Why AI writing is so generic, boring, and dangerous: Semantic ablation.
(We can measure semantic ablation through entropy decay. By running a text through successive AI "refinement" loops, the vocabulary diversity (type-token ratio) collapses.)
https://www.theregister.com/2026/02/16/semantic_ablation_ai_writing/
I see a lot of independent authors using machine translation for international markets.
While this might be initially cheaper, I suspect this will do them no favor in the long run.
@cstross I am sitting in a room different from the one you are in now. I am recording the sound of my speaking voice and I am going to play it back into the room again and again until the resonant frequencies of the room reinforce themselves...
@cstross hmmm, that might also explain why AI seems more effective for code.
For the most part you want a reversion to the mean in code. Novel solutions are only needed at the cutting edge where you trying to make the computer do something that’s not been done before.
@Jmj Yes. Also I suspect the semantic expressiveness of programming languages is far narrower than that of human languages: they're more precise, but it's much harder (though not impossible!) to write poetry in them. So there's less risk of losing something unique by generating output that tends to occupy the middle of the bell curve.
@cstross ironically got a Google cloud genAI and ML ad right in the middle of that.
@cstross No surprise that we see the textual equivalent of mad cow disease.
@cstross it's the textual equivalent of prions
@cstross I've previously described LLM-generated text as reading like "a middle management memo that no-one bothers reading". 🤷🏻♂️
@cstross by putting a measurable number on this feature, you have now made it possible to train out!
If you use an LLM to make “objective” decisions or treat it like a reliable partner, you’re almost inevitably stepping into a script that you did not consent to: the optimized, legible, rational agent who behaves in ways that are easy to narrate and evaluate. If you step outside of that script, you can only be framed as incoherent.
That style can masquerade as truth because humans are pattern-matchers: we often read smoothness as competence and friction as failure. But rupture in the form of contradiction, uncertainty, “I don’t know yet,” or grief that doesn’t resolve is often is the truthful shape of truth.
AI is part of the apparatus that makes truth feel like an aesthetic choice instead of a rupture. That optimization function operates as capture because it encourages you to keep talking to the AI in its format, where pain becomes language and language becomes manageable.
The only solution is to refuse legibility.
It's already beginning, where people speak the same words as always, but they don't mean the same things anymore from person to person.
New information from feedback that doesn't fit another's collapsed constraints for abstraction... can only be perceived as a threat. Because If you demand truth from a system whose objective is stability under stress, it will treat truth as destabilizing noise.
Reality is what makes a claim expensive. A model tries to make a claim cheap.
Systems that treat closure as safety will converge to smooth, repeatable outputs that erase the remainder. A useful intervention is one that increases the observer’s ability to detect and resist premature convergence—by exposing the hidden cost of smoothness and reinstating a legitimate place for uncertainty, contradiction, and falsifiability. But the intervention only remains non-doctrinal if it produces discriminative practice, not portable slogans.
Of cause it does. So the result becomes more and more readable for the deliberately uneducated masses. Style? Content? Facts? Who cares?
@cstross neat article, thanks.
I had a realization a while ago that LLM writing came at me with the same vibe I caught when I was briefly a teacher, and again in the workplace, where I dealt with people who had unacknowledged literacy challenges. Young folks who assembled written work by cribbing from others and rearranging words “by shape” to fulfill the requirements - always managed to convey zero meaningful thought.
I can’t help seeing in that elements of 1984 where Orwell describes successive reduction in vocabulary with the intended goal of making rebellious thought impossible
@cstross It is impossible to replace the human experience with a machine. The moment by its nature is sancrosanct; it's only in this atmosphere of gaming real estate insanity where life's nature is just another bitcoin to earn where we have lost our way.
@cstross there’s worse related things.
We come to see anything that the AI can not and does not produce as invalid and thus reading these bullshit, taupe texts shrinks our creative range, our sense of the possible, snd our willingness to forge out own path or follow someone else down there’s to knew territory.
Narrowing the range of semantics to an average is one thing.
Strangling our range of ideas is another.
@cstross As someone who, next to having a little knowledge about LLMs, was once complemented for choice of words by a native-speaking professor I hung around with at a conference for a few days, I am not surprised about this LLM fact. The professor however then was somewhat surprised of his own uttering and continued "but I had a few beers".
@cstross this is indeed a very neat explanation why the best possible outcome of an LLM is still terrible.
@cstross Is this article ai?
Although it's not within xyr academic field, Claudio Nastruzzi has touched upon the subject in at least one opinion piece before.
@cstross Well, that explains why every time I tried #AI for improving a text snippet, I was very disappointed.
It always wants to convince me that I should remove any uncommon sentence structures and replace them with generic ones, which often removes any personality from written text.
If you ask it to make a small addition to an existing text, it likes to rephrase everything in a more generic way and is unable to add a subtext layer. Honestly, it's just useless for writing in my opinion.
@cstross "beigification' is aspect of the recursive pollution problem
https://berryvilleiml.com/2026/01/10/recursive-pollution-and-model-collapse-are-not-the-same/
@cstross Blandness as a service.
Really great article, thanks for posting.
I know those who use ChatGPT to improve professional work letters. AI ChatGPT doesn't "improve" writing to refine meaning and nuance, it is designed to dumb down writing in successive loops to the most middling 6th grade level of comprehension thereby "improving" its "reach". It's the specifically designed feature of AI "editing".
I recently read a ChatGPT "improved" letter that may violate the Civil Rights Act by using a term that is a red flag for a lawsuit. I alerted the sender who confessed to using ChatGPT because her boss uses it and raves about it.
An ungraceful letter that keeps you legal is safer than a "clean" "professional" GPT letter that gets you sued. Know the risks!
Or in other words "Why AI is writing in corporate speech"
Reducing the text to the most common denominators?
@cstross on my rare exposure to the FB algo. feed, I see a long stream of heart-wrenching pleas for cats needing adoption, all matching a very standard and repetitive structure. I suspect the Cats are real and need help, but the pleas are "AI".
ActBlue (US Democratic political donation platform) suffered this even pre-LLM with its endless torrent of nearly identical begs for more money.
@cstross I find it hard not to wonder if the attraction of AI doesn't include so much forcible normalization of text that there is never anything unfamiliar, and thus never any reason to question a self-assessment of "smart".
(So much gets so messed up by mistaking responses for properties.)
