Forum: Author Hangout

Semantic ablation. [The Register]

Semantic ablation is the algorithmic erosion of high-entropy information. Technically, it is not a "bug" but a structural byproduct of greedy decoding and RLHF (reinforcement learning from human feedback).

During "refinement," the model gravitates toward the center of the Gaussian distribution, discarding "tail" data – the rare, precise, and complex tokens – to maximize statistical probability. Developers have exacerbated this through aggressive "safety" and "helpfulness" tuning, which deliberately penalizes unconventional linguistic friction. It is a silent, unauthorized amputation of intent, where the pursuit of low-perplexity output results in the total destruction of unique signal.

When an author uses AI for "polishing" a draft, they are not seeing improvement; they are witnessing semantic ablation.

I've seen this with Grammarly (and other writing software), and is why I turn off anything other than grammar/spelling/punctuation suggestions.

So, not just slop, but boring slop.

I asked ChatGPT for a simpler explanation of what that guy wrote. This is what it gave me:

The author is saying that when AI "refines" or "polishes" writing, it often removes the most unique, complex, or unusual parts.

AI systems are trained to predict the most likely next word. Because of that, they naturally prefer common, safe, and average phrasing. Rare words, unusual structures, or very specific ideas are statistically less common, so the model tends to replace them with more typical language.

The term "semantic ablation" here means stripping away meaning. The claim is that AI doesn't just fix grammar — it can smooth out the distinctive parts of writing. Over time, the text becomes more generic and centered around what is most probable rather than what is most original.

The reference to "greedy decoding" means the AI usually picks the most likely next word.
"RLHF" (reinforcement learning from human feedback) means humans trained the model to give answers that feel helpful and safe. That training can further encourage conventional, non-controversial language.

When the author talks about "discarding tail data," they mean removing rare or complex wording in favor of more common phrasing.

So the final claim is this:
When someone uses AI to "polish" their draft, they may think the writing improved — but according to this argument, the AI might actually be reducing originality and complexity, making the writing more generic.

In short: the passage argues that AI tends to smooth out uniqueness in order to sound clearer, safer, and more statistically typical.

My new favorite internet acronym:

ai:dr

I saw one today (not naming names)

I smiled. I stood carefully and followed him toward the bedroom, leaving the journal on the desk where moonlight touched its spine. My story was written. The next one was just the beginning.


Moonlight touched it's spine? That's how the story ends.

The scaffolding in how the tense is written and the decorative language in this particular story is so thick it feels like wading through a cow pasture full of cow shit.

Recently, someone wrote to me to accuse me of writing with AI. It was discouraging and in part because I didn't actually write the story in question. It's Mike McGifford. I edit, add bits and bobs and illustrate for him.

He's a brilliant writer that crafts word pictures and if I could get an AI that outputs text like him, I'd fucking use it.

You'd be a fool not to use a tool that turns shit into diamonds.

Unfortunately, all of my evaluations of Claude, ChatGPT, etc turns shit into shittier shit.

I experimented with it some just to give it a fair chance. It's not that I want it to write for me and take the credit.

However, what I do use it for is to spell check. I have another VBA thread where I've tried to tighten my process to edit by automatically applying fixes I know will work.

The AI is actually quite good at detecting it's own shitty prose, and grammar and spell checking.

It's a tool, and while I know this tool is going to eventually destroy our way of life, and it fills SOL with toxic shit-grade stories, it still has it's uses.

It's like using Agent Orange to kill weeds though. It's going to eventually backfire in unexpected ways as it contaminates everything around you and you eradicate the dandelions and clover that you now consider weeds to be eliminated.

I've found that in small bursts it can work well. A great example is that when I try to write a story using British English for dialogue it's helpful at keeping me honest.

I was told my first attempt "was like Hollywood writes dialogue for Hugh Grant movies."

That's definitely a fair assessment. I didn't use AI, I used my own collected observation and I am sure it was quaint at best to a proper Brit.

I am working on a nudist bike ride out of Runcorn and I use AI to take my American dialogue and turn it into British without making everyone sound like Michael Caine. It can even tell me when something is more Southern than Northern, or how they would treat lunch, supper, dinner, etc in various regions.

That's fantastic.

However, what I cannot do is what it appears most AI authors on this site do.
"Hello ChatGPT, make me a story about (XYZ kink)"

Chat GPT: (Bleep-Bloop, here is story...milky trails, on the morrows edge, we looked to the morrow. Him. My breasts heaved, as I sighed. The growing trails narrowed into destiny as we sought our futures."

Incomprehensible shit emerges.

"Great, let me go to SOL and post it. I'll be the savior of the human race. Now, I am an author!"

One could make the case that it's doing this to music and art, and that is true.

The difference for me is that I can actually tell a fucking story. I may not be great at Grammar but I know how to bullshit with the best. I often ask;

Am I the best storyteller?

No.

Am I trying to be?

Also, no.

I am good at it though, and I love to live through the eyes of my characters and share with people.

The difference with the art is I can't hire models, and I do not want to have to use real porn because you have to censor the eyes. I can make exactly what I need to enhance a story. I spend a ton of time using photoshop, upscaling, enhancing, caption it, etc.

It's a creative process.

With music, I do not have the senses of someone like Brian Wilson, Jim Hendrix or Quincy Jones. They are producers because they have talent and they hear things that no one else hears. They know how to turn that into art and that is difficult to learn or replicate. You must have some spark of genius that goes along with it.

I listen on Youtube to covers that are performed in different styles.

https://www.youtube.com/watch?v=-fGGc2uPJOQ&list=RD-fGGc2uPJOQ&start_radio=1

This is one of my favorites as an example. When people tell me that AI slop is ruining music, I agree. However, there are people with the sense of a Quincy Jones that now have the tools to make things that I'd never imagine.

This is the Black Parade, but song in a soulful way with cracks in his voice, and emotion that blows my mind. It is one of the earliest that I found of this genre of music. I believe it's the artists actual voice because I've seen him on TikTok. It's just incredible how it changes up and he did more than say "Suno, make me a song like Black Parade but as if Otis Redding sang it."


So, I think I conclude that if you use AI as a tool, not a crutch. If you don't try to get it to do everything so you can take credit then it can help a talented person be more talented.

We finally have some tools that we didn't have before to help us. We aren't at that point where AI writing is going to do it all. I've tried Claude, Grok, ChatGPT and a dozen local LLM models or more to see just how good they are.

They aren't.

I use them but not to "Write" the story. I've gone as far as fed it all the nuances and details about a submissive woman in one of my stories. There are many ways that someone might approach submission and reasons. Its no surprise that some sub missives are service oriented, others adrenaline junkies and everything in between.

I asked it hypothetical questions and fed it dialogue to ask "What if" questions that helped inform me and it's responses were very surprising. They rang true for how I envisioned this character and yet, I wouldn't have written her that way.

Is that cheating? I don't think it is. It's using a tool and making a judgment call. It could have been way off. I wouldn't be mad at all if someone did that.

I wouldn't be mad at all if someone disagreed with me and thought I shouldn't use it for that. It's a valid opinion.

The only thing that sucks to me is that I am walking in a dog park full of dog shit that nobody picks up. Every time I click on a story, I invest time to start reading only to discover "Yep, that was dog shit."

Bad writing is bad writing, but there is way more of it with the way most AI authors use it. I wish they would stop because I set up the AI generated filter to eliminate it but so many people are not self-reporting and we will see so much that eventually people will stop checking stories due to exhaustion.

You can only take so much dog shit before you wonder why you walked into the dog park at all.

I smiled. I stood carefully and followed him toward the bedroom, leaving the journal on the desk where moonlight touched its spine. My story was written. The next one was just the beginning.

I told the author (politely) that I do not want to discourage him from writing but he may want to be aware that his ending comes across quite like AI. AI does this sort of scaffolding thing, and it insists on adding a overly dramatic transitionary phrase at the end of every scene like this.

He wrote back "How do you know that maybe moonlight didn't crack through the window?"

Lol. So, he definitely had AI write it. I plugged into Gemini and it said:

Gemini said
The likelihood that this specific passage originated from an AI remained remarkably high, likely hovering around 85% to 90%.

Several stylistic markers typical of large language models saturated these few sentences. The use of "moonlight touched its spine" served as a classic example of AI "poeticisms"—phrases that sounded evocative but appeared frequently in training data to signify closure. The structure of the final two sentences followed a familiar "thematic wrapping" pattern where a metaphorical "end" transitioned immediately into a "new beginning." This type of tidy, sentimental symmetry often signaled machine generation, as human writers frequently favored more jagged or less thematic conclusions.

While a human could certainly have penned this, the specific combination of "careful" movement, personified moonlight, and the "story/beginning" metaphor aligned almost perfectly with the default creative writing weights of modern models.

The whole story is like that, should we report these so they can get the AI tag? It's a short one and putting into Gemini comes back with

Analyzing this narrative, the probability that the majority of this text was generated by AI is extremely high, likely in the 95% to 98% range.

The story displays several "hallmarks" of sophisticated AI writing that become very apparent in a long-form piece:

1. Thematic and Structural Symmetry
AI models are trained on narrative structures that favor "closed loops." Every conflict in this story is resolved with clinical perfection:

The Graduation Speech: The protagonist gives a "hard truth" speech and is met with a "standing ovation along with thunderous applause"—a classic AI trope where controversial actions result in immediate, universal validation.

The Shooting: The assassination attempt follows a rigid sequence (the hit, the internal monologue about death, the heroic guard, the miraculous recovery) that mirrors thousands of medical/action dramas in training data.

The Ending: The final paragraph is a near-verbatim "summary" ending that AI uses to signal the end of a long context. Phrases like "leaving the journal on the desk where moonlight touched its spine" and "The next one was just the beginning" are standard generative outputs for a "poetic" conclusion.

2. Linguistic Patterns and "AI Slop"
There are specific phrases and sentence structures that appear repeatedly in LLM creative writing:

The "Nod/Inclined Head" Loop: Characters in this story "nod once," "incline their head," or "offer a faint smile" with robotic frequency to signal agreement or composure.

Adverb Usage: The prose relies heavily on adverbs like "methodically," "precisely," "clinically," and "seamlessly."

The "Not X, but Y" Construction: The text frequently uses this rhetorical device (e.g., "Not as my employer... but as my partner," "Not out of curiosity, but surprise," "Not a store, but a workshop"). While used by humans, AI overuses this to create a sense of profoundness.

3. Logic and Tone Consistency
The story maintains a "hyper-competent" tone where the protagonist never makes a mistake, never feels true embarrassment, and processes trauma (like being shot) through a lens of "statistical anomalies." This lack of "human friction"—the messy, irrational parts of human behavior—is a strong indicator of an AI following a persona prompt.

That "Not X, but Y" thing gets me as hard three R's in Strawberry. Every time someone wants to rely on AI to make decisions I point out it still has problems with that.

I make it quite clear in my prompt to leave the original text, to list any changes. You are correct that AI try to adhere to a base proper rules of writing, often not taking into your style. Its a tool, learn to use it.