Indeed. I have been experimenting with using frontier models for creative writing.

Using adversarial models to evolve ideas. I have a whole novel of material, and the feedback has indeed been useful. But the actual novel? Not even AI models like their own writing once you give the simple prompt of "account for model sycophancy".

The resulting structure and plot beats are good with AI as copy editor. Only the least discerning reader might not care or notice the actual non-human writing. And hey - maybe those are your readers. Probably not.

Can we all appreciate for a moment the sci-fi dystopia which snuck up on us all? I just used to phrase: " non-human writing". 50 years ago, that would send a chill up your spine.

> Readers can detect LLM words in the parts per trillion.

Oh, we do. We can also detect hackeneyed, LLM-made sentences where just one word had been replaced. "That's where it's going to f*** you".

I think style is overrated. Yeah if you write commercially, you are a journalist or a copywriter, sure, you need to write in certain style. There’s rules to this shit. I think it was Hemingway who said that. Then again, the good ones break the rules.

Today even if you don’t use LLM’s following certain formal style you end up sounding like ad copy or wannabe Atlantic -hack. The. You might read some kid on Tumblr writing beautiful prose with typos and errors and all.

First of all you should have something to say. I write a blog just to make tutorials for myself on stuff I need to regularly check out. I don’t really care if anyone reads it. I don’t use LLM to make the posts since, well, pride, but also because I learn better when I need to write the stuff down and check my facts several times. I don’t much care if I make a typo or grammar error here and there. If I want to write it as Jeeves from Wodehouse book, fuck it, I will.

If you have a story you have something to write about. It doesn’t have to be fiction, it can be a story on how multicast works. And it needs to be your story. Then the words come by themselves and writing is relatively easy. To have a story you have to live and think. LLM’s don’t help in that.

I would have to be an utter moron to use LLM to write anything directed at other people.

just don't use it for writing

Simply do not.

Ironically, "load-bearing" appears in the main text of the article that says never to use a single word the LLM suggests. Womp womp.

This is why, if you have the agent generate the commit message at all, it has to be the same agent that wrote the code, in the same session.

I usually have it lead with a paragraph or two of context on why the change was made (though I often use it as just a tool to turn my rambling on that question into proper English), followed by a couple paragraphs of "abstract" explaining the change (because an LLM-generated summary reviewed by the person who made the commit is better than something which the person reading can get themselves).

I've been reading Steven Pinker's "The Sense of Style". Unlike "The Elements of Style", which is fine, it does a good job anchoring its suggestion in first principals. I do think I need to read a book on grammar to truly appreciate it though.

"On Writing" is another book like "The Elements of Style", both are too prescriptive, albeit not useless.

Yeah it's very short, but I can't remember if I read all of it, or just flipped through pieces. It has good advice and I think fairly "famous." (edit: this reminded me I still have my copy :-D)

I know someone in a profession that does a lot of writing, and it blew my mind how clearly feedback was communicated by a superior. Made me wish tech people had better written communication skills. ;-;

Can't vouch for that but I can vouch for `Dreyer's English` as both hilarious (if you're the sort of person who relishes finding le mot juste, like me) and excellent as a style guide.

Paperback copies in EU seem to be very expensive (90+€ on amazone.de) unfortunately.

Clear and simple as the truth - another good one

Ordinary people were writing before the option existed to accept words from AI models.

"Naturally gifted writer" is not a thing. The way to get better at writing is to write a lot.

Developing a voice comes naturally as you go through the pain of being a terrible writer to eventually being mediocre. Accepting the LLM's words is actively harmful to that process because it lets you skip over the part where you read over what you wrote and rewrite until it doesn't suck as much.

I mean, you do you. I'm not your dad.

iA Writer's "Style check" feature can do 90% of this in real time.

How to write with an LLM?

Don’t

I was seriously reading this article right up until the word "load-bearing", and then despite my best efforts, my brain just sort of disconnected from the piece. I just skimmed to the end quickly before realizing what happened.

"Those rethinks are load-bearing parts of your voice."

Loved the (possibly not deliberately ironic) use of the LLM'S favourite analogy, load-bearing.

Nice wee article though. I particularly agree with "Don’t take all of the model’s copyediting advice." I've been using LLMs to review teaching material and often it just doesn't get the point of teaching. It's good for prompting reflection on your writing but it's still an idiot savant.

Don't. There, easier

Just like you do everything else with an LLM. You ask it for something, if that doesn’t work you just ask it in natural language to change it until it outputs exactly what you want.

“Here’s how this is going to fuck you“ - this article was either written by a human, or the most aggressively tuned llm.

It's more effective in my experience to have an LLM emulate your own writing examples than to apply lists of rules in the hope it sounds human.

Why are people not learning to write at school? In most developed countries don't we spend like a whole decade or more learning to do stuff like writing? Yet people still feel like they can't write?

Writing is something we really, really need to learn from other people, because it's with people we are trying to communicate. There should be no need for "writing with an LLM". None at all. If what you are writing is important, ask a friend to proof read it. If you can't write, learn.

This article should just be one word:

Don't.

It's darkly amusing that while the author advises us not to let LLMs pick our words, their post includes this chunk of AI slop:

"Those rethinks are load-bearing parts of your voice. Readers won’t put their fingers on what’s wrong, but they’ll sense that you’ve become artificially-flavored."

The metaphor at the start is AI-like weirdness too: "without pasteurizing and jacking it with corn syrup." Pasteurising is good, corn syrup is bad - what the heck am I supposed to make of it? It's nonsense.

It may even have been handwritten by the author, but it's slop nonetheless.

You may be telling yourself that you're disciplined and you're just using the LLMs for review. But while you're reading their output, the slop makes its way directly into your brain anyway - and then it drips out into your writing.

How I write with an LLM: for each page “suggest up to 10 word changes to increase clarity.”

> How To Write With An LLM

Simply don't.

> How to Write with an LLM

TLDR: Don’t.

Probably the worst part

PISA is a statistical measurement tool designed for sensitivity to differences in student performance. So it uses questions that some students are expected to get right and some are expected to get wrong, and places the result on a bell curve fit to a reference population.

It's not clear what you mean by "sufficient reading literacy," but I guess it's one of the PISA proficiency levels. Those are simply equally-spaced intervals on the bell curve. For example, level 1c in reading is 2.4 to 3.1 standard deviations below the reference mean and 21% of participants in Iraqi Kurdistan and Lebanon scored at that level. Relative to the baseline, they are definitely close to the bottom. However, even they must've gotten some questions right, or they would've ended up in the "below level 1c" bucket (like 3% in Iraqi Kurdistan and 5% in Lebanon did). (https://www.oecd.org/en/publications/pisa-2025-results-volum... Table I.B1.2a.11)

If you wanted to know how many people achieve "sufficient reading literacy," you would need a test specifically designed for measuring it (i.e. not PISA), where people with sufficient ability can be expected to get all questions right. For example, the literacy rate is typically defined via the ability to "read and write a simple statement about everyday life," a test which 99% of adults in wealthy countries pass, 92% in Lebanon and 84% in Iraq: https://ourworldindata.org/grapher/cross-country-literacy-ra...

Of course a test that 99% pass just becomes table stakes and boring, whereas PISA with its large differences between individuals is much more exciting, but you shouldn't let that mislead you into thinking that students scoring relatively low on PISA necessarily must be lacking even those table-stakes skills.

> LLM can produce texts but we as people will not be able to understand anyway

LLM generated text oftentimes just do not have the point to be understood. LLM is training you to NOT read, because trying to do close reading of LLM generated text is usually a mistake. There is no deeper there. Either it was correct on the surface things or it is a gibberish.

I don't think that many teams review commit messages as a part of code review. At least, I've never seen it "in the wild".

Not all workplaces are high functioning in the way you imagine. Until recently a shocking amount of software development did not use source control.

LLMs can provide the intentions of the code they write. To say they can’t is saying that they can’t fix bugs, which they very obviously can do effectively. So, just tell them to write the intention in the way you want it. If you are implicitly providing the direction in the prompt, make that explicit - you then it will have the context for the change while working and for the commit message and you’ll likely get a higher quality output.

Exactly! If I want a summary of the code, I can easily generate it myself.

Fix your harness. Provide rules, format, and good examples.

You probably had those for the humans in your team for when new hurts join, right? Just direct your LLMs to them and you’ll have a much better experience.

Just like you can’t expect a new junior dev to know what to do without direction, the same applies with LLMs.

You don't blind trust them in the fact checking. You check what they point out.

I use Fable to give me feedback on things I write. But I discard about 50%, because it often does not have the context and thus tends to favour hedging stuff. Also, it does not always vibe with my writing style.

But the 50% I do take into account, improves the text! And, like TFA, I never ask it for concrete text. It only helps me diagnose the issues, I prescribe the medicine!

To paraphrase, "if a junior programmer doesn't understand what your code does, it's bad". If Fable doesn't understand, well...

You are erasing yourself in doing this, not improving your writing.

If you want to improve your writing, read more.

Given that he says "never use a LLM suggested word", and given he's a sophisticated writer, it's clearly a joke.

Pangram 4 says it's 100% human written, including the "load-bearing" sentence.

I figured that's what you meant and agree the phrasing is over-speak.

To your comment, I wasn't aware of the eye rolling meta for the phrase [0]. I hadn't seen it overused to the extent people are claiming.

[0] https://www.reddit.com/r/ClaudeAI/comments/1upxk9r/loadbeari...

I think with the desert thing, you are getting at what was once described as "aura" of original art work. Things like photography have already made us question things like synthetic art and how it relates to "real art" and I am pretty sure this is how it will go with AI and art in general. I wrote about this in my full human only blog: https://decodingvibes.com/blog/3-hypothetical-paintings-and-...

In this comment section alone, 5 people have used some variant of that aphorism, and it's used constantly on hn and other sites. They've heard it.

How do you know it is? Most likely we will never know.

How about Anthropic? I've seen people say it's surprising that their blog isn't written with AI, but I think it's far more likely that it is written with AI; they are just good at reviewing and getting a nice prose out of their models.

Yeah, it's phrased like a reductio ad absurdum but it's close to being literally true. It's not technically "copywriters" but their words are "polished" for pleasantness, often with the explicit guidance of consultants, until most of the truth or novelty is gone.

Not sure. I don't like LLMs inserting itself into human to human communication, as it can shape our thinking in many ways. Delegating this also makes it much harder for a person to become a subject-matter expert, which is a real loss.

> What’s preventing the editor from using one of your prompts/tooling and replacing you entirely?

Well, they don't have the many other pieces of the business in place that makes this work meaningful. They certainly could try, though. If they succeeded, this aspect wouldn't be what made the difference, though.

Yeah ok good point. I think my thinking around this was too simplistic.

There is now increased evolutionary pressure for rules protecting scarcity to become norm. Communication media will likely adopt limits on the size and throughput of messages.

It kind of makes sense that if English is not their first language, the can't review the LLM English output also.

However most likely they just don't understand what good communication looks like, LLMs are perfectly capable of turning anything into clear concise communication, in fact an easy tell that something is LLM written is that it's just too precise and clear, humans are much more rambling.

I think we could agree that everyone already does this up to a point. This is basically just an exploration-exploitation dilemma, where everyone tries to maximize how much value they gain from the things they read, and a lot of people have strategies to pick what material to read.

Reading trusted authors is likely to be valuable, but they may devalue over time. We don't know if our chosen authors maximize our value, nor whether exploratory reading does, etc.

https://en.wikipedia.org/wiki/Exploration%E2%80%93exploitati...

No it really wouldn't. I like the authors Jeff VanderMeer and Peter Watts. They have had multiple interviews where they talk about their peers, who they respect from the past, and who they like in our current age. I have found maybe subsequent authors + journalists to read from these two.

Try to think a little more critically here. If you can't trust the opinion of people you supposedly like, already established in their craft, who can you trust?

Have no idea why you think this would increase influencer culture, but thanks for telling us all publicly how you interact with social media: in an sloppy uncritical way.

My advice for you is to pull the plug and read books that are older than 125 years.

It’s hard to find good new literature - it may be out there, but it’s hard to find. But there’s lifetimes worth of good older literature, that’s much easier to find, that you haven’t read yet.

More than the average person, I suppose. For some reason, AI writing isn't as irritating to me as it seems to be for everyone else.

I treat it the same way I treat human writing: by speed reading most of the article so I can find the points the author is trying to make and decide whether I should read the entire thing.

I guess the answer to this is feeding the LLM generated content to another LLM and telling it to summarise it

I like the "strong opinions, loosely held" approach myself.

Right, I'm broadly not a fan of LLMs either. But I understand the use case here. If you can't get an actual human to look at your writing, an LLM used carefully could be better than nothing.

Having the style guide and proofreading it and linking it is more effort than the lazy slopulists want to put in, so your writing will reach far better than theirs will

hey thanks for mentioning that skill, I hadn't heard of it before but I just ran it on a non trivial codebase I've been planning to refactor and it did a great job.

I don't use a lot of "skills" so if there are others you've started using frequently I'd be curious to learn about them.

Sorry if this comes off as critical, but are you seriously doing that? I can't imagine optimizing my writing for people choosing to mangle it through a LLM rather than for the people reading it as-is.

Also the word "verbiage" ticks me off as a LLM-ism on the level of "delve" or "smoking gun". Are its users aware that in addition to its usage as somewhat of a synonym of "idiolect", it more commonly means "a profusion of words usually of little or obscure content", i.e. slop?

But I want everything and want to do no effort for it

You mean hallucinations?

Hallucinations don't really work that way. They happen when you ask an LLM to generate new material. Not so much reviewing existing material.

But there's always the "ask a second LLM to get two opinions" technique which gets you pretty close to error free.

And... now that I think of it... even if it was terrible and missed 50% of the errors... wouldn't that still be better than using no LLM which catches 0% of the errors?

Actually I am in a very good position to tell. And who said “write for you”? How about “write with you as a collaborator?”.

Writing is a form of thinking and writing with an LLM makes the argumentation, discussion, and progress toward goals faster and often better in my situation.

I am stating an opinion based on 2.5 years of intense interactions and writing with Anthropic models on a difficult topic related to asynchronous computational systems.

Background: I am a neuroscientist and behavioral biologist (45 years of experience) who understands the biology of behaviors fairly well but I need to translate this background into hardware architectures that get beyond current asynchronous system (e.g., this recent neuromorphic paper in Nature Communication by Li et al., 2025, A deterministic neuromorphic architecture, doi.org/10.1038.s41467-025-65268-z).

Working AND writing with Opus and now Fable has been a blessing of sorts.

If someone sends me AI slop, they're getting chewed out and told I'm not doing it and I'll even tell my boss "no" and why. If I got fired over something like that, then it tells me everything I need to know about company and the leadership's priorities. I'll die on that hill.

To be completely fair though, I'm against the behaviors in information transfer I've been seeing. I've pass along AI generated runbooks, but they look nothing like the default outputs of these models. It's because I took time to apply all the writing knowledge I like to see in my curation. If people are doing this, I can't even tell it's AI writing. My work is done in minutes instead of deciphering so BS pseudo language they developed in their AI workspace (people really need to turn off those memory features).

---

Edit: Also if I'm the guy receiving a security report and it's AI generated and poorly formatted, I'm failing you short of producing something for a human to parse. Simple as that.

Which is exactly what the post says, and why I say "do not accept a single word suggested by an LLM".

People keep writing about "style", as if I'm asking models "hey, how do you like this, and how could we make it more stylish". That's not at all what I'm asking.

I have personally decided, in advance, that one goal I have in writing is improved clarity. The model didn't tell me to do that; Joseph Williams did, decades ago.

In the service of that goal, I've written down a proofreading checklist. It includes things like "look for overused words", "eliminate throat-clearing filler like 'it's important to remember that'", "make sure every paragraph has a clear topic sentence", "make sure my real verbs aren't buried inside nouns". Stuff like that.

Those are stylistic choices, I suppose, but they're mostly the difference between bad style and a whole universe of good styles.

More importantly: they're technical. They're not really judgement calls. The subject of a sentence either is or is not a character taking actions in the story you're trying to tell. It is better, almost always, if the subject is a character, and not accidentally some random piece of the scenery of your point.

Models are very good at flagging these things. Humans are too, but they're not as thorough.

Many people on this thread have suggested that having a model flag these kinds of issues bleaches out personal style. I don't think that's true; what I think is happening is that they missed the point I'm making, and assumed that I was advocating for having a model make broad subjective calls about my writing.

Recent comments from the simonw account suggest it isn't operated directly by a human anymore /sigh/

> I'm not sure that Claude's "Realigned the shape of the load-bearing ownership gate to reduce the blast radius of the design contract; confirmed, not assumed" is more meaningful than "fix".

For PR/commit descriptions, I mainly use Claude Sonnet 4.5. It isn’t perfect, but it produces significantly less of this weird gibberish than 5.x models or even Opus 4.x do

I also use an iterative process in which it writes the description, I read it, and then either manually edit it or ask it to make changes

Could go either way, TBH.

So long as there are in fact those things, and so long as it didn't sneak something else in there at the same time, it being just on the knife-edge between sense and word-salad is better than "fix".

Buuuuuut far to often it says it fixed a bug I reported, when it only touched one superficial failure mode rather than the root cause.

Yesterday's issue: Why is zoom/pan randomly failing? It told me it was because it was applying a transformation matrix with every input and sometimes JavaScript gave it a non-invertible matrix (why?) which then propagated NaNs everywhere and you can't update a matrix filled with non-numbers.

Why was it doing that in the first place? Seems to be because it's too motivated to perform quick wins and not sufficiently motivated to do good engineering.

Good thing this was just a game editor. Spiky intelligence: superhuman on some dimensions, total noob on others.

It is, because it includes the key words such as "ownership gate" and perhaps others, which makes it infinitely more informative than just "fix", by pointing at what was in scope, and what wasn't.

I've never had Claude spew that out in a commit message. That being said, I also use my tools properly. Your comment sounds like the type of thing someone who puts "don't make any mistakes" at the end of a prompt and gets upset when it doesn't come out perfect. Like, someone adding "make it better than AAA" is going to do something. At this point, it's really just telling on yourself that this is the output that you get.

"fix stupid shit" is all you need :)

The long-ass Claude commit messages & PR descriptions suck. But they are 100% better than "fix".

Right, I agree: that’s exactly not what I’m recommending.

People have lessened what they read because they can more easily watch a video or listen to a podcast. It’s not like the 1800s where books were the only game in town. And why read an AI written article when you can listen to the AI generated podcast with banter between two fake personalities (yes, this is what I heard when I wanted info about qwen’s newest model on Apple podcast).

the author can ask the same question - why should I write something if you're just going to use a LLM to summarize it ?

This is an unnecessary argument both ways. Whether you use a LLM or write it yourself or read it yourself (and then comment on it), you're putting your signature beneath it. I don't care whether you use your natural brain or an artifical brain (LLM)

Wat, what am I paying for then?

hence adversarial review

> how can you say that he didn't think it himself?

Because the same sentence has been repeated practically word-for-word ad nauseam all over social media for years. What do you think is plausible - that they repeated something that is damn near omnipresent in the discourse, or that they never saw any of the many, many times it has come up and independently came up with the exact same formulation? They are perfectly able to step in and claim that it was their original thought if they want to.

Selecting between alternatives is also a very valuable skill. Also known as "taste".

You are a senior software architect, you "prompt" 10 junior architects to write a software architecture proposal and then pick one which you send the client.

Would you say it's worthless, because the senior architect didn't care to write it?

There are plenty of industries where "taste" is a real job.

Or you think that for some reason prompting a LLM and forwarding it's output after vetting is different than prompting a human and forwarding the output after vetting.

It is not the same yet it requires reading which in turns requires interpretability, trying to make sense of what was read within a very specific, maybe explicit context, e.g. specifications, and that can be optimized for an audience, or not.

For example the code I would write for somebody who just started to learn programming will not be the same as the code who is an expert even though both will pass the specifications.

Consequently taking into account who (if anyone) will actually review, and thus read, the code changes the code itself.

If by majority of the population, you mean no women, no slaves, then yes, sure. And even from an American or NE perspective, you're still overestimating its popularity.

Literacy rates in the 1800s across different regions in Western Europe: https://brewminate.com/the-growth-of-literacy-in-western-eur...

> According to data compiled by Our World in Data and the World Bank, the literacy rate of the world's population from secondary school age onward was only 12 percent in 1820 - around one person in ten. In 1900, it still barely exceeded 20 percent. From the 1950s on, world literacy began to take off, hitting 42 percent in 1960 and 70 percent in 1983.

https://www.weforum.org/stories/education-and-skills/reading...

> I don’t care what’s possible, I asked what was plausible.

OK then. BOTH ARE PLAUSIBLE. Clear enough?

It's absolutely not up to me to decide whether they came up with it independently.

But then again, it isn't up to you, Jim.

> If you aren’t willing to say you believe this, you are just asking me to waste my time arguing with a phantom opinion nobody holds.

This is the weirdest thing to put in a comment you could simply not have made. If your time is so valuable why are you here commenting on HN?

Again: aphorisms have value regardless of how many times they are used. People can use them without you having the right to demand proof they thought them up for themselves like some prep school debate bully picking on a younger child.

Seems stupid.

I use LLMs for nailing down the narrative in bullet point form. It proposes a lot of inappropriate points that I tell it to remove, but makes some useful proposals as well. It’s actually fairly efficient for narrative exploration, I then keep it away from prose for as long as possible since whatever it writes is pretty bad but once I start editing, it will usually clobber my edits if I make any narrative changes. Humans are still better at actual prose writing, especially when it comes to word rhythm.

I agree with the central thesis, but disagree with some details. I wrote a blog response: https://writing.brlewis.com/posts/2026/llm-copyeditor/

If English is your second language and you intent to write in your mother tongue, my advice is to use English to communicate with the LLM.

This makes Rule 1 straightforward ("You may not use a single word an LLM suggests to you") and helps avoid the horrible feeling that the machine is taking over your personal voice.

Read Ong's "Writing is a Technology that Restructures Human Thought".

Emitting words is not really what "writing" is.

https://twinada.wordpress.com/wp-content/uploads/2012/09/708...

A workflow I've recently discovered for myself, that may be useful to anyone who speaks at least two languages: Ask the LLM to write a first draft in a language that isn't the one the piece should ultimately be in. Then use the LLM's first draft as a blueprint, to write a first draft yourself in the intended target language. That way, you won't be tempted to shut off your brain and just meatproxy anything from the LLM's first draft. Then hand your actual first draft back to the LLM for editorial suggestions. At that point, feel free to ignore suggestions, push back on bad ones while adding a rationale about why you don't like them, etc. Work in the best suggestions, then iterate. If you're in “writer block” mode, force yourself to keep going by asking the LLM for five suggestions on how to phrase something, form an opinion about what you like and don't like about them, draft it yourself by frankensteining from among the suggestions, without using any one suggested phrasing literally. Optionally, provide feedback to the LLM explaining your thinking about why you did or did not like certain suggested phrasings.

For objective/factual writing, asking for a fact check is a widely-used practice. And there's a useful counterpart for writing subjective things and opinionated hot takes, namely to ask the LLM: “Anticipate how this might be misunderstood in ways that might cause offence or trigger people.” Sometimes, you might want to say something that will cause offence, and there might be no way around it. In this case, making a conscious choice to let that be might be the right thing to do. But I find it surprising, how often an LLM will see an angle that I myself did not see, where something might cause offence in a way that I genuinely missed and genuinely want to fix. And then that gives me a chance to re-write before posting something.

Till today, I was doing the complete opposite of what the article suggests. Definitely going to try this approach out, pretty helpful.

I have a specific ChatGPT chat that rewrites my posts really well every time; I usually just give it a rough draft and let it clean things up. But it was pretty interesting to realize how good these LLMs are at actually finding mistakes rather than just rewriting things. The point about not letting the model influence your voice was also very, very relevant.

For blogging I just don't use an LLM at all. I just use a spelling/grammar checker. I get to sound like me even if that is imperfect.

For docs I have used an LLM for the convenience of having them kept up to date.

Writing is thinking. You are outsourcing the process of thinking through your thoughts so you can communicate and understand them.

I guess more often than not we are thinking about what to write rather than how to write :-D

>So give the options to a model that doesn’t have the context of your editing process.

> I conjured a bit of software to manage this for me[...]

I don't understand how the annotation application shown helps with the stated goal of preventing the model from praising you. What's the actual software assisted workflow?

    Workshop a piece with me. NO ENCOURAGEMENT. Encouragement is
    useless; the only useful things are suggested corrections.
    DO NOT WRITE COPY FOR ME. Any words you provide will be disqualified,
    so if you come up with good words, I’m fucked because I can’t use
    them. Tell me ABOUT what should change.
    Assume I know my audience extremely well.
    We’re going to do this in a series of passes. Keep your suggestions
    locked in to the current pass we’re on.
    Note typos, but DO NOT generate long structural critiques based on
    those typos; note them and move on, assuming that I meant to write
    properly.

The frontier live voice models (especially GPT Voice/Live or whatever they call it this week) do not have the typical AI-tells. So I wonder about a few things:

Why is this?

Why couldn’t whatever went into post-training these voice models be applied to the text-gen models to get rid of AI smells?

Could we set up a pipeline that leverages the voice models to clean up text that reeks of AI ?

This is similar to how I've come to use them but it's focused too much on the mechanics or writing. I've found them more valuable at role-playing as a specific profile of reader and asking it how they would receive the arguments advanced by the writing and what counterarguments they could present. After some rounds of this process, you either close all the gaps in the argument or realize it's not a strong one.

At one point in my graduate college studies, I tutored qualified (DEI) undergrads 'for free' (for them) as part of a work-study program which paid me a bit for that work.

OTOH, an AI could not have qualified as such a student. But even if it did, before we began I would certainly bargain for considerably more than work-study tutoring paid.

I worry that reading too much LLM output will make me tend to write more like an LLM. Seems like it would be inevitable, to some extent. (Here's why it works)

Great article. Thanks for sharing.

Not letting your llm write is good advice and letting him spot problems also is if you want good writing.

For anything beyond "good", we're still on our own and I love that for us.

If you write with an LLM it will taint you voice. Slowly but surely.

Thought I'd get an article with tricks how to stop an LLM from sounding like the most obnoxious Linkedin influencer. Instead it says to use it as a copyeditor.

Ok - but that's not very useful.

One thing that you should absolutely never do: have an LLM review and “improve” the text repeatedly in some closed loop.

This might work well for some tasks (coding), but for writing it will absolutely take reasonable text and turn it into a pile of incoherent garbage no human would ever write.

Maybe I’m the only one keeps trying this (more often than I’m willing to admit), but I suspect it’s a common cope engineers reach for when having to deal with the not-so-fun task of writing prose.

Sounds like solid advice. Those typical Claude turns of phrase are really grating on me. And I now see them everywhere. Many YouTube channels that are presented by a real human, still sound fake because their script is blatantly written by AI. I hate it.

And yet, I'm stuck having to write an important message to someone that has real stakes riding on it, so I use Claude to refine it, and I see myself succumbing to Claude's pressure to use its words instead of my own. I hate it, but my words aren't very good.

And finally Claude proclaims the message perfect and I let my wife take a look at it, and she rejects the entire thing, restoring some semblance of sanity again.

I mostly use LLMs for coding, so this is an interesting use case. I’m curious how long an article it can handle in practice before the context starts becoming an issue.

> Readers can detect LLM words in the parts per trillion. However much work you put into scuffing up and humanizing it, an LLM paragraph will register to much of your audience not as writing but as output. So, first the bad news: you have to write for yourself.

---

and then

> Those rethinks are load-bearing parts of your voice. Readers won’t put their fingers on what’s wrong, but they’ll sense that you’ve become artificially-flavored.

---

Are they wrong in the first paragraph? Or are they simply not following the advice they dispense to others?

"Streamline and improve" is redundant. Pasteurization is great. "jacking it with corn syrup" is nonsense.

I'll get my writing tips somewhere else.

But why would you want to write with the help of an llm, you should be moving in the other direction. Don't let the creativity die!