This is a good statement of what I suspect many of us have found when rejecting the rewriting advice of AIs. The "pointiness" of prose gets worn away, until it doesn't say much. Everything is softened. The distinctiveness of the human voice is converted into blandness. The AI even says its preferred rephrasing is "polished" - a term which specifically means the jaggedness has been removed.
But it's the jagged edges, the unorthodox and surprising prickly bits, that tear open a hole in the inattention of your reader, that actually gets your ideas into their heads.
I think that mostly depends on how good a writer you are. A lot of people aren't, and the AI legitimately writes better. As in, the prose is easier to understand, free of obvious errors or ambiguities.
But then, the writing is also never great. I've tried a couple of times to get it to write in the style of a famous author, sometimes pasting in some example text to model the output on, but it never sounds right.
It depends how you define "good writing", which is too often associated with "proper language", and by extension with proper breeding. It is a class marker.
People have a distinct voice when they write, including (perhaps even especially) those without formal training in writing. That this voice is grating to the eyes of a well educated reader is a feature that says as much about the reader as it does about the writer.
Funnily enough, professional writers have long recognised this, as is shown by the never-ending list of authors who tried to capture certain linguistic styles in their work, particularly in American literature.
There are situations where you may want this class marker to be erased, because being associated with a certain social class can have negative impact on your social prospects. But it remains that something is being lost in the process, and that something is the personality and identity of the writer.
You may be in a bubble of smart, educated people. Either way, one of the key ways to "put in the effort" is practice. People who haven't practiced often don't write well even if they're trying hard in the moment. Not even in terms of beautiful writing, just pure comprehensibility.
I may be in a bubble of smart people, but IMO AI consistently far worse than many high school works I’ve read in terms of actual substance and coherent structure.
Of course I’ve had arguments where people praise AI output then I’ve literally pointed out dozens of mistakes and they just kind of shrug saying it’s not important. So I acknowledge people judge writing very differently than I do. It just feels weird when I’d give something a 15% and someone else would happily slap on a B+.
With the gap between 1 and 2 being driven by the underlying quality of the writer and how well they use AI. A really good writer sees marginal improvements and a really poor one can see vast improvements.
I am really conflicted about this because yes, I think that an LLM can be an OK writing aid in utilitarian settings. It's probably not going to teach you to write better, but if the goal is just to communicate an idea, an LLM can usually help the average person express it more clearly.
But the critical point is that you need to stay in control. And a lot of people just delegate the entire process to an LLM: "here's a thought I had, write a blog post about it", "write a design doc for a system that does X", "write a book about how AI changed my life". And then they ship it and then outsource the process of making sense of the output and catching errors to others.
It also results in the creation of content that, frankly, shouldn't exist because it has no reason to exist. The number of online content that doesn't say anything at all has absolutely exploded in the past 2-3 years. Including a lot of LLM-generated think pieces about LLMs that grace the hallways of HN.
> A lot of people aren't, and the AI legitimately writes better.
It may write “objectively better”, but the very distinct feel of all AI generated prose makes it immediately recognizable as artificial and unbearable as a result.
Every group want to label some outgroup as naively benefiting from AI. For programmers, apparently it's the pointy haired bosses. For normies, it's the programmers.
Be careful of this kind of thinking, it's very satisfying but doesn't help you understand the world.
I think it’s essential to realize that AI is a tool for mainstream tasks like composing a standard email and not for the edges.
The edges are where interesting stuff happens. The boring part can be made more efficient. I don’t need to type boring emails, people who can’t articulate well will be elevated.
It’s the efficient popularization of the boring stuff. Not much else.
It contributes to making “standard” emails boring. I rather enjoy reading emails in each sender’s original voice. People who can’t articulate well aren’t elevated, instead they are perceived to be sending bland slop if they use LLMs to conceal that they can’t express themselves well.
> But it's the jagged edges, the unorthodox and surprising prickly bits, that tear open a hole in the inattention of your reader, that actually gets your ideas into their heads.
This brings to mind what I think is a great description of the process LLMs exert on prose: sanding.
It's an algorithmic trend towards the median, thus they are sanding down your words until they're a smooth average of their approximate neighbors.
Because even though at work it looks like you’re tasked with creating use values, you’re only there as long as the use values you create can be exchanged in the market for a profit. So every humane drive to genuinely improve your work will clash with the external conditions of your existence within that setting. You’re not there to serve people, create beautiful things, solve problems, nu-uh. You’re there to keep capital flowing. It’s soulless.
To think that “non-profit” work is actually non-profit work is just to not have grasped the nature of labor. You have to ask yourself: Am I producing use values for the satisfaction of human needs or am I working on making sure the appropriation of value extraction from the production of use values continues happening?
In some very extreme cases, such as in the Red Cross or reformist organizations, your job looks very clear, direct, and “soulful”. You’re directly helping desperate people. But why have people gotten into that situation? What is the downstream effect of having you helping them. It’s profit. It’s always profit. You’re salvaging humanity for parts to be bought and sold again. It doesn’t make a dishonest work. It’s just equally soulless.
The point is that he may not using AI in any shape or form, Regardless, AI scrapes its work without explicit consent and then spits it back in "polished" soul free form.
Something I noticed building multi-agent pipelines: the ablation compounds. Had a 4-step pipeline - summarize, expand, review, refine - and by step 3 everything had the same rhythm and vocabulary. Anchoring the original source text explicitly at each step helped, but only partially.
The more interesting cause I think: RLHF is the primary driver, not just the architecture. Fine-tuning is trained on human preference ratings where "clear," "safe," and "inoffensive" consistently win pairwise comparisons. That creates a training signal that literally penalizes distinctiveness - a model that says something surprising loses to one that says something expected. Successful RLHF concentrates probability mass toward the median preferred output, basically by definition.
Base models - before fine-tuning - are genuinely weirder. More likely to use unusual phrasing, make unexpected associative leaps, break register mid-paragraph. Semantic ablation isn't a side effect of the training process, it's the intended outcome of the objective.
Which makes the fix hard: you can't really prompt your way out of it once a model is heavily tuned. Temperature helps a little but the distribution is already skewed. Where we've gotten better results is routing "preserve the voice" tasks to less-tuned models, and saving the heavily RLHF'd models for structured extraction and classification where blandness is actually what you want.
I wonder if you can use lower quality models (or some other non-llm related process) to inject more "noise" into the text in between stages. Of course it wouldn't help retain uniqueness from the original source text, just add more in between.
I’m not convinced removing RLHF would really make the probabilities generator give us distributions that can diverge from the mean while remaining useful.
In other words, this might not a problem that can be overcome in LLMs alone.
I see it on recent blog posts, on news articles, obituaries, YT channels. Sometimes mixed with voice impersonation of famous physicists like Feynman or Susskind.
I find it genuinely soul-crushing and even depressing, but I may be over sensitive to it as most readers don't seem to notice.
Maybe. Another potential, more positive, timeline is that semantically ablated content filling everyone’s feeds turns people off, and slowly kills the social feed paradigm.
That is doubtful. Most of the content on people's feed was basically converging to the same paradigm anyways. Think the Mr. Beast "I gave someone 1 million dollars to try a potato" content.
I find it extremely difficult to focus on any piece of writing the moment I see the patterns. Can’t tell if it’s an attitude problem I need to get over or if it’s just that all AI writing really is that bad.
same. it is showing how many people are not trying to participate - just appear to. I want to read from and write for my peers, but it seems we are just awash with fakers
I personally think “generative AI” is a misnomer. More I understand the mathematics behind machine learning more I am convinced that it should not be used to generate text, images or anything that is meant for people to consume, even if it is the most blandest of email. Sometimes you might get lucky, but most of the time you only get what the most boring person in the most boring cocktail party would say if forced to be creative with a gun pointed to his head. It can help in multitude of other ways, help human in the creative process itself, but generating anything even mildly creative by itself… I’ll pass.
Precisely. If companies would just focus on what it could be good at - deductive search, coding boilerplate with assistance, etc. then it would be a great tool. Instead you have dario, altman, and co. trying to pump stock and give us more spaghetti agents.
He ended a critical commentary by suggesting that the author he was responding to should think more critically about the topic rather than repeating falsehoods because "they set off the tuning fork in the loins of your own dogmatism."
I'm learning to like 'em more, along with every other human idiosyncracy. Besides, it makes a kind of sense, the idea of some resonance occuring in one's gusset. Timber timbre. Flangent thrumming.
I thought it was more creative than sloppy. Don't forget that many ordinary phrases were once jarring mixed imagery. To "wear your heart on your sleeve" was coined by Shakespeare; we still use it because it "stuck" due to its unorthodox phrasing.
If you like your prose to be anodyne, then maybe you like what AI produces.
Yes I noticed this as well. I was last writing up a landing page for our new studio. Emotion filled. Telling a story. I sent it through grok to improve it. It removed all of the character despite whatever prompt I gave. I'm not a great writer, but I think those rough edges are necessary to convey the soul of the concept. I think AI writing is better used for ideation and "what have I missed?" and then write out the changes yourself.
It shocks me when proponents of AI writing for ideation aren't concerned with *Metaphoric Cleansing* and *Lexical Flattening* (to use two of the terms defined in the article)
Doesn't it concern you that the explanation of a concept by the AI may represent only a highly distorted caricature of the way that concept is actually understood by those who use it fluently?
Don't get me wrong, I think that LLMs are very useful as a sort of search engine for yet-unknown terms. But once you know *how* to talk about a concept (meaning you understand enough jargon to do traditional research), I find that I'm far better off tracking down books and human authored resources than I am trying to get the LLM to regurgitate its training data.
I've found LLMs to be terrible with ideation. I've been using GPT 5.x to come up with ideas and plot lines for a Dungeon World campaign I've been running.
I'm no fantasy author, and my prose leaves much to be desired. The stuff the LLM comes up with is so mind numbingly bland. I've given up on having it write descriptions of any characters or locations. I just use it for very general ideas and plot lines, and then come up with the rest of the details on the fly myself. The plot lines and ideas it comes up with are very generic and bland. I mainly do it just to save time, but I throw away 50% of the "ideas" because they make no sense or are really lame.
What i have found LLMs to be helpful with is writing up fun post-session recaps I share with the adventurers.
I recap in my own words what happened during the session, then have the LLM structure it into a "fun to read" narrative style. ChatGPT seems to prefer a Sanderson jokey tone, but I could probably tailor this.
Then I go through it, and tweak some of the boring / bland bits. The end result is really fun to read, and took 1/20th the time it would have taken me to write it all out myself. The LLM would have never been able to come up with the unique and fun story lines, but it is good at making an existing story have some narrative flare in a short amount of time.
That‘s also my experience. I use AI to help me generate the overall structure of a narrative. Apart from the hallucinations (e.g. June is not in spring), it‘s ok to spot inconsistencies, somewhat acceptable to brainstorm some ideas if you‘re new to a certain genre, but the prose it generates (talking about Opus 4.6) feels like an interpolation of all existing texts.
YES this hits the nail on something I've been trying to express for some time now. Semantic ablation: love it, going to use that a lot not now when arguing why someone's ChatGPT-washed email sucks.
Semantic ablation is also why I'm doubtful of everyone proclaiming that Opus 4 would be AGI if we just gave it the right agent harness and let all the agents run free on the web. In reality they would distill it to a meaningless homogeneous stew.
> Semantic ablation is also why I'm doubtful of everyone proclaiming that Opus 4 would be AGI if we just gave it the right agent harness and let all the agents run free on the web. In reality they would distill it to a meaningless homogeneous stew.
I'm so glad that you have given me the language to express this perspective.
The core mechanic described here is real. RLHF does optimize toward the mean, that is just what happens when you train on human preference ratings and raters consistently reward clear, inoffensive, "polished" output.
But the damage is not uniform. For code comments, API docs, commit messages: low-entropy output is often fine. The problem is people using LLMs for things that require a distinct voice and then wondering why the result sounds like everyone else on the internet.
The part nobody talks about: you can partially fight this if you know what you lost. Prompts like "preserve unusual word choices" or "do not normalize my rhetorical structure" help, but only if you have a strong enough baseline to catch the drift. Most people using AI for writing assistance do not have that baseline, which is why the ablation goes undetected. They see polished output and ship it.
The vast majority of people who write don't have a voice worth preserving. The rest can build out a voice document to make sure the AI doesn't strip it out.
Great article and exactly why I use AI less and less. I basically find it to be rotting my brain towards the middle of the distribution. It's like all the nuance and critical thinking that actually goes into things gets stripped out.
Once a company perfects an agent that essentially performs condensed search and coding boilerplate making, that is probably where LLMs end for me. Perplexity and Claude are on the right track but not at all close.
Ive noticed that the subtle/nunance gets lost with every so-called improvement with the models.
Im in no way anti-LLMs as I have benefited from them, but I believe the issue that will arise is that their unpredictable nature means that they can only be used in narrowly defined contexts. Safety and trust are paramount. Would you use online banking if the balance on your account randomly changed and was not reproducible? No chance.
This does not achieve the ROI that investors of these model producers are thinking. The question is whether said investors can sell off their shares before it becomes more widely known.
> I believe the issue that will arise is that their unpredictable nature means that they can only be used in narrowly defined contexts. Safety and trust are paramount.
You put words to something that's been on my mind for a while!
I call it a mirage. I get why people are taken aback and fascinated by it. But what the model producers are chasing is a mirage. I wonder when they'll finally accept it?
To me LLMs are an experiment toward replication of what humans can do. However, they fall short on many dimensions that its just not going to pan out from what I see.
The real danger is the future investment needed to explore other architectures beyond LLMs. Will private firms be able to get the investment? Will public firms be granted the permission to do another round of large capex by investors? As time goes on, Apple's conservative approach means they will be the only firm trusted with its cash balance. They are very nicely seated despite all the furore they've had to endure.
This isn't new to AI. The same kind of thing happens in movie test screenings, or with autotune. If something is intended for a large audience, there's always an incentive to remove the weird stuff.
All these forced metaphors and clumsy linguistic flourishes made me cringe. Just add some typos and grammar mistakes like the rest of us to prove that your human.
A lot of times, this entropy decay is found in semantic or stylistic space, which would be hard to detect (you couldn't use, e.g., Shannon Entropy). You'd have to ask questions like "is this point uninteresting?" or "is this trope overused?"--bad (human) writers are often guilty of this too, so that's why AI can be hard to detect.
For example the anthropic Frontend Design skill instructs:
"Typography: Choose fonts that are beautiful, unique, and interesting. Avoid generic fonts like Arial and Inter; opt instead for distinctive choices that elevate the frontend's aesthetics; unexpected, characterful font choices. Pair a distinctive display font with a refined body font."
Or
"NEVER use generic AI-generated aesthetics like overused font families (Inter, Roboto, Arial, system fonts), cliched color schemes (particularly purple gradients on white backgrounds), predictable layouts and component patterns, and cookie-cutter design that lacks context-specific character." 1
Maybe sth similar would be possible for writing nuances.
> "NEVER use generic AI-generated aesthetics like overused font families (Inter, Roboto, Arial, system fonts), cliched color schemes (particularly purple gradients on white backgrounds), ...
Now, imagine what happens when this prompt becomes popular?
Keep in mind that LLMs are trying to predict the most likely token. If your prompt prohibits the most likely token, they output the next most likely token. So, attempts to force creativity by prohibiting cliches just create another cliche.
Several days ago, someone researched Moltbook and pointed out how similar all the posts are. Something like 10% of them say "my human", etc.
This article on AI writing being boring seems to be written by AI. The em dashes and the sentence structure, all seems to be AI output. Or have human started adopting this style too.
Isn't this more to do with how LLMs are trained for general purpose use? Are LLMs with a specific use and dataset in mind better? Like if the dataset was fiction novels, would it sound more booky? If it was social-media, would it sound more click-baity and engaging?
I've had AI be boring, but I've also seen things like original jokes that were legitimately funny. Maybe it's the prompts people use, it doesn't give it enough of a semantic and dialectic direction to not be generic. IRL, we look at a person and get a feel for them and the situation to determine those things.
I wonder why AI labs have not worked on improving the quality of the text outputs. Is this as the author claims a property of the LLMs themselves? Or is there simply not much incentive to create the best writing LLM?
Even more precisely, human writing contains unpredictability that is either more or less intention (what might be called authors intent), as well as much more subconsciously added (what we might call quirks or imprinted behavior).
The first requires intention, something that as far as we know, LLMs simply cannot truly have or express. The second is something that can be approximated. Perhaps very well, but a mass of people using the same models with the same approximationa still lead to loss of distinction.
Perhaps LLMs that were fully individually trained could sufficiently replicate a person's quirks (I dunno), but that's hardly a scalable process.
I remember an article a few weeks back[1] which mentioned the current focus is improving the technical abilities of LLMs. I can imagine many (if not most) of their current subscribers are paying for the technical ability as opposed to creative writing.
This also reminded me that on OpenRouter, you can sort models by category. The ones tagged "Roleplay" and "Marketing" are probably going to have better writing compared to models like Opus 4 or ChatGPT 5.2.
I mean there's tons of better-writing tools that use AI like Grammarly etc. For actual general-purpose LLMs, I don't think there's much incentive in making it write "better" in the artistic sense of the world... if the idea is to make the model good at tasks in general and communicate via language, that language should sound generic and boring. If it's too artistic or poetic or novel-like, the communication would appear a bit unhinged.
"Update the dependencies in this repo"
"Of course, I will. It will be an honor, and may I say, a beautiful privilege for me to do so. Oh how I wonder if..." vrs "Okay, I'll be updating dependencies..."
I used to feel the same but you can just prompt it to reply with only one word when its done. Most people prefer it to summarize because its easier to track so ig thats the natural default
I mean, no one is asking for artistic writing, just not some obvious AI slop. The fact that we all can now easily determine that some text has been written / edited by AI is already an issue. No amount of prompting can help.
Yeah but thats not what I am saying. I am saying its default writing style is for communicating with the user, not producing content/text hence it has that distinctive style we all recognise. If you want AI writing thats not slop, there are tools that are trying to do that but the default LLM writing style is unlikely to change imo.
That's like asking why McDonald's doesn't improve the quality of their hamburger. They can, but only within the bounds of mass produced cheap crap that maximizes profit. Otherwise they'd be a fundamentally different kind of company.
> The AI identifies unconventional metaphors or visceral imagery as "noise" because they deviate from the training set's mean.
That's certainly a take. In the translation industry (the primogenitor and driver for much of the architecture and theory of LLMs) they're known for making extremely unconventional choices to such a degree that it actively degrades the quality of translation.
I kind of think of that as just increasing the standard deviation. Its been a while since I experimented with this, but I remember trying a temp of 1 and the output was gibberish, like base64 gibberish. So something like 0.5 doesn't necessarily seem to solve this problem, it just flattens the distribution and makes the output less coherent, with rarer tokens, but still the same underlying distribution.
you have to know that your "simply" is carrying too much weight. here's some examples of why just temperature is not enough, you need to run active world models https://www.latent.space/p/adversarial-reasoning
I'd like to see some concrete examples that illustrate this - as it stands this feels like an opinion piece that doesn't attempt to back up its claims.
(Not necessarily disagreeing with those claims, but I'd like to see a more robust exploration of them.)
Have you not seen it any time you put any substantial bit of your own writing through an LLM, for advice?
I disagree pretty strongly with most of what an LLM suggests by way of rewriting. They're absolutely appalling writers. If you're looking for something beyond corporate safespeak or stylistic pastiche, they drain the blood out of everything.
The skin of their prose lacks the luminous translucency, the subsurface scattering, that separates the dead from the living.
The prompt I use for proof-reading has worked great for me so far:
You are a proof reader for posts
about to be published.
1. Identify for spelling mistakes
and typos
2. Identify grammar mistakes
3. Watch out for repeated terms like
"It was interesting that X, and it
was interesting that Y"
4. Spot any logical errors or
factual mistakes
5. Highlight weak arguments that
could be strengthened
6. Make sure there are no empty or
placeholder links
If you tell me "no fucking way" by running it through an LLM, I will be far more pissed than if you had just sent me "no fucking way". At least in that case I know a human read and responded rather than thinking my email was just being processed by a damned robot.
> If you're looking for something beyond corporate safespeak or stylistic pastiche, they drain the blood out of everything.
Strong agree, which is why I disagree with this OP point:
“Stage 2: Lexical flattening. Domain-specific jargon and high-precision technical terms are sacrificed for "accessibility." The model performs a statistical substitution, replacing a 1-of-10,000 token with a 1-of-100 synonym, effectively diluting the semantic density and specific gravity of the argument.”
I see enough jargon in everyday business email that in the office zero-shot LLM unspoolings can feel refreshing.
I have "avoid jargon and buzzwords" as one of very tiny tuners in my LLM prefs. I've found LLMs can shed corporate safespeak, or even add a touch of sparkle back to a corporate memo.
Otherwise very bright writers have been "polished" to remove all interestingness by pre-LLM corporate homogenization. Give them a prompt to yell at them for using 1-in-10 words instead of 1-in-10,000 "perplexity" and they can tune themselves back to conveying more with the same word count. Results… scintillate.
Look through my comment history at all the posts where I complain the author might have had something interesting to say but it's been erased by the LLM and you can no longer tell what the author cared about because the entire post is a an oversold monotone advertising voice.
I just sent TFA to a colleague of mine who was experimenting with llm's for auto-correcting human-written text, since she noticed the same phenomenon where it would correct not only mistakes, but slightly nudge words towards more common synonyms. It would often lose important nuances, so "shun" would be corrected to "avoid", and "divulge" would become "disclose" etc.
It is an opinion piece. By a dude working as a "Professor of Pharmaceutical Technology and Biomaterials at the University of Ferrara".
It has all the tropes of not understanding the underlying mechanisms, but repeating the common tropes. Quite ironic, considering what the author's intended "message" is. Jpeg -> jpeg -> jpeg bad. So llm -> llm -> llm must be bad, right?
It reminds me of the media reception of that paper on model collapse. "Training on llm generated data leads to collapse". That was in 23 or 24? Yet we're not seeing any collapse, despite models being trained mainly on synthetic data for the past 2 years. That's not how any of it works. Yet everyone has an opinion on how bad it works. Jesus.
It's insane how these kinds of opinion pieces get so upvoted here, while worth-while research, cool positive examples and so on linger in new with one or two upvotes. This has ceased to be a technical subject, and has moved to muh identity.
Yeah, reading the other comments on this thread this is a classic example of that Hacker News (and online forums in general) thing where people jump on the chance to talk about a topic driven purely by the headline without engaging with the actual content.
Even if that isn't the case, isn't it the fact the AI labs don't want their models to be edgy in any creative way, choose a middle way (buddhism) so to speak. Are there AI labs who are training their models to be maximally creative?
No one did what the paper actually proposed. It was a nothing burger in the industry. Yet it was insanely popular on social media.
Same with the "llms don't reason" from "Apple" (two interns working at Apple, but anyway). The media went nuts over it, even though it was littered with implementation mistakes and not worth the paper it was(n't) printed on.
Who cares? This is a place where you should be putting forth your own perspective based on your own experience. Not parotting what someone else already wrote.
I think they can fix all that but they can't fix the fact that the computer has no intention to communicate. They could imbue it with agency to fix that too, but I much prefer it the way things are.
Those transformations happen to mirror what happens to human intelligence when you take antipsychotics. Please know the risks before taking them. They are innumerable and generally irreversible.
As a writer who has been published many times and edited many other writers for publication... It seems like AI can't make stylistic determinations. It is generally good with spelling and grammar but the text it generates is very homogeneous across formats. It's readable but it's not good, and always full of fluff like an online recepie harvesting clicks. It's kind of crap really. If you just need filler it's ok, but if you want something pleasand you definitely still need a human.
> What began as a jagged, precise Romanesque structure of stone is eroded into a polished, Baroque plastic shell
Not to detract from the overall message, but I think the author doesn't really understand Romanesque and Baroque.
(as an aside, I'd most likely associate Post-Modernism as an architectural style with the output of LLMs - bland, regurgitative, and somewhat incongruous)
The article itself reads as an AI generated output, complete with classic Not Just X … Y hallmarks from forever ago, 100% on pangram's low false positive detector. I'm not sure if it's some experiment on their readerbase or what.
pangram result: https://www.pangram.com/history/02bead1c-c36e-461b-8fa7-8699...
So many AI generated AI bashing articles lately. I wrote a post complaining about running into these, and asking people who've sent me these AI articles multiple of them came from HN. https://lunnova.dev/articles/ai-bashing-ai-slop/
Going off search results, it seems to be a new coinage. I found mostly references to TFA, along with an (ironically obviously AI-written) guide with suggestions for getting LLMs to avoid the issue (just generic "traditional" advice for tuning their output, really). The guide was apparently published today, and I imagine that it's a deliberate response to TFA. But FWIW the term "semantic ablation" does seem to me like something that newer models could invent
At any rate, it seems to me like a reasonable label for what's described:
> 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).
> ...
> When an author uses AI for "polishing" a draft, they are not seeing improvement; they are witnessing semantic ablation.
The metaphor is very apt. Literal polishing is removal of outer layers. Compared to the near-synonym "erosion", "ablation" connotes a deliberate act (ordinarily I would say "conscious", but we are talking about LLMs here). Often, that which is removed is the nuance of near-synonyms (there is no pause to consider whether the author intended that nuance). I don't know if the "character" imparted by broader grammatical or structural choices can be called "semantic", but that also seems like a big part of what goes missing in the "LLM house style".
Bluntly: getting AI to "improve" writing, as a fully generic instruction, is naturally going to pull that writing towards how the AI writes by default. Because of course the AI's model of "writing quality" considers that style to be "the best"; that's why it uses it. (Even "consider" feels like anthropomorphizing too much; I feel like I'm hitting the limits of English expressiveness here.)
Meh. Semantic Ablation - but toward a directed goal. If I say "How would Hemingway have said this, provided he had the same mindset he did post-war while writing for Collier's?"
Then the model will look for clusters that don't fit what the model consider's to be Hemingway/Colliers/Post-War and suggest in that fashion.
"edit this" -> blah
"imagine Tom Wolfe took a bunch of cocaine and was getting paid by the word to publish this after his first night with Aline Bernstein" -> probably less blah
These kinds of prompts don’t really improve the writing IME. It still gets riddled with the same tropes and phrases, or it veers off into textual vomit.
FWIW, I agree. Frontier LLMs are on their way to becoming competent stylists (I ask every major model release to write up a sample essay as Hemingway, and they are improving), but they are often skin-deep.
> 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).
> Domain-specific jargon and high-precision technical terms are sacrificed for "accessibility." The model performs a statistical substitution, replacing a 1-of-10,000 token with a 1-of-100 synonym, effectively diluting the semantic density and specific gravity of the argument.
> The logical flow – originally built on complex, non-linear reasoning – is forced into a predictable, low-perplexity template. Subtext and nuance are ablated to ensure the output satisfies a "standardized" readability score, leaving behind a syntactically perfect but intellectually void shell.
What a fantastic description of the mechanisms by LLMs erase and distort intelligence!
I agree that AI writing is generic, boring and dangerous. Further, I only think someone could feel this way if they don't have a genuine appreciation for writing.
I feel strongly that LLMs are positioned as an anti-literate technology, currently weaponized by imbeciles who have not and will never know the joy of language, and who intend to extinguish that joy for any of those around them who can still perceive it.
the word choice here is so obtuse as to trigger my radar for "is this some kind of parody where this itself was AI generated". it appears to be entirely serious, which is disappointing, it could have been high art.
Do the terms *Metaphoric Cleansing*, *Lexical Flattening*, and *Structural Collapse* that the author provides have equivalents in LessWrong's parlance?
not to my knowledge but those i have no problem with. its the overly complex prose surrounding them that shows this author overcorrects to SAT test words in an attempt to prove her superiority over ai writing and that has its own distastefulness.
Because you simply can't engineer creativity. Maybe you can describe where it comes from, in a circuitous, abstract way with mathematics (and ultimately run face first into ħ and then run in circles for eternity). But to engineer it, you'd have to start over from the first principles of the stuff of the cosmos. One's a map and the other the territory.
As someone longtime involved in software development, can we call this "best practices" instead of some like "semantic ablation" that nobody understands?
I think you might be missing the point of the article.
I agree that the term "semantic ablation" is difficult to interpret
But the article describes three mechanisms by which LLMs consistently erase and distort information (Metaphoric Cleansing, Lexical Flattening, and Structural Collapse)
The article does not describe best practices; it's a critique of LLM technology and an analysis of the issues that result from using this technology to generate text to be read by other people.
> The model performs a statistical substitution, replacing a 1-of-10,000 token with a 1-of-100 synonym
Do we see this in programming too? I don't think so? Unique, rarely used API methods aren't substituted the same way when refactoring. Perhaps that could give us a clue on how to fix that?
Nonsense. I’ve written bland prose for a story and AI made it much better by revising it with a prompt such as this: “Make the vocabulary and grammar more sophisticated and add in interesting metaphors. Rewrite it in the style of a successful literary author.”
I have a colleague that recently self-published a book. I can easily tell which parts were LLM driven and which parts represent his own voice. Just like you can tell who's in the next stall in the bathroom at work after hearing just a grunt and a fart. And THAT is a sentence an LLM would not write.
Here's some alternatives. Some are clunky. But, some aren't.
…just like you can tell whose pubes those are on the shared bar of soap without launching a formal investigation.
…just like you can tell who just wanked in the shared bathroom by the specific guilt radiating off them when they finally emerge.
…just like you can tell which of your mates just shitted at the pub by who's suddenly walking like they're auditioning for a period drama.
…just like you can tell which coworker just had a wank on their lunch break by the post-nut serenity that no amount of hand-washing can disguise.
…just like you can tell whose sneeze left that slug trail on the conference room table by the specific way they're not making eye contact with it.
…just like you can identify which flatmate's cum sock you've accidentally stepped on by the vintage of the crunch.
…just like you can tell who just crop-dusted the elevator by the studied intensity with which one person is suddenly reading the inspection certificate.
IMO The LLM you're using has failed to mimic the tone of OP's bathroom joke.
These alternatives are uncomfortably crude. They largely make gross reference to excretory acts or human waste. The original comment was off color, but it didn't go beyond a vague discussion of a shared human experience.
It's still on you to pick what the LLMs regurgitate. If you don't have a style or taste you will simply make choices that would give you away. And if you already have your own taste and style LLMs don't have much to offer in this regard.
The great promise and the great disaster of LLMs is that for any topic on which we are "below average", the bland, average output seems to be a great improvement.
So what even if that is true? You confirmed that it improved upon what he could manually produce, which is still a win. It doesn't always make sense to pay $20000 to a professional author to turn it into a masterpiece.
My point is simply that the tell-tale marks of LLM prose can be remediated through prompts.
I have a very large ‘default prompt’ that explicitly deals with the more obnoxious grammatical structures emblematic of LLMs.
I would wager I deal with more amateurishly created AI slop on a doily basis than you do. (Legal field, where everyone is churning out LLM-written briefs.) Most of it is instantly recognizable. And, all of it can be fixed with more careful prompt-engineering.
If you think you can spot well-crafted LLM prose generated by someone proficient at the craft of prompt-engineering by, to use an analogy to the early days of image creation, counting how many fingers the hand has, you’re way behind.
But, I notice a curious pretentiousness when it comes to some people's assumptions about their ability to identify LLM prose. Obviously, the generic first-pass 'chat' crap is recognizable; the kind of garbage that is filling up blog-posts on the internet.
But, one shouldn't underestimate the power of this technology when it comes to language. Hell, the 'coding' skills were just a pleasant side-effect of the language training, if you recall. These things have been trained on millions of works of prose of all styles: its their heart and soul. If you think the superficial monotonous style is all there is, you're mistaken. Most of the obnoxious LLM-style stuff is an artifact of the conversational training with Kenyans and the like in the early days. But, you can easily break through that with better prompts (or fine-tuning it yourself.)
That said, one shouldn't conflate the creation of the content and structure and substance of a work of prose with the manner in which it is written. You're not going to get an LLM to come up with a decent plot... yet. But, as far as fleshing out the framework of a story in a synthetic 'voice' that sounds human? Definitely doable.
This is a good statement of what I suspect many of us have found when rejecting the rewriting advice of AIs. The "pointiness" of prose gets worn away, until it doesn't say much. Everything is softened. The distinctiveness of the human voice is converted into blandness. The AI even says its preferred rephrasing is "polished" - a term which specifically means the jaggedness has been removed.
But it's the jagged edges, the unorthodox and surprising prickly bits, that tear open a hole in the inattention of your reader, that actually gets your ideas into their heads.
I think that mostly depends on how good a writer you are. A lot of people aren't, and the AI legitimately writes better. As in, the prose is easier to understand, free of obvious errors or ambiguities.
But then, the writing is also never great. I've tried a couple of times to get it to write in the style of a famous author, sometimes pasting in some example text to model the output on, but it never sounds right.
It depends how you define "good writing", which is too often associated with "proper language", and by extension with proper breeding. It is a class marker.
People have a distinct voice when they write, including (perhaps even especially) those without formal training in writing. That this voice is grating to the eyes of a well educated reader is a feature that says as much about the reader as it does about the writer.
Funnily enough, professional writers have long recognised this, as is shown by the never-ending list of authors who tried to capture certain linguistic styles in their work, particularly in American literature.
There are situations where you may want this class marker to be erased, because being associated with a certain social class can have negative impact on your social prospects. But it remains that something is being lost in the process, and that something is the personality and identity of the writer.
I find most people can write way better than AI, they simply don’t put in the effort.
Which is the real issue, we’re flooding channels not designed for such low effort submissions. AI slop is just SPAM in a different context.
You may be in a bubble of smart, educated people. Either way, one of the key ways to "put in the effort" is practice. People who haven't practiced often don't write well even if they're trying hard in the moment. Not even in terms of beautiful writing, just pure comprehensibility.
I may be in a bubble of smart people, but IMO AI consistently far worse than many high school works I’ve read in terms of actual substance and coherent structure.
Of course I’ve had arguments where people praise AI output then I’ve literally pointed out dozens of mistakes and they just kind of shrug saying it’s not important. So I acknowledge people judge writing very differently than I do. It just feels weird when I’d give something a 15% and someone else would happily slap on a B+.
My experience has been
(ordered from best to worst)
1. Author using AI well
2. Author not using AI
3. Author using AI poorly
With the gap between 1 and 2 being driven by the underlying quality of the writer and how well they use AI. A really good writer sees marginal improvements and a really poor one can see vast improvements.
I am really conflicted about this because yes, I think that an LLM can be an OK writing aid in utilitarian settings. It's probably not going to teach you to write better, but if the goal is just to communicate an idea, an LLM can usually help the average person express it more clearly.
But the critical point is that you need to stay in control. And a lot of people just delegate the entire process to an LLM: "here's a thought I had, write a blog post about it", "write a design doc for a system that does X", "write a book about how AI changed my life". And then they ship it and then outsource the process of making sense of the output and catching errors to others.
It also results in the creation of content that, frankly, shouldn't exist because it has no reason to exist. The number of online content that doesn't say anything at all has absolutely exploded in the past 2-3 years. Including a lot of LLM-generated think pieces about LLMs that grace the hallways of HN.
Even if they “stay in control and own the result”, it’s just tedious if all communication is in that same undifferentiated sanded-down language.
> A lot of people aren't, and the AI legitimately writes better.
It may write “objectively better”, but the very distinct feel of all AI generated prose makes it immediately recognizable as artificial and unbearable as a result.
I think it is also fairly similar to the kind of discourse a manager in pretty much any domain will produce.
He lacks (or lost thru disuse) technical expertise on the subject, so he uses more and more fuzzy words, leaky analogies, buzzwords.
This maybe why AI generated content has so much success among leaders and politicians.
Every group want to label some outgroup as naively benefiting from AI. For programmers, apparently it's the pointy haired bosses. For normies, it's the programmers.
Be careful of this kind of thinking, it's very satisfying but doesn't help you understand the world.
I think it’s essential to realize that AI is a tool for mainstream tasks like composing a standard email and not for the edges.
The edges are where interesting stuff happens. The boring part can be made more efficient. I don’t need to type boring emails, people who can’t articulate well will be elevated.
It’s the efficient popularization of the boring stuff. Not much else.
It contributes to making “standard” emails boring. I rather enjoy reading emails in each sender’s original voice. People who can’t articulate well aren’t elevated, instead they are perceived to be sending bland slop if they use LLMs to conceal that they can’t express themselves well.
Mediocrity as a Service
I liked mediocrity as a service better when it was fast food restaurants and music videos.
Bryan Cantrill referred to it as "normcore" on a podcast, and that's the perfect description.
> But it's the jagged edges, the unorthodox and surprising prickly bits, that tear open a hole in the inattention of your reader, that actually gets your ideas into their heads.
This brings to mind what I think is a great description of the process LLMs exert on prose: sanding.
It's an algorithmic trend towards the median, thus they are sanding down your words until they're a smooth average of their approximate neighbors.
I'm sure this can be corrected by AI companies.
The question is… why? What is the actual human benefit (not monetary).
If AI wrote and thought better by default then I wouldn't have to read the AI slop my co-workers send me.
Just let my work have a soul, please.
That is NOT possible.
Why not?
Because even though at work it looks like you’re tasked with creating use values, you’re only there as long as the use values you create can be exchanged in the market for a profit. So every humane drive to genuinely improve your work will clash with the external conditions of your existence within that setting. You’re not there to serve people, create beautiful things, solve problems, nu-uh. You’re there to keep capital flowing. It’s soulless.
Unless you work in the public sector, non-profit or charity.
To think that “non-profit” work is actually non-profit work is just to not have grasped the nature of labor. You have to ask yourself: Am I producing use values for the satisfaction of human needs or am I working on making sure the appropriation of value extraction from the production of use values continues happening?
In some very extreme cases, such as in the Red Cross or reformist organizations, your job looks very clear, direct, and “soulful”. You’re directly helping desperate people. But why have people gotten into that situation? What is the downstream effect of having you helping them. It’s profit. It’s always profit. You’re salvaging humanity for parts to be bought and sold again. It doesn’t make a dishonest work. It’s just equally soulless.
Eh, it's not __that__ simple.
It is, just don’t use a thing with no soul like ai if soul is what you’re after.
The point is that he may not using AI in any shape or form, Regardless, AI scrapes its work without explicit consent and then spits it back in "polished" soul free form.
Great comment. It really is that simple.
Something I noticed building multi-agent pipelines: the ablation compounds. Had a 4-step pipeline - summarize, expand, review, refine - and by step 3 everything had the same rhythm and vocabulary. Anchoring the original source text explicitly at each step helped, but only partially.
The more interesting cause I think: RLHF is the primary driver, not just the architecture. Fine-tuning is trained on human preference ratings where "clear," "safe," and "inoffensive" consistently win pairwise comparisons. That creates a training signal that literally penalizes distinctiveness - a model that says something surprising loses to one that says something expected. Successful RLHF concentrates probability mass toward the median preferred output, basically by definition.
Base models - before fine-tuning - are genuinely weirder. More likely to use unusual phrasing, make unexpected associative leaps, break register mid-paragraph. Semantic ablation isn't a side effect of the training process, it's the intended outcome of the objective.
Which makes the fix hard: you can't really prompt your way out of it once a model is heavily tuned. Temperature helps a little but the distribution is already skewed. Where we've gotten better results is routing "preserve the voice" tasks to less-tuned models, and saving the heavily RLHF'd models for structured extraction and classification where blandness is actually what you want.
I wonder if you can use lower quality models (or some other non-llm related process) to inject more "noise" into the text in between stages. Of course it wouldn't help retain uniqueness from the original source text, just add more in between.
I’m not convinced removing RLHF would really make the probabilities generator give us distributions that can diverge from the mean while remaining useful.
In other words, this might not a problem that can be overcome in LLMs alone.
The "AI voice" is everywhere now.
I see it on recent blog posts, on news articles, obituaries, YT channels. Sometimes mixed with voice impersonation of famous physicists like Feynman or Susskind.
I find it genuinely soul-crushing and even depressing, but I may be over sensitive to it as most readers don't seem to notice.
> The "AI voice" is everywhere now.
Maybe I'm going crazy but I can smell it in the OP as well.
Yeah the article smells extremely strongly of AI to me, but I've been told here before that that's just the register's house style, so I have no idea.
Yeah I started seeing it too, the article is just full of AI clues.
Literally the worst thing that happened to the internet after addictive doomscroll feeds and ads everywhere.
And, the worst part is noone will ever make a new internet because of the founder effect. We are basically in the worst timeline.
Maybe. Another potential, more positive, timeline is that semantically ablated content filling everyone’s feeds turns people off, and slowly kills the social feed paradigm.
That is doubtful. Most of the content on people's feed was basically converging to the same paradigm anyways. Think the Mr. Beast "I gave someone 1 million dollars to try a potato" content.
I find it extremely difficult to focus on any piece of writing the moment I see the patterns. Can’t tell if it’s an attitude problem I need to get over or if it’s just that all AI writing really is that bad.
I just close the tab. My reading backlog is too long as it is.
Yes, I get more and more visceral reactions to it. I'm reminded of JPEG artifacts - unnoticeable in 1993!
I like to consider all the different dimensions in which our breath stinks (metaphorically) and we just don't know it yet.
same. it is showing how many people are not trying to participate - just appear to. I want to read from and write for my peers, but it seems we are just awash with fakers
The internet is a post-truth space now that you can spin up a million different agents to push whatever narrative you choose.
I personally think “generative AI” is a misnomer. More I understand the mathematics behind machine learning more I am convinced that it should not be used to generate text, images or anything that is meant for people to consume, even if it is the most blandest of email. Sometimes you might get lucky, but most of the time you only get what the most boring person in the most boring cocktail party would say if forced to be creative with a gun pointed to his head. It can help in multitude of other ways, help human in the creative process itself, but generating anything even mildly creative by itself… I’ll pass.
Precisely. If companies would just focus on what it could be good at - deductive search, coding boilerplate with assistance, etc. then it would be a great tool. Instead you have dario, altman, and co. trying to pump stock and give us more spaghetti agents.
People want the real thing, not artificially flavored tokens.
I would rather read the prompt than the generative output, even if it’s just disjointed words and sentence fragments.
I hope "just the prompt please" becomes a common phrase
Regurgitative AI
Degenerative AI
> most of the time you only get what the most boring person in the most boring cocktail party would say
don't be mean, it's median AI à la mode
Bible Scholar and youtube guy Dan McClellan had an amazing "high entropy" phrase that slayed me a few days ago.
https://youtu.be/605MhQdS7NE?si=IKMNuSU1c1uaVCDB&t=730
He ended a critical commentary by suggesting that the author he was responding to should think more critically about the topic rather than repeating falsehoods because "they set off the tuning fork in the loins of your own dogmatism."
Yeah, AI could not come up with that phrase.
> Yeah, AI could not come up with that phrase.
Agreed.
"AI" would never say "loins" (too sexual)
"AI" would never say "dogmatism" (encroaches on the "AI" provider's own marketing scheme)
A sloppy mixed metaphor?
I'm learning to like 'em more, along with every other human idiosyncracy. Besides, it makes a kind of sense, the idea of some resonance occuring in one's gusset. Timber timbre. Flangent thrumming.
Tuning fork in loins just makes me think of that chess cheating scandal with a vibrating butt plug.
It just makes me think of that time I saw someone recovering from eye surgery and I had a visceral reaction.
I thought it was more creative than sloppy. Don't forget that many ordinary phrases were once jarring mixed imagery. To "wear your heart on your sleeve" was coined by Shakespeare; we still use it because it "stuck" due to its unorthodox phrasing.
If you like your prose to be anodyne, then maybe you like what AI produces.
I thought it was quite an effective metaphor!
Yes I noticed this as well. I was last writing up a landing page for our new studio. Emotion filled. Telling a story. I sent it through grok to improve it. It removed all of the character despite whatever prompt I gave. I'm not a great writer, but I think those rough edges are necessary to convey the soul of the concept. I think AI writing is better used for ideation and "what have I missed?" and then write out the changes yourself.
> I think AI writing is better used for ideation
It shocks me when proponents of AI writing for ideation aren't concerned with *Metaphoric Cleansing* and *Lexical Flattening* (to use two of the terms defined in the article)
Doesn't it concern you that the explanation of a concept by the AI may represent only a highly distorted caricature of the way that concept is actually understood by those who use it fluently?
Don't get me wrong, I think that LLMs are very useful as a sort of search engine for yet-unknown terms. But once you know *how* to talk about a concept (meaning you understand enough jargon to do traditional research), I find that I'm far better off tracking down books and human authored resources than I am trying to get the LLM to regurgitate its training data.
I've found LLMs to be terrible with ideation. I've been using GPT 5.x to come up with ideas and plot lines for a Dungeon World campaign I've been running.
I'm no fantasy author, and my prose leaves much to be desired. The stuff the LLM comes up with is so mind numbingly bland. I've given up on having it write descriptions of any characters or locations. I just use it for very general ideas and plot lines, and then come up with the rest of the details on the fly myself. The plot lines and ideas it comes up with are very generic and bland. I mainly do it just to save time, but I throw away 50% of the "ideas" because they make no sense or are really lame.
What i have found LLMs to be helpful with is writing up fun post-session recaps I share with the adventurers.
I recap in my own words what happened during the session, then have the LLM structure it into a "fun to read" narrative style. ChatGPT seems to prefer a Sanderson jokey tone, but I could probably tailor this.
Then I go through it, and tweak some of the boring / bland bits. The end result is really fun to read, and took 1/20th the time it would have taken me to write it all out myself. The LLM would have never been able to come up with the unique and fun story lines, but it is good at making an existing story have some narrative flare in a short amount of time.
That‘s also my experience. I use AI to help me generate the overall structure of a narrative. Apart from the hallucinations (e.g. June is not in spring), it‘s ok to spot inconsistencies, somewhat acceptable to brainstorm some ideas if you‘re new to a certain genre, but the prose it generates (talking about Opus 4.6) feels like an interpolation of all existing texts.
YES this hits the nail on something I've been trying to express for some time now. Semantic ablation: love it, going to use that a lot not now when arguing why someone's ChatGPT-washed email sucks.
Semantic ablation is also why I'm doubtful of everyone proclaiming that Opus 4 would be AGI if we just gave it the right agent harness and let all the agents run free on the web. In reality they would distill it to a meaningless homogeneous stew.
> Semantic ablation is also why I'm doubtful of everyone proclaiming that Opus 4 would be AGI if we just gave it the right agent harness and let all the agents run free on the web. In reality they would distill it to a meaningless homogeneous stew.
I'm so glad that you have given me the language to express this perspective.
The core mechanic described here is real. RLHF does optimize toward the mean, that is just what happens when you train on human preference ratings and raters consistently reward clear, inoffensive, "polished" output.
But the damage is not uniform. For code comments, API docs, commit messages: low-entropy output is often fine. The problem is people using LLMs for things that require a distinct voice and then wondering why the result sounds like everyone else on the internet.
The part nobody talks about: you can partially fight this if you know what you lost. Prompts like "preserve unusual word choices" or "do not normalize my rhetorical structure" help, but only if you have a strong enough baseline to catch the drift. Most people using AI for writing assistance do not have that baseline, which is why the ablation goes undetected. They see polished output and ship it.
The vast majority of people who write don't have a voice worth preserving. The rest can build out a voice document to make sure the AI doesn't strip it out.
For those who haven't seen it yet, this Wiki page also has what I think is very good advice about writing:
https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
While the page's purpose is to help editors detect AI contributions, you can also detect yourself doing these same things sometimes, and fix them.
This matches what I saw when I tried using AI as an editor for writing.
It wanted to replace all the little bits of me that were in there.
Great article and exactly why I use AI less and less. I basically find it to be rotting my brain towards the middle of the distribution. It's like all the nuance and critical thinking that actually goes into things gets stripped out.
Once a company perfects an agent that essentially performs condensed search and coding boilerplate making, that is probably where LLMs end for me. Perplexity and Claude are on the right track but not at all close.
The original title of the article is: "Why AI writing is so generic, boring, and dangerous"
Why was the title of of the link on HackerNews updated to remove the term "Dangerous"?
The term was in the link on HackerNews for the first hour or so that this post was live.
In recent months(?) I've more often noticed HN story titles changing over time.
I'm not sure what's driving this. It reminds me of SEO.
In this case, the edited title appears to be an attempt to neuter the article's political claim.
Race to the middle really sums up how I feel about AI.
The middle gets lower and lower with every passing day.
Well yeah, when you race to the middle, the middle becomes the top, and then you've got to go find the new middle.
Ive noticed that the subtle/nunance gets lost with every so-called improvement with the models.
Im in no way anti-LLMs as I have benefited from them, but I believe the issue that will arise is that their unpredictable nature means that they can only be used in narrowly defined contexts. Safety and trust are paramount. Would you use online banking if the balance on your account randomly changed and was not reproducible? No chance.
This does not achieve the ROI that investors of these model producers are thinking. The question is whether said investors can sell off their shares before it becomes more widely known.
> I believe the issue that will arise is that their unpredictable nature means that they can only be used in narrowly defined contexts. Safety and trust are paramount.
You put words to something that's been on my mind for a while!
People felt this was already happening. I remenber reading filter world which described this pre Ai
I call it the great blur.
I call it a mirage. I get why people are taken aback and fascinated by it. But what the model producers are chasing is a mirage. I wonder when they'll finally accept it?
I think the LLM providers are selling the ability to create a mirage.
LLMs are a tool for marketers or state departments who want to be create FUD on a moment's notice.
The obvious truth is that LLMs basically suck for writing code.
The real marketing scheme is the ability to silence and stifle that obvious truth.
To me LLMs are an experiment toward replication of what humans can do. However, they fall short on many dimensions that its just not going to pan out from what I see.
The real danger is the future investment needed to explore other architectures beyond LLMs. Will private firms be able to get the investment? Will public firms be granted the permission to do another round of large capex by investors? As time goes on, Apple's conservative approach means they will be the only firm trusted with its cash balance. They are very nicely seated despite all the furore they've had to endure.
This isn't new to AI. The same kind of thing happens in movie test screenings, or with autotune. If something is intended for a large audience, there's always an incentive to remove the weird stuff.
All these forced metaphors and clumsy linguistic flourishes made me cringe. Just add some typos and grammar mistakes like the rest of us to prove that your human.
The part about a change in entropy was interesting.
Is there an easy way to get / compare the entropy of two passages? (e.g. to see if it has indeed dropped after gen ai manipulation).
And could this be used to flag AI-gen text (or at least, boring, soulless sounding text)
A lot of times, this entropy decay is found in semantic or stylistic space, which would be hard to detect (you couldn't use, e.g., Shannon Entropy). You'd have to ask questions like "is this point uninteresting?" or "is this trope overused?"--bad (human) writers are often guilty of this too, so that's why AI can be hard to detect.
I wonder how much of it could be prompted away.
For example the anthropic Frontend Design skill instructs:
"Typography: Choose fonts that are beautiful, unique, and interesting. Avoid generic fonts like Arial and Inter; opt instead for distinctive choices that elevate the frontend's aesthetics; unexpected, characterful font choices. Pair a distinctive display font with a refined body font."
Or
"NEVER use generic AI-generated aesthetics like overused font families (Inter, Roboto, Arial, system fonts), cliched color schemes (particularly purple gradients on white backgrounds), predictable layouts and component patterns, and cookie-cutter design that lacks context-specific character." 1
Maybe sth similar would be possible for writing nuances.
1 https://github.com/anthropics/skills/blob/main/skills/fronte...
> "NEVER use generic AI-generated aesthetics like overused font families (Inter, Roboto, Arial, system fonts), cliched color schemes (particularly purple gradients on white backgrounds), ...
Now, imagine what happens when this prompt becomes popular?
Keep in mind that LLMs are trying to predict the most likely token. If your prompt prohibits the most likely token, they output the next most likely token. So, attempts to force creativity by prohibiting cliches just create another cliche.
Several days ago, someone researched Moltbook and pointed out how similar all the posts are. Something like 10% of them say "my human", etc.
Many have tried, it does not work. Regression to the mean always sets in.
This article on AI writing being boring seems to be written by AI. The em dashes and the sentence structure, all seems to be AI output. Or have human started adopting this style too.
Isn't this more to do with how LLMs are trained for general purpose use? Are LLMs with a specific use and dataset in mind better? Like if the dataset was fiction novels, would it sound more booky? If it was social-media, would it sound more click-baity and engaging?
I've had AI be boring, but I've also seen things like original jokes that were legitimately funny. Maybe it's the prompts people use, it doesn't give it enough of a semantic and dialectic direction to not be generic. IRL, we look at a person and get a feel for them and the situation to determine those things.
What a weird use of "Romanesque" and "Baroque". Doesn't compute for me at all.
I wonder why AI labs have not worked on improving the quality of the text outputs. Is this as the author claims a property of the LLMs themselves? Or is there simply not much incentive to create the best writing LLM?
The argument is that the best writing is the unexpected, while an LLM's function is to deliver the expected next token.
Even more precisely, human writing contains unpredictability that is either more or less intention (what might be called authors intent), as well as much more subconsciously added (what we might call quirks or imprinted behavior).
The first requires intention, something that as far as we know, LLMs simply cannot truly have or express. The second is something that can be approximated. Perhaps very well, but a mass of people using the same models with the same approximationa still lead to loss of distinction.
Perhaps LLMs that were fully individually trained could sufficiently replicate a person's quirks (I dunno), but that's hardly a scalable process.
Yeah, that makes banana.
What was the name of the last book you read?
I remember an article a few weeks back[1] which mentioned the current focus is improving the technical abilities of LLMs. I can imagine many (if not most) of their current subscribers are paying for the technical ability as opposed to creative writing.
This also reminded me that on OpenRouter, you can sort models by category. The ones tagged "Roleplay" and "Marketing" are probably going to have better writing compared to models like Opus 4 or ChatGPT 5.2.
[1]: https://www.techradar.com/ai-platforms-assistants/sam-altman...
I mean there's tons of better-writing tools that use AI like Grammarly etc. For actual general-purpose LLMs, I don't think there's much incentive in making it write "better" in the artistic sense of the world... if the idea is to make the model good at tasks in general and communicate via language, that language should sound generic and boring. If it's too artistic or poetic or novel-like, the communication would appear a bit unhinged.
"Update the dependencies in this repo"
"Of course, I will. It will be an honor, and may I say, a beautiful privilege for me to do so. Oh how I wonder if..." vrs "Okay, I'll be updating dependencies..."
I wish it would just say "k, updated xyz to 1.2.3 in Cargo.toml" instead of the entire pages it likes to output. I don't want to read all of that!
I used to feel the same but you can just prompt it to reply with only one word when its done. Most people prefer it to summarize because its easier to track so ig thats the natural default
I mean, no one is asking for artistic writing, just not some obvious AI slop. The fact that we all can now easily determine that some text has been written / edited by AI is already an issue. No amount of prompting can help.
Yeah but thats not what I am saying. I am saying its default writing style is for communicating with the user, not producing content/text hence it has that distinctive style we all recognise. If you want AI writing thats not slop, there are tools that are trying to do that but the default LLM writing style is unlikely to change imo.
That's like asking why McDonald's doesn't improve the quality of their hamburger. They can, but only within the bounds of mass produced cheap crap that maximizes profit. Otherwise they'd be a fundamentally different kind of company.
> The AI identifies unconventional metaphors or visceral imagery as "noise" because they deviate from the training set's mean.
That's certainly a take. In the translation industry (the primogenitor and driver for much of the architecture and theory of LLMs) they're known for making extremely unconventional choices to such a degree that it actively degrades the quality of translation.
Could we invert a sign somewhere and get the opposite effect?
(Obviously a different question from "is an AI lab willing to release that publicly” ;)
It's a hard problem and so far not a profitable one (I hope the solution will emerge as a byproduct of another innovation)
https://nostalgebraist.tumblr.com/post/778041178124926976/hy...
https://nostalgebraist.tumblr.com/post/792464928029163520/th...
Couldn't you simply increase the temperature of the model to somewhat mitigate this effect?
I kind of think of that as just increasing the standard deviation. Its been a while since I experimented with this, but I remember trying a temp of 1 and the output was gibberish, like base64 gibberish. So something like 0.5 doesn't necessarily seem to solve this problem, it just flattens the distribution and makes the output less coherent, with rarer tokens, but still the same underlying distribution.
you have to know that your "simply" is carrying too much weight. here's some examples of why just temperature is not enough, you need to run active world models https://www.latent.space/p/adversarial-reasoning
When applied to insightful writing, that is much more likely to dull the point rather than preserve or sharpen it.
I'd like to see some concrete examples that illustrate this - as it stands this feels like an opinion piece that doesn't attempt to back up its claims.
(Not necessarily disagreeing with those claims, but I'd like to see a more robust exploration of them.)
Have you not seen it any time you put any substantial bit of your own writing through an LLM, for advice?
I disagree pretty strongly with most of what an LLM suggests by way of rewriting. They're absolutely appalling writers. If you're looking for something beyond corporate safespeak or stylistic pastiche, they drain the blood out of everything.
The skin of their prose lacks the luminous translucency, the subsurface scattering, that separates the dead from the living.
The prompt I use for proof-reading has worked great for me so far:
> If you're looking for something beyond corporate safespeak
AI has been great for removing this stress. "Tell Joe no f'n way" in a professional tone and I can move on with my day.
If you tell me "no fucking way" by running it through an LLM, I will be far more pissed than if you had just sent me "no fucking way". At least in that case I know a human read and responded rather than thinking my email was just being processed by a damned robot.
Yeah but does it make sense to have invested all this money for this?
Lol no. Might be great for you as a consumer who is using these products for free. But expand the picture more.
> Yeah but does it make sense to have invested all this money for this?
No, but it's here. Why wouldn't I use it?
> If you're looking for something beyond corporate safespeak or stylistic pastiche, they drain the blood out of everything.
Strong agree, which is why I disagree with this OP point:
“Stage 2: Lexical flattening. Domain-specific jargon and high-precision technical terms are sacrificed for "accessibility." The model performs a statistical substitution, replacing a 1-of-10,000 token with a 1-of-100 synonym, effectively diluting the semantic density and specific gravity of the argument.”
I see enough jargon in everyday business email that in the office zero-shot LLM unspoolings can feel refreshing.
I have "avoid jargon and buzzwords" as one of very tiny tuners in my LLM prefs. I've found LLMs can shed corporate safespeak, or even add a touch of sparkle back to a corporate memo.
Otherwise very bright writers have been "polished" to remove all interestingness by pre-LLM corporate homogenization. Give them a prompt to yell at them for using 1-in-10 words instead of 1-in-10,000 "perplexity" and they can tune themselves back to conveying more with the same word count. Results… scintillate.
Look through my comment history at all the posts where I complain the author might have had something interesting to say but it's been erased by the LLM and you can no longer tell what the author cared about because the entire post is a an oversold monotone advertising voice.
https://news.ycombinator.com/item?id=46583410#46584336
https://news.ycombinator.com/item?id=46605716#46609480
https://news.ycombinator.com/item?id=46617456#46619136
https://news.ycombinator.com/item?id=46658345#46662218
https://news.ycombinator.com/item?id=46630869#46663276
https://news.ycombinator.com/item?id=46656759#46663322
https://news.ycombinator.com/item?id=46661936#46663362
https://news.ycombinator.com/item?id=46748077#46749699
I just sent TFA to a colleague of mine who was experimenting with llm's for auto-correcting human-written text, since she noticed the same phenomenon where it would correct not only mistakes, but slightly nudge words towards more common synonyms. It would often lose important nuances, so "shun" would be corrected to "avoid", and "divulge" would become "disclose" etc.
Kaffee: Corporal, would you turn to the page in this book that says where the mess hall is, please?
Cpl. Barnes: Well, Lt. Kaffee, that's not in the book, sir.
Kaffee: You mean to say in all your time at Gitmo, you've never had a meal?
Cpl. Barnes: No, sir. Three squares a day, sir.
Kaffee: I don't understand. How did you know where the mess hall was if it's not in this book?
Cpl. Barnes: Well, I guess I just followed the crowd at chow time, sir.
Kaffee: No more questions.
It is an opinion piece. By a dude working as a "Professor of Pharmaceutical Technology and Biomaterials at the University of Ferrara".
It has all the tropes of not understanding the underlying mechanisms, but repeating the common tropes. Quite ironic, considering what the author's intended "message" is. Jpeg -> jpeg -> jpeg bad. So llm -> llm -> llm must be bad, right?
It reminds me of the media reception of that paper on model collapse. "Training on llm generated data leads to collapse". That was in 23 or 24? Yet we're not seeing any collapse, despite models being trained mainly on synthetic data for the past 2 years. That's not how any of it works. Yet everyone has an opinion on how bad it works. Jesus.
It's insane how these kinds of opinion pieces get so upvoted here, while worth-while research, cool positive examples and so on linger in new with one or two upvotes. This has ceased to be a technical subject, and has moved to muh identity.
Yeah, reading the other comments on this thread this is a classic example of that Hacker News (and online forums in general) thing where people jump on the chance to talk about a topic driven purely by the headline without engaging with the actual content.
(I'm frequently guilty of that too.)
Even if that isn't the case, isn't it the fact the AI labs don't want their models to be edgy in any creative way, choose a middle way (buddhism) so to speak. Are there AI labs who are training their models to be maximally creative?
> Yet we're not seeing any collapse, despite models being trained mainly on synthetic data for the past 2 years.
Maybe because researchers learned from the paper to avoid the collapse? Just awareness alone often helps to sidestep a problem.
No one did what the paper actually proposed. It was a nothing burger in the industry. Yet it was insanely popular on social media.
Same with the "llms don't reason" from "Apple" (two interns working at Apple, but anyway). The media went nuts over it, even though it was littered with implementation mistakes and not worth the paper it was(n't) printed on.
Who cares? This is a place where you should be putting forth your own perspective based on your own experience. Not parotting what someone else already wrote.
I think they can fix all that but they can't fix the fact that the computer has no intention to communicate. They could imbue it with agency to fix that too, but I much prefer it the way things are.
Those transformations happen to mirror what happens to human intelligence when you take antipsychotics. Please know the risks before taking them. They are innumerable and generally irreversible.
How much money would it take for me to take an open weight model, treat it nice, and go have some fun? Maybe some thousands, right?
As a writer who has been published many times and edited many other writers for publication... It seems like AI can't make stylistic determinations. It is generally good with spelling and grammar but the text it generates is very homogeneous across formats. It's readable but it's not good, and always full of fluff like an online recepie harvesting clicks. It's kind of crap really. If you just need filler it's ok, but if you want something pleasand you definitely still need a human.
> What began as a jagged, precise Romanesque structure of stone is eroded into a polished, Baroque plastic shell
Not to detract from the overall message, but I think the author doesn't really understand Romanesque and Baroque.
(as an aside, I'd most likely associate Post-Modernism as an architectural style with the output of LLMs - bland, regurgitative, and somewhat incongruous)
The article itself reads as an AI generated output, complete with classic Not Just X … Y hallmarks from forever ago, 100% on pangram's low false positive detector. I'm not sure if it's some experiment on their readerbase or what. pangram result: https://www.pangram.com/history/02bead1c-c36e-461b-8fa7-8699...
So many AI generated AI bashing articles lately. I wrote a post complaining about running into these, and asking people who've sent me these AI articles multiple of them came from HN. https://lunnova.dev/articles/ai-bashing-ai-slop/
Sematic ablation... that's some technobable.
Going off search results, it seems to be a new coinage. I found mostly references to TFA, along with an (ironically obviously AI-written) guide with suggestions for getting LLMs to avoid the issue (just generic "traditional" advice for tuning their output, really). The guide was apparently published today, and I imagine that it's a deliberate response to TFA. But FWIW the term "semantic ablation" does seem to me like something that newer models could invent
At any rate, it seems to me like a reasonable label for what's described:
> 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).
> ...
> When an author uses AI for "polishing" a draft, they are not seeing improvement; they are witnessing semantic ablation.
The metaphor is very apt. Literal polishing is removal of outer layers. Compared to the near-synonym "erosion", "ablation" connotes a deliberate act (ordinarily I would say "conscious", but we are talking about LLMs here). Often, that which is removed is the nuance of near-synonyms (there is no pause to consider whether the author intended that nuance). I don't know if the "character" imparted by broader grammatical or structural choices can be called "semantic", but that also seems like a big part of what goes missing in the "LLM house style".
Bluntly: getting AI to "improve" writing, as a fully generic instruction, is naturally going to pull that writing towards how the AI writes by default. Because of course the AI's model of "writing quality" considers that style to be "the best"; that's why it uses it. (Even "consider" feels like anthropomorphizing too much; I feel like I'm hitting the limits of English expressiveness here.)
Meh. Semantic Ablation - but toward a directed goal. If I say "How would Hemingway have said this, provided he had the same mindset he did post-war while writing for Collier's?"
Then the model will look for clusters that don't fit what the model consider's to be Hemingway/Colliers/Post-War and suggest in that fashion.
"edit this" -> blah
"imagine Tom Wolfe took a bunch of cocaine and was getting paid by the word to publish this after his first night with Aline Bernstein" -> probably less blah
These kinds of prompts don’t really improve the writing IME. It still gets riddled with the same tropes and phrases, or it veers off into textual vomit.
FWIW, I agree. Frontier LLMs are on their way to becoming competent stylists (I ask every major model release to write up a sample essay as Hemingway, and they are improving), but they are often skin-deep.
Even if it would work good luck writing with a new style.
See also https://en.wikipedia.org/wiki/Regression_toward_the_mean
> 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).
> Domain-specific jargon and high-precision technical terms are sacrificed for "accessibility." The model performs a statistical substitution, replacing a 1-of-10,000 token with a 1-of-100 synonym, effectively diluting the semantic density and specific gravity of the argument.
> The logical flow – originally built on complex, non-linear reasoning – is forced into a predictable, low-perplexity template. Subtext and nuance are ablated to ensure the output satisfies a "standardized" readability score, leaving behind a syntactically perfect but intellectually void shell.
What a fantastic description of the mechanisms by LLMs erase and distort intelligence!
I agree that AI writing is generic, boring and dangerous. Further, I only think someone could feel this way if they don't have a genuine appreciation for writing.
I feel strongly that LLMs are positioned as an anti-literate technology, currently weaponized by imbeciles who have not and will never know the joy of language, and who intend to extinguish that joy for any of those around them who can still perceive it.
People haven't really spoken about the obvious token manipulation that will be on the horizon once any model producer has some semblance of lock-in.
If you thought Google's degredation of search quality was strategic manipulation, wait till you see what they do with tokens.
the word choice here is so obtuse as to trigger my radar for "is this some kind of parody where this itself was AI generated". it appears to be entirely serious, which is disappointing, it could have been high art.
the words TFA is looking for is mode collapse https://www.lesswrong.com/posts/t9svvNPNmFf5Qa3TA/mysteries-... and the author could herself learn to write more clearly.
It almost certainly is AI generated. It reads like it is, pangram thinks it is, https://www.pangram.com/history/02bead1c-c36e-461b-8fa7-8699... and pangram's unlikely to give false positives https://www.pangram.com/blog/third-party-pangram-evals
Do the terms *Metaphoric Cleansing*, *Lexical Flattening*, and *Structural Collapse* that the author provides have equivalents in LessWrong's parlance?
not to my knowledge but those i have no problem with. its the overly complex prose surrounding them that shows this author overcorrects to SAT test words in an attempt to prove her superiority over ai writing and that has its own distastefulness.
Because you simply can't engineer creativity. Maybe you can describe where it comes from, in a circuitous, abstract way with mathematics (and ultimately run face first into ħ and then run in circles for eternity). But to engineer it, you'd have to start over from the first principles of the stuff of the cosmos. One's a map and the other the territory.
As someone longtime involved in software development, can we call this "best practices" instead of some like "semantic ablation" that nobody understands?
I think you might be missing the point of the article.
I agree that the term "semantic ablation" is difficult to interpret
But the article describes three mechanisms by which LLMs consistently erase and distort information (Metaphoric Cleansing, Lexical Flattening, and Structural Collapse)
The article does not describe best practices; it's a critique of LLM technology and an analysis of the issues that result from using this technology to generate text to be read by other people.
> The model performs a statistical substitution, replacing a 1-of-10,000 token with a 1-of-100 synonym
Do we see this in programming too? I don't think so? Unique, rarely used API methods aren't substituted the same way when refactoring. Perhaps that could give us a clue on how to fix that?
I think that's different because refactoring usually involves calling the same functions/methods albeit in a bit more readable way.
When not given a clear guideline to "just" refactor, I have had problems with LLMs hallucinating functions that don't exist.
Nonsense. I’ve written bland prose for a story and AI made it much better by revising it with a prompt such as this: “Make the vocabulary and grammar more sophisticated and add in interesting metaphors. Rewrite it in the style of a successful literary author.”
Etc.
Have you considered that your analysis skills may not be keen enough to detect generic or boring prose?
Is it possible that what is a good result to you is a pity to someone with more developed taste?
I have a colleague that recently self-published a book. I can easily tell which parts were LLM driven and which parts represent his own voice. Just like you can tell who's in the next stall in the bathroom at work after hearing just a grunt and a fart. And THAT is a sentence an LLM would not write.
> And THAT is a sentence an LLM would not write.
Really?
Here's some alternatives. Some are clunky. But, some aren't.
…just like you can tell whose pubes those are on the shared bar of soap without launching a formal investigation.
…just like you can tell who just wanked in the shared bathroom by the specific guilt radiating off them when they finally emerge.
…just like you can tell which of your mates just shitted at the pub by who's suddenly walking like they're auditioning for a period drama.
…just like you can tell which coworker just had a wank on their lunch break by the post-nut serenity that no amount of hand-washing can disguise.
…just like you can tell whose sneeze left that slug trail on the conference room table by the specific way they're not making eye contact with it.
…just like you can identify which flatmate's cum sock you've accidentally stepped on by the vintage of the crunch.
…just like you can tell who just crop-dusted the elevator by the studied intensity with which one person is suddenly reading the inspection certificate.
IMO The LLM you're using has failed to mimic the tone of OP's bathroom joke.
These alternatives are uncomfortably crude. They largely make gross reference to excretory acts or human waste. The original comment was off color, but it didn't go beyond a vague discussion of a shared human experience.
It's still on you to pick what the LLMs regurgitate. If you don't have a style or taste you will simply make choices that would give you away. And if you already have your own taste and style LLMs don't have much to offer in this regard.
The great promise and the great disaster of LLMs is that for any topic on which we are "below average", the bland, average output seems to be a great improvement.
Counter intuitively... this is a disaster.
We dont need more average stuff - below average output serves as a proxy for one to direct their resources towards producing output of higher-value.
So what even if that is true? You confirmed that it improved upon what he could manually produce, which is still a win. It doesn't always make sense to pay $20000 to a professional author to turn it into a masterpiece.
My point is simply that the tell-tale marks of LLM prose can be remediated through prompts.
I have a very large ‘default prompt’ that explicitly deals with the more obnoxious grammatical structures emblematic of LLMs.
I would wager I deal with more amateurishly created AI slop on a doily basis than you do. (Legal field, where everyone is churning out LLM-written briefs.) Most of it is instantly recognizable. And, all of it can be fixed with more careful prompt-engineering.
If you think you can spot well-crafted LLM prose generated by someone proficient at the craft of prompt-engineering by, to use an analogy to the early days of image creation, counting how many fingers the hand has, you’re way behind.
Why don't you post it so we can see how much better the AI made it?
Because HN isn't a literary forum.
Maybe it sucks. Maybe it doesn't.
But, I notice a curious pretentiousness when it comes to some people's assumptions about their ability to identify LLM prose. Obviously, the generic first-pass 'chat' crap is recognizable; the kind of garbage that is filling up blog-posts on the internet.
But, one shouldn't underestimate the power of this technology when it comes to language. Hell, the 'coding' skills were just a pleasant side-effect of the language training, if you recall. These things have been trained on millions of works of prose of all styles: its their heart and soul. If you think the superficial monotonous style is all there is, you're mistaken. Most of the obnoxious LLM-style stuff is an artifact of the conversational training with Kenyans and the like in the early days. But, you can easily break through that with better prompts (or fine-tuning it yourself.)
That said, one shouldn't conflate the creation of the content and structure and substance of a work of prose with the manner in which it is written. You're not going to get an LLM to come up with a decent plot... yet. But, as far as fleshing out the framework of a story in a synthetic 'voice' that sounds human? Definitely doable.