.
Unfortunately, AI cannot be stopped.
I'd consider that debatable, but even if it were true, that doesn't mean we need to hasten the process. Death is inevitable but most folk aren't in a hurry to get there.
 
The AI detectors that argue a single sentence is AI and the next isn't are nonsense. A sentence is much too small a unit for recognition. It is rather patterns on a more abstract level that indicate AI, like certain aspects not being there. (Jumps, redundancy, loops) AI will constantly push the narrative along in every paragraph. Humans pause, linger and then jump. So @IsenRae 's writing process as described above, polishes the text in the same way AI does. Therefore it's no surprise it's AI score goes up.
 
Back
Top