From TFA: "Eric Schwitzgebel writes that . . . There’s a huge cognitive difference between nodding along while reading something and actually productively generating a text. Two reasons: First, once the text is on the page, it’s easy to passively let the approximate word suffice, rather than thinking about word choice in the same effortful, active way we do when generating prose de novo. Second, as I suggested above, I doubt that human beings, even experts, have a good sense of all the factors that shape word choice -- everything they’re being sensitive to. You would have phrased it slightly differently, and even if you don’t know that, or why, a different signal is sent and received."
This is the best articulation I've seen of why simply reviewing and copy-editing does not provide remotely the same value as writing from scratch. I spent a considerable amount of time over the past two weeks reviewing and improving a work document that was the output of an LLM. Given the number of people involved and the final level of effort, I'm firmly convinced that writing it manually would have been faster and resulted in a higher quality product. Getting the wording right matters.
I really did like the author's framing there, but I think there is a different, simpler way to put it:
Why would you think that asking someone else (that is, another human) to write something (and then reviewing it) is the same thing as writing it yourself?
You may trust the other writer's opinions and knowledge, but it will not have the same tone, structure, word choice, understanding, or narrative flow as it would if you were to write it yourself.
And when it's an LLM, you should not trust it's "opinions" and "knowledge", because it does not have either of those things. The appearance of those things is just that, an appearance.
Except that there are such things as speech writers, whose job is literally to write prose in the voice of the person who’s going to read it.
And there are countless other jobs out there where people are the “voice” of another party. From those who manage social media presence, to PR firms, to copyrighters, all who put statements out on others behalf.
There was already an industry of professionals whose job it was to have others write things on our behalf, and that existed long before LLMs were a thing. What LLMs did was make that that service available much more cheaply.
Speech writers do not write in the voice of someone else, they create the public image of say a president who is inarticulate and make them seem polished and articulate, if anything they lend their voice and ideas to someone else. That’s why they are hired.
And no LLMs are not producing ‘the same thing’ as a speechwriter. If you are unable to tell the difference I’d suggest doing some a lot more reading of human books. LLMs produce stultifying pablum without coherence or style.
> Speech writers do not write in the voice of someone else, they create the public image of say a president who is inarticulate and make them seem polished and articulate, if anything they lend their voice and ideas to someone else
Nope. The exact opposite in most cases. Eg Jon Favreau (Obamas speechwriter) has said:
“As a speechwriter, your ego has to take a backseat. The goal is to make the speaker sound like the best version of themselves, rather than to showcase your own cleverness."
> And no LLMs are not producing ‘the same thing’ as a speechwriter.
I didn’t say it was the same. I said it was an existing industry that AI has stepped into and AI is taking the low end of market.
It’s just like how AI has captured the low-end of many tech roles too.
> If you are unable to tell the difference I’d suggest doing some a lot more reading of human books.
“Some a lot more”? Very ironically timed editing error there. You can bet an LLM wouldn’t have made that mistake ;)
Speech writers and shadow writers in general are writing on behalf of other people, but they're not writing the same thing. They can't! Usually they're writing something better, with the proper voice and authentic to the views of the person, but still not the same as what they would have written.
Depends on the individual. It’s usually a highly collaborative process. But in general, you’re correct that you’d expect a good writer to write better than the person who hired them.
The key difference is that while an encyclopedia holds facts themselves, LLMs trained on that source material encode something more like a highly probable facsimile of those facts - the original fact was lost, LLMs are lossy, but can often be generated again with a decent level of accuracy by churning through stats about words, concepts, and relationships between them.
The whole catch is that they can often be regenerated. But LLMs (on their own, in their parametric memory - which is the result of training) don't have any conception of whether what they've generated is a real reproduction of some training material or whether they've invented something false that seemed probable based on their encoded stats. When the probability produces something contrary to what was in the training material, you get hallucinations.
They're very, very good predictive text models and can be very, very powerful when hooked up to other tools or outside databases. But its fundamentally lossy technology and all the books having been fed in doesn't guarantee all of the knowledge from those books can be spat back out.
Humans can learn texts by heart, even lots of text (I have some expertise on that, having studied opera singing). They can reproduce these texts accurately, deterministically and repeatably. An LLM is a statistical machine. It does not know any text by heart and it is by pure chance that it sometimes reproduces existing texts verbatim.
Break it into small chunks and practice. Practice more. Put the chunks together and practice even more. Like any skill, it isn't something people are magically good at beyond minor proficiency.
Every performance- singing a song, playing an instrument, performing a stand-up routine, giving a speech, performing a theatrical role are all things that are best done from memory but require practice.
There are pneumonic tricks you can use- I've seen some people do it for tricks like memorizing the order of a deck of cards- but it's less useful for long term recital because it helps with order but not comprehension or fast indexing.
Humans sometimes don't remember things 100%. LLMs can produce certain text and be run deterministically.
I don't really understand this side of the debate other than as a gotcha tbh.
If LLM use atrophies your brain and skills that's bad. If it has a repulsive writing style that's bad.
I'm not sure what the debate about whether an AI is a statistical parrot unlike humans accomplishes. Is relying 100% on a bad human speechwriter somehow better?
It comes up frequently when discussing whether LLMs actually have intelligence or qualify as AI, especially in discussions on the path to achieving AGI.
The primary complaint seems to be that LLMs are held to a higher standard than humans, though I don't particularly buy that line of reasoning.
It's funny. I always thought that writing was meant to inform, persuade, or entertain about the subject at hand.
But in professional settings, a lot more of the informativeness is about the author, and a lot more of the persuasiveness is I'm worth your time and money. So, if the author is an LLM, and obviously so, what exactly are you informing your audience of (about yourself), and what are you persuading them to do (with your article).
People using LLMs not in a gruntwork manner tend to confuse style and tone with facts and knowledge.
LLMs are great to make drafts if you give them the source materials. They're great at validation if you give them the tools. They are great at layouting if you give them linters.
Encode architecture and decisions in your workflow, then proofread what your agents have been working on. Not the other way around.
Better validation and testing means more work will transform from exhaustive decision work to automateable gruntwork.
> First, once the text is on the page, it’s easy to passively let the approximate word suffice, rather than thinking about word choice in the same effortful, active way we do when generating prose de novo.
Meanwhile, me, as an English non-native speaker, ended up discussing two sentences I want send to HR for ten minutes while applying to a job.
I do believe there's generally a bias to accept something that's already written. The much bigger reason though is why you let somebody else write it to begin with.
It might just be that not thinking carefully about every sentence/wording was the exact thing that made you use AI to begin with.
> I spent a considerable amount of time over the past two weeks reviewing and improving a work document that was the output of an LLM.
I had the same experience with texts where I have a very detailed expectation of the desired result.
This is just a general limitation. For code, there's the saying "the precise description of the solution is already the code". Describing X is a simplification of X, oftentimes it's fine guess the gaps. But when it's not, it didn't help to describe X, you have to manifest X itself.
A workflow I've recently discovered for myself that provides a kind of middle ground: I'll ask the LLM to write a first draft in a language that isn't the one the piece should ultimately be in. Then I'll use the LLM's first draft as a blueprint, to write a first draft myself in the intended target language. This fights my brain's temptation to just shut off, and forces every word choice to actually be mine. Then I'll hand that over to the LLM for editorial suggestions and iterate from there.
It's probably annoying the way I oversimplify things, but you reminded me of something I read somewhere.
Some Jazz musician was asked to define Jazz.
He couldn't really, his answer was something like: 'I don't know. But I'll know when I hear it'.
To add something to the discussion directly: thinking requires vocabulary. Vocabulary is the currency of thought.
You can usually express an idea with a few thoughts, or many. The audience, and amount of details chosen should always be kept in mind. Writing helps you to remember vocabulary and word choice when expressing ideas.
While I mostly agree with you, what is the real difference between an editor saying, 'when generating prose de novo' sounds pretentious to an American audience so let's use 'when writing from scratch' instead, versus an LLM giving the same advice?
Not that it's necessarily better but it's the kind of thing an editor might pick up and so why would you reject the advice if it's a machine giving it rather than a human?
TFA is about having AI do the writing, not about having AI suggest editorial improvements. The parent comment argues against having the AI write the text and having the human merely review it, not against having the AI review human-written text (and the human deciding which of the AI’s suggestions they might apply).
I don’t see any conflict with what the root comment wrote. There are probably some differences between human and AI editorial suggestions that could be discussed, but that wasn’t the topic here.
Same flow as translating from your native tongue to a foreign language.
Going from foreign to native is easier than native to foreign. The latter requires completely different brain paths and a lot more understanding of the language to actually get to something correct.
Use AI to write things for you to read that you wish someone else had written. ‘Give me a summary of the research on this topic’; ‘Write a report on this data to help me make a decision’. ‘Take this transcript of a meeting and write me the email it could have been’.
Don’t use AI to write things that you are producing for someone else to consume.
I think that's somewhat reasonable, but not always: the effort of researching and of writing yourself will teach you better then reading something someone else wrote, especially an LLM that will get things wrong in surprising ways that a human would not.
I'm not saying you should never do this: our time is valuable, and we shouldn't spend it on things that are not genuinely worthwhile to us if we can help it. But we're still losing something by having an LLM write for us, even if the intended audience is just ourselves.
I do not personally even think AI is useful for editorial purposes; I would rather read and re-read my own text and edit it down (which, despite my overlong comments here, I actually do), than have someone else do it. The reason is simple: editing is my search for the clearest way to express my thoughts.
I tend to turn grammar checkers off, but leave spelling checkers on, because they catch typos and a handful of my spelling gremlins like "liaise".
I am someone else, if AI can write something that I want to and can consume then it can do the same for other’s consumption. Perhaps a disclaimer would be nice, “co-written with ____” but I see no difference between the two though of course I understand what you are trying to convey conceptually
When I seek input (whether passively or actively; it doesn't matter) from another person, then I generally want that person's input. I want their words.
They can be broken and jumbled, or terse AF, or combine to be Pulitzer-quality prose. Either way: I want to consume their own unique human expression of a concept.
There's value in raw human expression, including in the missteps.
If I instead want the regurgitated waggyings of a bot, then: I know how to get that on my own.
We've all got web browsers and pocket supercomputers. We've all (well, most of us) been alive and aware of LLMs since their recent rise from infancy.
It's a no-brainer for us to paste some paragraphs into our favorite chatbot and get a summary or an artificial expansion or whatever else we wish to have. If that's what our goal is, then we can do that on a whim -- and we can still retain the original expression.
But doing it on our behalf is deleterious, unsettling, and unhelpful. It has negative value to the beholder.
> When I seek input (whether passively or actively; it doesn't matter) from another person, then I generally want that person's input. I want their words.
There's a (probably subconscious) reason for you wanting that (and I just posted this elsewhere on a different story, but it bears repeating): https://news.ycombinator.com/item?id=49767867
this is where I perhaps differ from other commenters, I do not have higher tolerance, regardless of whether I asked it or am reading something someone else posted, fully or partially written by LLM.
> this is where I perhaps differ from other commenters, I do not have higher tolerance, regardless of whether I asked it or am reading something someone else posted, fully or partially written by LLM.
Perhaps English is not your first language? To people who are well-read, LLM-ism grate on their nerves like the high-pressure sales pitch you hear on 3am shopping channels.
But the reader cannot be sure they ARE understanding what you are trying to convey. Maybe the LLM added something you never thought of, maybe it made form to your various thoughts in a way that isn't actually true to the way you would have communicated them if you did it yourself.
The exercise of writing is an important step of actually understanding your own thoughts properly. Readers are reading to get a sense of your own experience and ideas on the topic, if it turns out you didn't actually write the final thing yourself they'll always feel a sense – in whatever way – of being conned by the author. Intentional or not.
Technical documentation, where what the reader wants to come away with is the clear facts on what something does and how to use it/whatever, is a different matter. But even then editing is key – overlong text, unnecessary information, these will again fight against what the reader is there to get.
With that said, if you do want to do it, a disclaimer is at least upfront and may well show there's no intentional attempt to deceive.
> Technical documentation, where what the reader wants to come away with is the clear facts on what something does and how to use it/whatever, is a different matter.
I don’t agree with this one. Writing technical documentation can be a slog, but quite often you will discover very annoying design decisions in the process. Like you’ll write a tutorial for a CLI and realize that actually some flag makes no sense, or some default should be changed.
This also goes for dogfooding. If you’re using AI to write and run the tools, try to use them yourself and see how nice they are. AI will happily run whatever jank it’s working with.
Docstrings perhaps, but even then I find myself editing heavily to avoid cruft sneaking in.
That's logical, yes. If person A wants to read a summary (even a generated one) outlining the state of modern space dinosaur research, or whatever, and person B also wants the exact same thing, then person B should be happy to read the summary that person A prompted.
The same is true if the thing they both want is a stale cupcake bought from a gas station. But this still doesn't mean that person B wants to be presented with person A's stale thing, because the overhead of the social transaction isn't worth the object. Which is to say, it makes a terrible gift.
The words of a machine are not of interest to me. I want to hear what you, the human, have to say, imperfections and all. Reading something written by an LLM is actively annoying, and it's not very respectful to inflict that upon others.
There's a difference between me getting AI to write something I want to read, and getting it to write something you want to read. The difference is that I know me, but I don't know you.
The other difference is that, if I get AI to write something for me, I expect something AI-written. If I am writing something for you, you expect something that I wrote, not something that AI wrote.
Moreover, when you wrote it yourself, you are intellectually responsible, you can defend it, while if the LLM wrote it for you, nobody is really responsible: The LLM didn't write it "out of its own motivation", it was merely extrapolating your prompt; and you only provided the prompt, not the extrapolation.
So the intellectual responsibility is diluted to a substantial degree: The text is not the opinion of anyone who can be expected to honestly defend it. It's bullshit in the technical sense.
TLDR: Communication from a human is not only the material being communicated, it's also a signal that the human understands the message they are sending you. When you pass that message via an intermediary (regardless of whether the intermediary is a human or an LLM), the receiver is never sure that what they received is what you intended.
It's doubly worse when you review it, notice a few additions and think "Great - that makes me sound smart and knowledgeable about this topic", because the receiver now has an inaccurate view about your skill.
At any rate, using an LLM to construct a message for someone else to read is basically sending the wrong signals about your skill levels.
(This probabyl explains why so many people are so ready to use an LLM to write the material - in their minds, it makes them sound smarter than they actually are, but as we've seen from people on this forum vigorously defending the practice with "pangram is wrong" and similar nonsense, they themselves are not skilled enough to recognise slop)
> TLDR: Communication from a human is not only the material being communicated, it's also a signal that the human understands the message they are sending you. When you pass that message via an intermediary (regardless of whether the intermediary is a human or an LLM), the receiver is never sure that what they received is what you intended.
100% right. The fundamental point here is:
- it's OK to be wrong
- it's OK to be wrong and not know
- it's a little less OK to be wrong when you should have known
- it's much less OK to be wrong when you haven't put any effort into being right
- it's far less OK to be wrong when you are pretending to know either way
Relying on AI writing risks pushing you further down the path to worse kinds of wrongness.
"AI writing is vague and wrong in hard-to-notice ways"
This one felt to me, at work I lost a lot of time re-reading and correcting AI prose that was launched to "summarize" a collective white paper, and where subtlety and nuance disappeared in ways that harmed the initial ideas/arguments I was writing about, but in sneaky ways.
AI is very bad at properly handling statements that make heavy use of vague quantifiers (e.g. "some", "most") and also commits a lot of pretty serious logical fallacies. It is also bad at handling subtle logical negation, generally.
One of the most egregious negation issues I run into a lot is when I (or someone) makes a statement of the form: "not X" or "X is thus not true", and the AI then proceeds to interpret or summarize this as 'whatever is the opposite of X is the case'". This will cause it to go down a useless path investigating or disputing the opposite of X, which generally has no relevance or bearing on anything.
It also often very harmfully will replace your carefully chosen words with weirdly specific academic operationalizations or formalisms, then again waste huge amounts of text refuting / showing "problems" that result from that formalism, all of which again have no bearing or relevance on the original statement. An example would be you saying something like "intelligence, generally, must surely explain some of the differences in X", and then it will go "actually IQ does not correlate with X", unless you specifically tell it not to conflate psychometric IQ with intelligence generally.
Sometimes this is helpful, but the more specific / technical the domain, the more often you specifically have to prevent it from going down stupid paths that should be obvious given the expert context and wording, because it can seem almost hungry to try to catch you in some kind of insipid 'gotcha'. Much of these issues often clearly arise immediately from the first-pass "reword what the user said" part, given the reasoning traces.
Most of us have worked our whole lives to be writers in some domain: Code, documentation, specifications, product requirements, corporate communication, etc.
If we don't want to be writers, then we have to be editors. And editing is an entirely different job and it's not an easy one. In many ways it's harder.
Especially when LLMs love writing novels when all we need is a short story or less.
Instead of asking AI to write something substansive for you, ask it to critique your writing. Then use your own judgement as to which bits of advice to follow and which to ignore.
The LLM will always give you a full rewrite — don't use it. It always does too much, and persuading it to tone it down is a constant battle.
Here's how I use it to write. I dump an unedited stream of consciousness into the prompt. Sometimes it's just me having a conversation with myself thinking through what I actually want to say. Then, I have it ask me to clarify anything that's unclear. Then I have it write me a rough draft. I read that and then just rewrite most if not all of it while I edit the output. I use this workflow for work stuff or dealing with businesses in personal life. I'd never use this to generate a letter or email to a friend or colleague.
Hmmm, I've never actually checked that. Just had codex run the numbers excluding the boilerplate in some of the documents output is on average ~13% shorter than input, I only checked the ones that were all of the same type (proposals) since it was easier to have codex run the numbers on that.
I compared at final edited output less boilerplate vs sum(prompts).
If I had to guess I'd say far more people are better at editing than they are at writing de novo. My approach probably limits the creativity of the output but I'm not writing a novel.
I can't help but think that some of the arguments here apply equally well to writing _code_ as they to to other substantive writings.
Writing code is thinking, AI code is often vague and wrong in hard-to-notice ways, and the (human) reader of code is the one that pays the cost for this later.
The cost/benefit analysis may still work out differently for code though...
I never use AI for writing but do so regularly for code. They are quite different to me.
My reason: code can be checked objectively. I can run it and confirm it works. I don't get attached to it. I don't feel pride in it (even when I write it by hand). Code just is. It's lifeless, inert, and entirely replaceable.
How do I do the equivalent for prose? How can I tell if my words "work"? Do they clearly convey my ideas to the intended audience? There's an element of subjectivity here forces me to identify personally with the prose.
Code has no such personality. I don't tie my identity or ego to code the same way I would an essay.
> My reason: code can be checked objectively. I can run it and confirm it works.
Running the code only confirms that it works with the precise input, in the precise environment, under the precise circumstances you run it under. It doesn’t ensure that the code is correct. Thinking through the code, on the other hand, lets you consider all possible cases. It’s the difference between experiment and (mathematical) proof.
For an objective correctness proof, using a formal language is indispensable.
Even thinking through the code is not complete because our brains make mistakes. They also make assumptions, thousands of them, that you don’t know about. It’s tricky, because every thought you have has an entire sea of preconceived notions and knowledge behind it, which you can’t identify. In that way, it’s similar to an LLM.
In practice, for code, testing code is the way to go. Formal proofs work, too, but the barrier to entry is high and it’s overkill for most business applications. LLMs can be very good at writing tests, if you read the test thoroughly and analyze them.
I understand the distinction you're drawing, but in my experience, it gets a lot blurrier as a project evolves. Caveat that I use LLMs all day and have for quite some time, so I'm coming from a positive perspective.
The danger for me in LLM code is the same as in writing, it's just that I'm not typically writing at the same scale as when I'm building something. The final piece when I'm writing is usually a message or a 1,000 word article at most. So I'm naturally going to analyze it quite intensely, because I can afford to. And I don't really use LLMs for this at all. I use them for things around writing (research, interrogating ideas, situationally specific stuff, mapping, visuals, publishing, etc.) And the code equivalent to an essay or message would probably be something like a single script, or the sort of thing I'd write as example code when I'm teaching. In those settings, again, LLMs can be helpful, but I'm still going to be really opinionated at a highly detailed resolution.
But a codebase is more comparable to a novel than an essay. Or more directly, the writing in a codebase is usually the documentation, which grows commensurately with the codebase. And the real LLM risk here is the drift that can happen over the course of many epics or "chapters" as the LLM writes "code that works but is imprecise and probably shouldn't work this way" or introduces weird new terminology that neither of us can precisely define. Worse, this usually becomes obvious down the line, and I have to parse through the verbose constructed world the agent has created to trace the issue back. That's a big cognitive tax, because I'm holding these weird parallel worlds of "How did the LLM's alien brain get here within the bounds of the contracts" and "What do I really want this to look like".
So I think it's fundamentally the same phenomenon, and we're all developing our skills around working with it in real time.
> I don't feel pride in it (even when I write it by hand).
Hard for me to imagine. You feel no pride in using a tool to accomplish a goal?
> Code has no such personality. I don't tie my identity or ego to code the same way I would an essay
Code certainly does have a personality. When working with teams for a while you can absolutely get a sense for which person wrote what code in a codebase, just by subtle little tells.
You may not tie your identity or ego to it, bully for you, but for me I take a lot of pride in writing clear and maintainable code that contributes to big projects in meaningful ways.
Maybe the problem with software is there's too many people who treat writing code as a mere means to an end, instead of a very important part of the process.
Well, there's some nuance. I'm not really proud of my code. I'm proud of the software that gets built with it.
Code is just sitting in a git repo somewhere, not necessarily running. That's a big distinction for me. Consider that code volume has increased 14x on github this year, but we see nowhere near that increase in the actual usable software. Code is cheap and getting cheaper. Running software ain't.
Another way to put it: I'm only interested in code so far as the value it provides. That value, not the code itself, is the source of pride. If I can provide similar value without any code at all, I'd gladly do so.
But code is the mean to get an idea across to someone. It’s the difference between “I think this can work” and “I’ve worked on this and it does work”. I’m part of the OpenBSD mailing list and it quite nice someone sharing a diff for an idea or experiment. It may not get merged, but it’s better than realm of deliberations.
> I can't help but think that some of the arguments here apply equally well to writing _code_ as they to to other substantive writings.
All of the arguments here apply to writing code; Why would the quality of writing differe when writing prose vs code?
I cannot read LLM slop prose without getting mentally fatigued trying to figure out what it is saying, and I've discovered that the quality of the code it produces has the same effect on me.
And also to think in different directions than you would have gone in isolation. Regardless, this is a great heuristic / rule that I will be sharing and keeping in mind.
I don't know what everyone else is doing but I've certainly been writing a lot more often because of my interactions with AI. Maybe some people used to write more but have been leaning on AI so they can write less but I don't think that's universally the case.
I agree that being lazy about writing and simply using a short prompt or list is detrimental if you are replacing your own output but you don't have to do that and you don't have to accept any of the output either. The best chats I've had usually start with a large amount of my own writing up front, thinking through the idea, listing several alternative ideas, asking a lot of questions, jotting down related topics, etc. and then reading the resulting output and critiquing it, asking for clarification, doing my own research on it, even just discussing it with the AI and doing this for a number of turns until I feel like I've exhausted the topic in the chat. I then often take what I've shaped in my own mind from that process and write something myself, either that or take the best parts of the output and edit, rearrange, reinterpret, and add to it in order to produce something.
I guess it all comes down to whether you are actively engaging with the material, regardless if that material is the result of a Google search, pulled from a book or generated by AI. If you just read it or copy paste it and don't engage with it and think about it then it doesn't do much good.
Communication between two people is hard enough. Involving another actor as an intermediary rarely helps. It quickly becomes a game of telephone. Doubly so if the intermediary is an LLM.
I built a place to continue the disagreement around Grunewald's article as a structured one-on-one discussion. Its an experiment in finding better ways to have deeper conversations online if anyone here wants to try it: https://hwgo.io/p/7
Good or bad please let me know what you think of this format!
Is editing really any less 'thinking' than writing the original copy? I don't use LLMs to generate prose but I can compare to writing code. Let's say I use an LLM to write code to some specification. After it meets the spec I go over everything with a fine-toothed comb - tweaking things to a style I like better, rewriting pieces or interfaces that I think would be better a different way, etc. Is the final work product any different than if I wrote it from scratch without absolutely any LLM use? Personally I would say no, if I do enough editing to be perfectly satisfied with it then it's the exact same output in the end.
I think it's hard to generalize because people have different approaches to writing, but to me, there's a marked difference between iteratively polishing my own text and editing for someone else (and that includes LLM output produced by my own prompts).
In the second case, these sentences were never a "voice in your head", and you never invested any effort into shaping them, so it feels more like janitorial work / drive-by nitpicking than craftsmanship.
I offended a coworker by asking him to not send me AI written things because they’re wrong in subtle ways and overly verbose. I’m glad someone else shares my opinion.
Any form of thoughtful writing is usually meaningful when it's deeply human. Having said that, I've found it helpful to use AI as a sounding board and planner before actually getting into the writing. For me, the hardest part of writing a piece is always in getting the actual writing started, and AI has helped me unblock that a few times.
I don't find AI helpful for truly constructive or creative prose, it devolves to the median output very quickly and it's almost painful to hear in my inner voice when I read out an AI-ified idea. The "not this but that" formulation is especially bad or cloying attempt at pretending to refine an idea or lightly correct something to start from a position of authority, very annoying.
That said, I have a lot of situations at work where I am asked to simplify something I am an expert in or convey something for a different audience, particularly as I prepare for presentations and I do find it helpful in helping me step down my writing or work. YMMV.
I also skip words a lot and can miss that even after 2-3 editorial passes and it's very helpful at that vs. say normal spellcheck.
One way I use AI to help me write that I've not seen anyone else talk about is this: I'll say "give me multiple versions of saying xyz". I'll almost never use any of what it outputs verbatim, but seeing it said in different ways helps me clarify how I'd prefer to approach writing it.
When I was a graduate 20 years ago, the partner would sit with a print out of every word I wrote and a red pen marking every change required, as the partner knew anything that was written in his name had to be exactly what he would have said.
Same needs to go for LLMs, leverage them for productivity both make sure you are owning their output.
Every time there are some articles like this, some people replied with: "Yeah I agree, but I use LLMs 'this way' and my way doesn't suffer from the problem mentioned." I'm not saying these people are wrong but it's interesting that people always think their way of using LLMs is the best (and other people are using LLMs in the wrong way).
I’ve found Gemini and Codex can do pretty well by just telling them to follow my writing style, be brief, be less robotic, be professional, be semi-professional, use less technical jargon, no em-dashes, etc. depending on the task. But, I only use that for short summaries of things I am deeply familiar with to send a slack message about or similar. I also edit before sending. It takes maybe a couple of prompts to get 90% of the way there.
Claude on the other hand is atrocious at sounding human, it’s like it’s deliberately following LLM -speak stereotypes no matter how you tell it to write. I finally told it to stop using “Claudeisms,” which gave better though not great results.
Irony an article written by AI, wants to say do not use AI for writing. Following sentence looks like written by an AI, human do not use double dash, it seems AI instead of emdash wrote double dash;
“I think you should almost never use AI to write -- that is, to do the thing you’re doing when you type words on a page -- whether for a blog post, a research report, a memo,”
This absolutely doesn't look like AI writing and this is a pretty weird signal to use.
I have a much better theory: as it happens, Substack automatically converts "--" to em-dash as you type. I'm guessing the author knew that and wanted to use this feature, but composed the article in another editor and then copied-and-pasted the whole thing. Because of how Substack implements the feature, the substitution doesn't actually kick in on Ctrl-V, so all these "--" remained in place.
Haha, I do actually use double dash, as I write in org-mode. I then convert that org-mode text to Markdown and copy and paste the Markdown into the Substack editor, which preserves the double dash. In case it needs to be said, the article is 100% human-written.
In some of your other writing can see use of emdash probably thats also intentional, emacs might have very special key combo for it to use unicode emdash instead of ASCII “-“. I commented about double dash (“--“) because because thats what LLM does when asked not to use emdash for writing.
“Another promising approach is to conduct on-site inspections regularly—say, every six months— to make sure chips are not just installed to fool inspectors and then removed and smuggled, as has happened before. However, conducting on-site inspections at this scale would likely require additional budget for export enforcement, or for US authorities to implement a third-party export auditor program.”
Yes, for that newsletter, where I write in my day job as a policy researcher, we use em dashes. For my personal writing (example: https://www.lesswrong.com/users/erich_grunewald), I use double dashes.
Double dash has long been the way to clearly indicate in plain ASCII text an em dash as distinct from a simple hyphen. So I think the author here is here emphatically reclaiming the use of the em dash while pointedly not using the Unicode character so beloved by LLMs.
I think about this a lot. I started my teenage-to-young-adult life in the literary world as a poet who loved programming, and made a living as a ghostwriter. At some point, I fell in love with mathematics and wound up in ML in research/engineering for the last 8 years or so. So, I've thought a lot about writing and ML and their intersection.
I think one of the underappreciated things about writing "substantive" work is that the work you see at the end isn't the first attempt. And I don't mean the first draft of the piece. I mean that almost always, writers iterate on the same topic many times, either with complete published pieces or abandoned drafts or even just conversations and sessions of unproductive daydreaming. It's a cliche that your best work typically also comes out fastest, but it's not because of divine inspiration, it's because you've whittled the big idea you actually care about down so much in your mind that you instinctively know exactly how to write it.
My experience has been that for people who don't work this way or don't write a lot, LLMs can give them this incredible feeling of leaping straight from inkling to "substantive" writing. And because they haven't built up those muscles or "taste", they don't immediately recognize that it's imprecise and hard to follow.
That's not to say they're not brilliant in their own right, just that they haven't spent a lot of time on this particular thing. Sort of like a very gifted programmer who doesn't have a ton of experience yet (speaking as someone who is gifted at nothing and frequently has to do things they're inexperienced at).
So I don't think the problem is that LLMs just write bad. It's that LLMs are so wonderfully powerful that they allow you to confidently leap forward to create something that is a little beyond your experience.
And that's why I have different reactions to heavily AI generated writing. When it feels like marketing at scale, it grosses me out. But when I feel like it's just someone who is excited to write an idea and maybe doesn't have a lot of experience doing it, I'm not judgemental. My hope is that it makes them more excited about writing, and that trying to make their next piece better will lead them inevitably to start thinking about where the last piece fell short. And if there's some slop along the way, eh, I'm not compelled to read it.
This advice is too coarse. I believe LLMs have made me a better writer of technical reports. I have a keen ear for language, and I can usually tell when a sentence or a paragraph is orbiting around a point rather than landing on it firmly. Some writers insist that unclear prose is dispositive proof of unclear ideas -- a sign the writer has not thought things through. My experience is different: I often have a clear idea in mind that I cannot quite find the right idiom to express. Sometimes I can find it if I keep revising, but staring at a sentence too long can entangle you in a particular linguistic frame that can be hard to escape.
LLMs are helpful because their idiom-list is vast and they are indefatigable and infinitely patient. You can keep iterating on a sentence and it will keep giving you fresh takes. Eventually, whether through careful steering or brute-force iteration, it will come up with something that has the right resonance. Like a word on the tip of your tongue, you recognize it when you hear it.
For me, the end result of this writing-process is often a document where much of the language first came from the agent, but the voice is recognizably mine, and I feel a clear sense of authorship. It is still my work because of the microscopic attention I paid to every word. The marble is the agent's, but the chisel and mallet are in my hands.
> Sometimes I can find it if I keep revising, but staring at a sentence too long can entangle you in a particular linguistic frame that can be hard to escape.
This expresses quite well an experience that I have sometimes had. Your first words can trap you in a box that you know isn't right, but you can't figure out how to escape.
I understand this is a confusing time, and we all feel vulnerable to being duped. But please be a bit more circumspect with these confident accusations, because they can be genuinely hurtful. You are factually wrong here: I don't use LLMs to write or revise personal content.
I thought it was an interesting comment and I suffer from the same issue when trying to get a clear idea from my head into writing. I sacrificed some Pangram credits to reply to the parent comment, fwiw.
> I think so because (1) the writing process is an essential part of the thinking process, (2) AI writing is vague and wrong in hard-to-notice ways, and (3) writing with AI (and not labeling it as such) is rude and misleading. I’ll explain these points in more detail below, but first, a few throat clearings.
1. Agree writing process is essential, but just getting your thoughts down is the start.
2. Human writing is very often vague and wrong in hard to notice ways, but I would say it's more likely to be wrong in easy to notice ways, which is ... better?
3. I think it's rude to not use the best tools to convey the message most clearly and most respectfully of my time. I often use AI to help me write emails and nearly every time it provides more concise well structured versions of what I have to say. 99% of the time I'm just trying to convey a message clearly, that's all. When I write my draft there is often words I can delete or phrases that I can shorten. It's hard to spot but when you have AI point them out it's obvious. This is the role of an editor (check out Stephen King book On Writing which makes this point). Unfortunately I don't have an editor, but I have AI that helps trim the fat, get to the point and save my readers time.
If these tools are available to you and you don't use them, then I think that would be disrespectful. I don't take any solace thinking how much time and effort went into what someone wrote me. If I could save them time and have AI write or edit, and save me time reading it since it would be more concise and better structured, it's a win-win.
Just a couple of days ago, a customer rep at a company we partner with dropped a long doc that had basic reasoning errors once you got into the individual sections.
I would have much preferred if the rep had written a shorter document and reached out about things they were uncertain about. Instead thet wasted my and my team's time as we first tried to make sense of the document on good faith before realizing the problem.
I don't care if people use LLMs per se, but if this is functionally the result, then in this aspect of my work things would go much better if people did not use them. They are a great interface for talking to a machine, but due to the laziness they breed, they are a terrible interface for talking to other humans unless approached with a great deal of discipline.
Yes, you hit on the main problem with LLMs for writing. It doesn't cost you anything. It can do the reasoning for you and creating long messaging with dodgy reasoning.
I think you should write down your thoughts, tell it to make the points you spelled out and be concise and respectful of the reader's time and attention. When I do this, it works very well. But I agree if you just have it respond and reason it tends to often produces bad responses.
I think your point on “cutting down on words”, so to speak, is relative to personal writing style. I find that AI tends to write far more words than I personally do, which is sometimes useful, but in the opposite direction that it works for you.
This resonated with me - I’ve been avoiding things that affect my word choice since mobile messaging apps started introducing “next logical word choice” features for the exact reasons outlined here. If writing approximates thought, I don’t want some third party process constantly butting in and derailing that train, even if I’m just texting someone.
But as a thought experiment, replace “writing” in this article with “coding”, and ask yourself why the same points wouldn’t hold? Maybe it’s because I enjoyed writing essays in college, before I wrote software as a job, but I’ve never seen much difference between writing code and writing natural language. They’re both a means to an end, some lossy way of encoding thought. One is machine executable and one isn’t (arguable maybe, at this point) but I think they both suffer from the same pathologies outlined here when AI takes over, eg, “writing is thought”, “AI writing can be subtly wrong” etc.
Yet the preponderance of hot takes I see these days are along the lines of “code review is dead”, “writing code by hand is dead”. Perhaps people largely don’t see code as communication? It bothers me to imagine a world where the vast majority of textual communication produced becomes machine generated, whether that’s natural language or code.
LLM’s can be soooo derailing. Better ask a few friends to give feedback on your writing. Yes, it will take two weeks before you get any, but it is waaay more valuable than any LLM and absolutely worth the time (also gives you some fresh eyes ;) ). Dont forget to give them a gift, copy, credit.
i like to write and then workshop with the ai. i find that i have to push back on tendencies it (has less of these days) has to rewrite my work in massive ways. it also can help with writers block sometimes, or parts of a paragraph i'm not that interested in.
This is just some sort of Luddism all over again. Sorry to be blunt, but nobody cares about your proofreading, because your text will most likely be read by another AI anyway. Your boss doesn't care about what's more convenient for you either. If you generate X amount of value for the company and your colleague generates 3X, you'll soon find yourself out on the street. It's like saying you don't like Googling things and instead prefer spending the whole day at the library reading through articles and books. LLMs are incredibly powerful tools, and everyone's task right now is to learn how to maximize their productivity by leveraging their advantages and mitigating their flaws, rather than urging people not to use AI. The genie is out of the bottle, and the world will never be the same.
LLMs are incredibly powerful and useful tools, but offloading your writing to them oftentimes means you’re placing an undue burden on whoever you’re sending the generated output to.
To take it to the extreme, a “3x worker” is by definition relative, and someone malicious could achieve that by overloading a colleague with generated text that they themselves hadn’t bothered to read in full. You can stay at your baseline but be “3x” if you’re sufficiently able to sabotage your coworkers.
I’ve been on the receiving end of needing to review “documentation” that was generated almost entirely by LLM. It wasn’t anything that was done with malicious intent, but it’s still exhausting, and you need to exercise an even greater degree of caution because of the potential subtle mistakes or inaccuracies that a real human would never make.
I agree that sending someone a text you haven't read and understood is a sign of poor skills, much like forwarding someone else's message that has nothing to do with the issue being discussed. But we have to acknowledge that the market for speechwriters and corporate copywriters has existed for centuries. Surely you don't think a president is being dishonest when delivering a speech that someone else wrote for them?
Unfortunately the sad truth is that any good writing on the internet these days is going to be drowned by a tidal wave of SEO optimized slop. The other unfortunate thing is that LLMs write better than most people I know, so the number of people using LLMs to write, even when the purpose is to inform or persuade, will only grow.
I know I'm an outlier but i truly just can't imagine using an LLM for absolutely anything outside programming itself. I'm something of an expert in programming so i can spot the hallucinations and nonsense. I'm not an expert in lots of other things where i wouldn't be able to spot the problems. That and the writing is so grating and awful, the docs are so verbose yet vague. Reading the responses i explicitly asked it to generate is bad enough, i could never ask someone else to do it with something they could just generate themselves.
If you're talking about the general plan, structuring paragraphs, organizing your thoughts then I totally agree. Claude, in particular, is very hard to read.
But if English is a second language for you and you still get confused about what is the proper preposition or phrasal verb then you need AI. I suggest Grammarly.
I disagree, but think that there's a specific way to use AI to write. I use AI to get critiques on my writing, but never want it to actually produce text for me. I experimented with this by making https://www.writelucid.cc, and I really like it, I use it for all my writing nowadays.
Getting comments on specific sentences/passages and being able to debate them has been great for my thinking, and I've learned a lot about what works and what doesn't work while writing with it.
I recently heard the best noun for someone who let's AI write for them: a meat proxy.
Apart from the extremely good points made in the article; there's also whether you care that someone reads what you (or rather, the AI) has written. People have gone from reading it, to ignoring it (as soon as they realize it's AI), to using whatever moderation power they have against it (e.g. downvoting).
My 75 y/o mother wrote a kids' book about the animals on my brother's farm, and she was singing praises about how much AI was helping her. I reminded her that I would rather have her grandchildren reading _her_ words and not some clanker's words. I don't think I am alone in valuing the human soul.
People were self-conscious about their writing (quality, speed) before LLMs. AI seems like a way out of that weakness. That's just the nature of a tool: it seems to extend our ability. People are going to use it if they don't have a better option.
The only cure: be the person people can turn to instead of the machine, whether that's helping them write or helping them become better writers. And, yeah, you'll have to do that cheaper than the AI (netted against the social cost of being the guy/gal who got caught slopping).
So, I think I agree with all of this, but want to use this post to host my take on AI writing, a take I don't think I can get across on the big thread here on my post about AI writing:
I think people are stuck in the idea that the way you "use AI to write" is to let it generate alternative words for your thoughts. That's a terrible way to write. I advocate for a rule: "any word an LLM suggests to you is disqualified, even if it's better than the word you've chosen". Nobody is vigilant enough to keep LLM word selection from bleaching out their personal style.
But LLMs can do things for language (natch) that deterministic programs can't. Those things are helpful and you should consider taking advantage of them. Here's a short list of ways an LLM can potentially improve your writing:
* It can instantly spot overused words and turns of phrase, or, better still, worthless filler and throat-clearing like "just" and "very" and "it's important to note".
* It can rescue active verbs trapped inside nouns, where sentences are wrapped in zombie verbs like "make" or "reach" or "start" or "have", carbonite-frozen verbs like "decision" or "agreement" or "distortion".
* It can match the subjects of your sentences with actual characters in the action of your story or argument, flagging all the times you accidentally nail the subject of a sentence down to some random part of the scenery. It can check to see whether the new detail each sentence adds (if it adds any at all --- something else it can check) is in the stress position of the sentence.
* It can check your transitions and flag places where the openings of sentences and paragraphs are abrupt. For that matter: it can check the topic sentences of paragraphs and the flow from graf to graf.
* It can check for passive voice, but also note the (many) instances where passive is the right call for what you're trying to say.
The model never gets tired. It generally never forgets the rules. It can instantaneously diagram out a sentence and work out an accurate model of the semantics of your writing. It's doing things you cannot do with a grammar checker.
You're not going to want to act on everything an LLM flags. LLMs have an idiosyncratic sense of style (I keep saying they write every sentence like it's the headline of a magazine article). Some of these quirks of writing will be part of your voice, and you'll need to keep them.
You're not going to want the LLM to give you alternate words and sentences. That way lies Velveeta. And many times, the LLM will flag a usage issue and your response won't be to act on the suggestion, but just to rewrite the sentence or paragraph entirely --- or, better yet: just delete it, which is an awesome feeling.
This is using an LLM as a copyeditor. I've worked with professional copyeditors, and the feeling of working with an LLM copyeditor is comparable, except that the LLM is much more thorough, and I don't feel bad about ignoring it when we disagree. This style of working doesn't allow the model to infect your writing; it's just making you better informed about the words you're choosing. I think more people should try it.
Writing is pretty hard for a lot of people, maybe especially so if they are more non-verbal thinkers, and then doubly so again if one must write not in one's native language.
Depends on the writing. Writing a blog post or something is challenging because you want to have an engaging style. But I see people at work using LLMs to fill out tickets, which is the easiest thing in the world. It's just a plain description of things, it requires no skill at writing at all. It baffles me that people are using LLMs for something like that, which should take less than a minute of your time to fill out and will be more pleasant for the recipient than the slop the model produces.
Here I would maybe argue that the simplicity makes these kinds of tasks tedious (like doing taxes), so it also makes a lot of sense for people to want to just throw AI at it. Tedious, easy rote tasks can be much more unpleasant than engaging ones.
I think people massively over-rely on AI and I really worry about the consequences of this. But there is nothing baffling at all about the basic appeal, IMO.
I don't think I have to speculate why comments like yours get downvoted so much.
Anyway, I completely agree. Writing isn't something to be delegated.
The problem is not new with LLMs. Businesses have been delegating writing to idiots who don't really care for the entire history of business. Authors and musicians have traded their souls for inauthentic crowd-pleasing results for about as long.
No, the problem with LLMs is that people are all using the same generic models. This only further shows that the future of LLMs is locally trained and locally run. Personal LLMs on a personal computer. Business LLMs on-prem. We will continue see them flourish in ways that generate absolutely no money whatsoever on their own, but do add marginal value when combined with a heavy dose of human creativity.
All the clueless old farts that are still alive by this point will continue to lecture the rest of us with more condescending gee whiz "whaddyaknow" and "whodathunkit" nonsense. As if they've even had their finger on the pulse since the early 2000s.
Who cares? AI gives me lots of text and it's not like it's ever going to go away. Any time in my life in the future when I need text, I can just have AI generate it. Thinking is solved.
From TFA: "Eric Schwitzgebel writes that . . . There’s a huge cognitive difference between nodding along while reading something and actually productively generating a text. Two reasons: First, once the text is on the page, it’s easy to passively let the approximate word suffice, rather than thinking about word choice in the same effortful, active way we do when generating prose de novo. Second, as I suggested above, I doubt that human beings, even experts, have a good sense of all the factors that shape word choice -- everything they’re being sensitive to. You would have phrased it slightly differently, and even if you don’t know that, or why, a different signal is sent and received."
This is the best articulation I've seen of why simply reviewing and copy-editing does not provide remotely the same value as writing from scratch. I spent a considerable amount of time over the past two weeks reviewing and improving a work document that was the output of an LLM. Given the number of people involved and the final level of effort, I'm firmly convinced that writing it manually would have been faster and resulted in a higher quality product. Getting the wording right matters.
I really did like the author's framing there, but I think there is a different, simpler way to put it:
Why would you think that asking someone else (that is, another human) to write something (and then reviewing it) is the same thing as writing it yourself?
You may trust the other writer's opinions and knowledge, but it will not have the same tone, structure, word choice, understanding, or narrative flow as it would if you were to write it yourself.
And when it's an LLM, you should not trust it's "opinions" and "knowledge", because it does not have either of those things. The appearance of those things is just that, an appearance.
Except that there are such things as speech writers, whose job is literally to write prose in the voice of the person who’s going to read it.
And there are countless other jobs out there where people are the “voice” of another party. From those who manage social media presence, to PR firms, to copyrighters, all who put statements out on others behalf.
There was already an industry of professionals whose job it was to have others write things on our behalf, and that existed long before LLMs were a thing. What LLMs did was make that that service available much more cheaply.
Speech writers do not write in the voice of someone else, they create the public image of say a president who is inarticulate and make them seem polished and articulate, if anything they lend their voice and ideas to someone else. That’s why they are hired.
And no LLMs are not producing ‘the same thing’ as a speechwriter. If you are unable to tell the difference I’d suggest doing some a lot more reading of human books. LLMs produce stultifying pablum without coherence or style.
> Speech writers do not write in the voice of someone else, they create the public image of say a president who is inarticulate and make them seem polished and articulate, if anything they lend their voice and ideas to someone else
Nope. The exact opposite in most cases. Eg Jon Favreau (Obamas speechwriter) has said:
“As a speechwriter, your ego has to take a backseat. The goal is to make the speaker sound like the best version of themselves, rather than to showcase your own cleverness."
> And no LLMs are not producing ‘the same thing’ as a speechwriter.
I didn’t say it was the same. I said it was an existing industry that AI has stepped into and AI is taking the low end of market.
It’s just like how AI has captured the low-end of many tech roles too.
> If you are unable to tell the difference I’d suggest doing some a lot more reading of human books.
“Some a lot more”? Very ironically timed editing error there. You can bet an LLM wouldn’t have made that mistake ;)
Speech writers and shadow writers in general are writing on behalf of other people, but they're not writing the same thing. They can't! Usually they're writing something better, with the proper voice and authentic to the views of the person, but still not the same as what they would have written.
Depends on the individual. It’s usually a highly collaborative process. But in general, you’re correct that you’d expect a good writer to write better than the person who hired them.
There's knowledge in books, and it's ingested all the books, so it does hold knowledge. But I guess that's not how you mean it.
The key difference is that while an encyclopedia holds facts themselves, LLMs trained on that source material encode something more like a highly probable facsimile of those facts - the original fact was lost, LLMs are lossy, but can often be generated again with a decent level of accuracy by churning through stats about words, concepts, and relationships between them.
The whole catch is that they can often be regenerated. But LLMs (on their own, in their parametric memory - which is the result of training) don't have any conception of whether what they've generated is a real reproduction of some training material or whether they've invented something false that seemed probable based on their encoded stats. When the probability produces something contrary to what was in the training material, you get hallucinations.
They're very, very good predictive text models and can be very, very powerful when hooked up to other tools or outside databases. But its fundamentally lossy technology and all the books having been fed in doesn't guarantee all of the knowledge from those books can be spat back out.
Your criticism of LLMs also applies to humans and so implies humans don’t possess knowledge.
Humans can learn texts by heart, even lots of text (I have some expertise on that, having studied opera singing). They can reproduce these texts accurately, deterministically and repeatably. An LLM is a statistical machine. It does not know any text by heart and it is by pure chance that it sometimes reproduces existing texts verbatim.
Ant tips for doing so? I've never been able to do that, even as a kid. I remember the meaning but not the exact words.
Break it into small chunks and practice. Practice more. Put the chunks together and practice even more. Like any skill, it isn't something people are magically good at beyond minor proficiency.
Every performance- singing a song, playing an instrument, performing a stand-up routine, giving a speech, performing a theatrical role are all things that are best done from memory but require practice.
There are pneumonic tricks you can use- I've seen some people do it for tricks like memorizing the order of a deck of cards- but it's less useful for long term recital because it helps with order but not comprehension or fast indexing.
Humans sometimes don't remember things 100%. LLMs can produce certain text and be run deterministically.
I don't really understand this side of the debate other than as a gotcha tbh.
If LLM use atrophies your brain and skills that's bad. If it has a repulsive writing style that's bad.
I'm not sure what the debate about whether an AI is a statistical parrot unlike humans accomplishes. Is relying 100% on a bad human speechwriter somehow better?
It comes up frequently when discussing whether LLMs actually have intelligence or qualify as AI, especially in discussions on the path to achieving AGI.
The primary complaint seems to be that LLMs are held to a higher standard than humans, though I don't particularly buy that line of reasoning.
It's funny. I always thought that writing was meant to inform, persuade, or entertain about the subject at hand.
But in professional settings, a lot more of the informativeness is about the author, and a lot more of the persuasiveness is I'm worth your time and money. So, if the author is an LLM, and obviously so, what exactly are you informing your audience of (about yourself), and what are you persuading them to do (with your article).
I think we now know.
People using LLMs not in a gruntwork manner tend to confuse style and tone with facts and knowledge.
LLMs are great to make drafts if you give them the source materials. They're great at validation if you give them the tools. They are great at layouting if you give them linters.
Encode architecture and decisions in your workflow, then proofread what your agents have been working on. Not the other way around.
Better validation and testing means more work will transform from exhaustive decision work to automateable gruntwork.
It doesn't hold knowledge it has the ability to seive through a lossy latent representation of knowledge.
> First, once the text is on the page, it’s easy to passively let the approximate word suffice, rather than thinking about word choice in the same effortful, active way we do when generating prose de novo.
Meanwhile, me, as an English non-native speaker, ended up discussing two sentences I want send to HR for ten minutes while applying to a job.
I do believe there's generally a bias to accept something that's already written. The much bigger reason though is why you let somebody else write it to begin with.
It might just be that not thinking carefully about every sentence/wording was the exact thing that made you use AI to begin with.
> I spent a considerable amount of time over the past two weeks reviewing and improving a work document that was the output of an LLM.
I had the same experience with texts where I have a very detailed expectation of the desired result. This is just a general limitation. For code, there's the saying "the precise description of the solution is already the code". Describing X is a simplification of X, oftentimes it's fine guess the gaps. But when it's not, it didn't help to describe X, you have to manifest X itself.
A workflow I've recently discovered for myself that provides a kind of middle ground: I'll ask the LLM to write a first draft in a language that isn't the one the piece should ultimately be in. Then I'll use the LLM's first draft as a blueprint, to write a first draft myself in the intended target language. This fights my brain's temptation to just shut off, and forces every word choice to actually be mine. Then I'll hand that over to the LLM for editorial suggestions and iterate from there.
Excellent idea, I'll try this. And for those who don't speak a second language other than English, you could use leet.
Wow, love this! Could even match with a language I am actively learning so it helps that as well!
I use an antagonistic agent to review and refute its findings, works great.
Writing is thinking. In many situations where we are called on to write, what is actually needed is thought.
It's probably annoying the way I oversimplify things, but you reminded me of something I read somewhere.
Some Jazz musician was asked to define Jazz.
He couldn't really, his answer was something like: 'I don't know. But I'll know when I hear it'.
To add something to the discussion directly: thinking requires vocabulary. Vocabulary is the currency of thought.
You can usually express an idea with a few thoughts, or many. The audience, and amount of details chosen should always be kept in mind. Writing helps you to remember vocabulary and word choice when expressing ideas.
While I mostly agree with you, what is the real difference between an editor saying, 'when generating prose de novo' sounds pretentious to an American audience so let's use 'when writing from scratch' instead, versus an LLM giving the same advice?
Not that it's necessarily better but it's the kind of thing an editor might pick up and so why would you reject the advice if it's a machine giving it rather than a human?
TFA is about having AI do the writing, not about having AI suggest editorial improvements. The parent comment argues against having the AI write the text and having the human merely review it, not against having the AI review human-written text (and the human deciding which of the AI’s suggestions they might apply).
See also this existing comment: https://news.ycombinator.com/item?id=49768564, which I completely agree with and which is complementary to the above root comment.
True, but I was responding to the comment rather than the article.
I expanded my comment since.
Sure , but that seems like you've agreeing that an AI suggestion is as good as an editor suggestion.
What do you see as the difference?
I don’t see any conflict with what the root comment wrote. There are probably some differences between human and AI editorial suggestions that could be discussed, but that wasn’t the topic here.
It was the topic I was discussing with the person I replied to.
If you want to discuss something else just let me know.
Same flow as translating from your native tongue to a foreign language.
Going from foreign to native is easier than native to foreign. The latter requires completely different brain paths and a lot more understanding of the language to actually get to something correct.
Use AI to write things for you to read that you wish someone else had written. ‘Give me a summary of the research on this topic’; ‘Write a report on this data to help me make a decision’. ‘Take this transcript of a meeting and write me the email it could have been’.
Don’t use AI to write things that you are producing for someone else to consume.
I think that's somewhat reasonable, but not always: the effort of researching and of writing yourself will teach you better then reading something someone else wrote, especially an LLM that will get things wrong in surprising ways that a human would not.
I'm not saying you should never do this: our time is valuable, and we shouldn't spend it on things that are not genuinely worthwhile to us if we can help it. But we're still losing something by having an LLM write for us, even if the intended audience is just ourselves.
This is 100% my take. It's nice and easy.
I do not personally even think AI is useful for editorial purposes; I would rather read and re-read my own text and edit it down (which, despite my overlong comments here, I actually do), than have someone else do it. The reason is simple: editing is my search for the clearest way to express my thoughts.
I tend to turn grammar checkers off, but leave spelling checkers on, because they catch typos and a handful of my spelling gremlins like "liaise".
I am someone else, if AI can write something that I want to and can consume then it can do the same for other’s consumption. Perhaps a disclaimer would be nice, “co-written with ____” but I see no difference between the two though of course I understand what you are trying to convey conceptually
When I seek input (whether passively or actively; it doesn't matter) from another person, then I generally want that person's input. I want their words.
They can be broken and jumbled, or terse AF, or combine to be Pulitzer-quality prose. Either way: I want to consume their own unique human expression of a concept.
There's value in raw human expression, including in the missteps.
If I instead want the regurgitated waggyings of a bot, then: I know how to get that on my own.
We've all got web browsers and pocket supercomputers. We've all (well, most of us) been alive and aware of LLMs since their recent rise from infancy.
It's a no-brainer for us to paste some paragraphs into our favorite chatbot and get a summary or an artificial expansion or whatever else we wish to have. If that's what our goal is, then we can do that on a whim -- and we can still retain the original expression.
But doing it on our behalf is deleterious, unsettling, and unhelpful. It has negative value to the beholder.
> When I seek input (whether passively or actively; it doesn't matter) from another person, then I generally want that person's input. I want their words.
There's a (probably subconscious) reason for you wanting that (and I just posted this elsewhere on a different story, but it bears repeating): https://news.ycombinator.com/item?id=49767867
There is a higher tolerance for LLM-isms if you asked an LLM to write something for you
Elsewhere it’s painful and irksome when you didn’t ask for it
this is where I perhaps differ from other commenters, I do not have higher tolerance, regardless of whether I asked it or am reading something someone else posted, fully or partially written by LLM.
> this is where I perhaps differ from other commenters, I do not have higher tolerance, regardless of whether I asked it or am reading something someone else posted, fully or partially written by LLM.
Perhaps English is not your first language? To people who are well-read, LLM-ism grate on their nerves like the high-pressure sales pitch you hear on 3am shopping channels.
But the reader cannot be sure they ARE understanding what you are trying to convey. Maybe the LLM added something you never thought of, maybe it made form to your various thoughts in a way that isn't actually true to the way you would have communicated them if you did it yourself.
The exercise of writing is an important step of actually understanding your own thoughts properly. Readers are reading to get a sense of your own experience and ideas on the topic, if it turns out you didn't actually write the final thing yourself they'll always feel a sense – in whatever way – of being conned by the author. Intentional or not.
Technical documentation, where what the reader wants to come away with is the clear facts on what something does and how to use it/whatever, is a different matter. But even then editing is key – overlong text, unnecessary information, these will again fight against what the reader is there to get.
With that said, if you do want to do it, a disclaimer is at least upfront and may well show there's no intentional attempt to deceive.
> Technical documentation, where what the reader wants to come away with is the clear facts on what something does and how to use it/whatever, is a different matter.
I don’t agree with this one. Writing technical documentation can be a slog, but quite often you will discover very annoying design decisions in the process. Like you’ll write a tutorial for a CLI and realize that actually some flag makes no sense, or some default should be changed.
This also goes for dogfooding. If you’re using AI to write and run the tools, try to use them yourself and see how nice they are. AI will happily run whatever jank it’s working with.
Docstrings perhaps, but even then I find myself editing heavily to avoid cruft sneaking in.
That's logical, yes. If person A wants to read a summary (even a generated one) outlining the state of modern space dinosaur research, or whatever, and person B also wants the exact same thing, then person B should be happy to read the summary that person A prompted.
The same is true if the thing they both want is a stale cupcake bought from a gas station. But this still doesn't mean that person B wants to be presented with person A's stale thing, because the overhead of the social transaction isn't worth the object. Which is to say, it makes a terrible gift.
The words of a machine are not of interest to me. I want to hear what you, the human, have to say, imperfections and all. Reading something written by an LLM is actively annoying, and it's not very respectful to inflict that upon others.
so you never use LLMs like at all?
There's a difference between me getting AI to write something I want to read, and getting it to write something you want to read. The difference is that I know me, but I don't know you.
The other difference is that, if I get AI to write something for me, I expect something AI-written. If I am writing something for you, you expect something that I wrote, not something that AI wrote.
Moreover, when you wrote it yourself, you are intellectually responsible, you can defend it, while if the LLM wrote it for you, nobody is really responsible: The LLM didn't write it "out of its own motivation", it was merely extrapolating your prompt; and you only provided the prompt, not the extrapolation.
So the intellectual responsibility is diluted to a substantial degree: The text is not the opinion of anyone who can be expected to honestly defend it. It's bullshit in the technical sense.
I just explained why this is a bad idea: https://news.ycombinator.com/item?id=49767867
TLDR: Communication from a human is not only the material being communicated, it's also a signal that the human understands the message they are sending you. When you pass that message via an intermediary (regardless of whether the intermediary is a human or an LLM), the receiver is never sure that what they received is what you intended.
It's doubly worse when you review it, notice a few additions and think "Great - that makes me sound smart and knowledgeable about this topic", because the receiver now has an inaccurate view about your skill.
At any rate, using an LLM to construct a message for someone else to read is basically sending the wrong signals about your skill levels.
(This probabyl explains why so many people are so ready to use an LLM to write the material - in their minds, it makes them sound smarter than they actually are, but as we've seen from people on this forum vigorously defending the practice with "pangram is wrong" and similar nonsense, they themselves are not skilled enough to recognise slop)
> TLDR: Communication from a human is not only the material being communicated, it's also a signal that the human understands the message they are sending you. When you pass that message via an intermediary (regardless of whether the intermediary is a human or an LLM), the receiver is never sure that what they received is what you intended.
100% right. The fundamental point here is:
- it's OK to be wrong
- it's OK to be wrong and not know
- it's a little less OK to be wrong when you should have known
- it's much less OK to be wrong when you haven't put any effort into being right
- it's far less OK to be wrong when you are pretending to know either way
Relying on AI writing risks pushing you further down the path to worse kinds of wrongness.
"AI writing is vague and wrong in hard-to-notice ways" This one felt to me, at work I lost a lot of time re-reading and correcting AI prose that was launched to "summarize" a collective white paper, and where subtlety and nuance disappeared in ways that harmed the initial ideas/arguments I was writing about, but in sneaky ways.
AI is very bad at properly handling statements that make heavy use of vague quantifiers (e.g. "some", "most") and also commits a lot of pretty serious logical fallacies. It is also bad at handling subtle logical negation, generally.
One of the most egregious negation issues I run into a lot is when I (or someone) makes a statement of the form: "not X" or "X is thus not true", and the AI then proceeds to interpret or summarize this as 'whatever is the opposite of X is the case'". This will cause it to go down a useless path investigating or disputing the opposite of X, which generally has no relevance or bearing on anything.
It also often very harmfully will replace your carefully chosen words with weirdly specific academic operationalizations or formalisms, then again waste huge amounts of text refuting / showing "problems" that result from that formalism, all of which again have no bearing or relevance on the original statement. An example would be you saying something like "intelligence, generally, must surely explain some of the differences in X", and then it will go "actually IQ does not correlate with X", unless you specifically tell it not to conflate psychometric IQ with intelligence generally.
Sometimes this is helpful, but the more specific / technical the domain, the more often you specifically have to prevent it from going down stupid paths that should be obvious given the expert context and wording, because it can seem almost hungry to try to catch you in some kind of insipid 'gotcha'. Much of these issues often clearly arise immediately from the first-pass "reword what the user said" part, given the reasoning traces.
Most of us have worked our whole lives to be writers in some domain: Code, documentation, specifications, product requirements, corporate communication, etc.
If we don't want to be writers, then we have to be editors. And editing is an entirely different job and it's not an easy one. In many ways it's harder.
Especially when LLMs love writing novels when all we need is a short story or less.
A big part of the reason I'm deeply resentful of AI is that I want to be an artist not an editor
Can you be one without the other? Serious question – I’ve written five books, and the first draft was like 5% of the work.
Instead of asking AI to write something substansive for you, ask it to critique your writing. Then use your own judgement as to which bits of advice to follow and which to ignore.
The LLM will always give you a full rewrite — don't use it. It always does too much, and persuading it to tone it down is a constant battle.
AI can't critique because it has no critical faculty. It can only offer critique-shaped language. That may be useful for some, though.
That’s not entirely true. High end models will point out things that are misrepresented or plain wrong, and this will often be accurate.
Potayto, potahto
It blows my mind that anyone would ever use an LLM to write, if communication is their goal.
LLM writing takes your prompt and adds stuff you didn't write, burying your meaning and making it harder for the reader to get your message.
Way more effective to just publish the prompt.
Of course, it's great if you're just trying to fill space or satisfy demands of a bullshit job.
Here's how I use it to write. I dump an unedited stream of consciousness into the prompt. Sometimes it's just me having a conversation with myself thinking through what I actually want to say. Then, I have it ask me to clarify anything that's unclear. Then I have it write me a rough draft. I read that and then just rewrite most if not all of it while I edit the output. I use this workflow for work stuff or dealing with businesses in personal life. I'd never use this to generate a letter or email to a friend or colleague.
Definite ‘work stuff’ because most work stuff is read by colleagues, otherwise why was it written?
Interesting. Does your output end up being shorter than your input?
Hmmm, I've never actually checked that. Just had codex run the numbers excluding the boilerplate in some of the documents output is on average ~13% shorter than input, I only checked the ones that were all of the same type (proposals) since it was easier to have codex run the numbers on that.
I compared at final edited output less boilerplate vs sum(prompts).
If I had to guess I'd say far more people are better at editing than they are at writing de novo. My approach probably limits the creativity of the output but I'm not writing a novel.
I can't help but think that some of the arguments here apply equally well to writing _code_ as they to to other substantive writings.
Writing code is thinking, AI code is often vague and wrong in hard-to-notice ways, and the (human) reader of code is the one that pays the cost for this later.
The cost/benefit analysis may still work out differently for code though...
I never use AI for writing but do so regularly for code. They are quite different to me.
My reason: code can be checked objectively. I can run it and confirm it works. I don't get attached to it. I don't feel pride in it (even when I write it by hand). Code just is. It's lifeless, inert, and entirely replaceable.
How do I do the equivalent for prose? How can I tell if my words "work"? Do they clearly convey my ideas to the intended audience? There's an element of subjectivity here forces me to identify personally with the prose.
Code has no such personality. I don't tie my identity or ego to code the same way I would an essay.
> My reason: code can be checked objectively. I can run it and confirm it works.
Running the code only confirms that it works with the precise input, in the precise environment, under the precise circumstances you run it under. It doesn’t ensure that the code is correct. Thinking through the code, on the other hand, lets you consider all possible cases. It’s the difference between experiment and (mathematical) proof.
For an objective correctness proof, using a formal language is indispensable.
Code is a formal language, and can be partially (or rarely fully) verified by a typing system. There is no equivalent in natural language.
Even thinking through the code is not complete because our brains make mistakes. They also make assumptions, thousands of them, that you don’t know about. It’s tricky, because every thought you have has an entire sea of preconceived notions and knowledge behind it, which you can’t identify. In that way, it’s similar to an LLM.
In practice, for code, testing code is the way to go. Formal proofs work, too, but the barrier to entry is high and it’s overkill for most business applications. LLMs can be very good at writing tests, if you read the test thoroughly and analyze them.
Programs must be written for people to read, and only incidentally for machines to execute. -- Hal Abelson
I understand the distinction you're drawing, but in my experience, it gets a lot blurrier as a project evolves. Caveat that I use LLMs all day and have for quite some time, so I'm coming from a positive perspective.
The danger for me in LLM code is the same as in writing, it's just that I'm not typically writing at the same scale as when I'm building something. The final piece when I'm writing is usually a message or a 1,000 word article at most. So I'm naturally going to analyze it quite intensely, because I can afford to. And I don't really use LLMs for this at all. I use them for things around writing (research, interrogating ideas, situationally specific stuff, mapping, visuals, publishing, etc.) And the code equivalent to an essay or message would probably be something like a single script, or the sort of thing I'd write as example code when I'm teaching. In those settings, again, LLMs can be helpful, but I'm still going to be really opinionated at a highly detailed resolution.
But a codebase is more comparable to a novel than an essay. Or more directly, the writing in a codebase is usually the documentation, which grows commensurately with the codebase. And the real LLM risk here is the drift that can happen over the course of many epics or "chapters" as the LLM writes "code that works but is imprecise and probably shouldn't work this way" or introduces weird new terminology that neither of us can precisely define. Worse, this usually becomes obvious down the line, and I have to parse through the verbose constructed world the agent has created to trace the issue back. That's a big cognitive tax, because I'm holding these weird parallel worlds of "How did the LLM's alien brain get here within the bounds of the contracts" and "What do I really want this to look like".
So I think it's fundamentally the same phenomenon, and we're all developing our skills around working with it in real time.
> I don't feel pride in it (even when I write it by hand).
Hard for me to imagine. You feel no pride in using a tool to accomplish a goal?
> Code has no such personality. I don't tie my identity or ego to code the same way I would an essay
Code certainly does have a personality. When working with teams for a while you can absolutely get a sense for which person wrote what code in a codebase, just by subtle little tells.
You may not tie your identity or ego to it, bully for you, but for me I take a lot of pride in writing clear and maintainable code that contributes to big projects in meaningful ways.
Maybe the problem with software is there's too many people who treat writing code as a mere means to an end, instead of a very important part of the process.
Well, there's some nuance. I'm not really proud of my code. I'm proud of the software that gets built with it.
Code is just sitting in a git repo somewhere, not necessarily running. That's a big distinction for me. Consider that code volume has increased 14x on github this year, but we see nowhere near that increase in the actual usable software. Code is cheap and getting cheaper. Running software ain't.
Another way to put it: I'm only interested in code so far as the value it provides. That value, not the code itself, is the source of pride. If I can provide similar value without any code at all, I'd gladly do so.
But code is the mean to get an idea across to someone. It’s the difference between “I think this can work” and “I’ve worked on this and it does work”. I’m part of the OpenBSD mailing list and it quite nice someone sharing a diff for an idea or experiment. It may not get merged, but it’s better than realm of deliberations.
I was about to reply with the exact same sentiment. How often may be different from writing but writing code is often thinking.
> I can't help but think that some of the arguments here apply equally well to writing _code_ as they to to other substantive writings.
All of the arguments here apply to writing code; Why would the quality of writing differe when writing prose vs code?
I cannot read LLM slop prose without getting mentally fatigued trying to figure out what it is saying, and I've discovered that the quality of the code it produces has the same effect on me.
If I may propose a broader rule, that I think this post presents a special case of;
Only ever use AI to make yourself think harder, and more.
And also to think in different directions than you would have gone in isolation. Regardless, this is a great heuristic / rule that I will be sharing and keeping in mind.
This is brilliant, I'll be using that. Thanks.
I don't know what everyone else is doing but I've certainly been writing a lot more often because of my interactions with AI. Maybe some people used to write more but have been leaning on AI so they can write less but I don't think that's universally the case.
I agree that being lazy about writing and simply using a short prompt or list is detrimental if you are replacing your own output but you don't have to do that and you don't have to accept any of the output either. The best chats I've had usually start with a large amount of my own writing up front, thinking through the idea, listing several alternative ideas, asking a lot of questions, jotting down related topics, etc. and then reading the resulting output and critiquing it, asking for clarification, doing my own research on it, even just discussing it with the AI and doing this for a number of turns until I feel like I've exhausted the topic in the chat. I then often take what I've shaped in my own mind from that process and write something myself, either that or take the best parts of the output and edit, rearrange, reinterpret, and add to it in order to produce something.
I guess it all comes down to whether you are actively engaging with the material, regardless if that material is the result of a Google search, pulled from a book or generated by AI. If you just read it or copy paste it and don't engage with it and think about it then it doesn't do much good.
Communication between two people is hard enough. Involving another actor as an intermediary rarely helps. It quickly becomes a game of telephone. Doubly so if the intermediary is an LLM.
I built a place to continue the disagreement around Grunewald's article as a structured one-on-one discussion. Its an experiment in finding better ways to have deeper conversations online if anyone here wants to try it: https://hwgo.io/p/7
Good or bad please let me know what you think of this format!
Is editing really any less 'thinking' than writing the original copy? I don't use LLMs to generate prose but I can compare to writing code. Let's say I use an LLM to write code to some specification. After it meets the spec I go over everything with a fine-toothed comb - tweaking things to a style I like better, rewriting pieces or interfaces that I think would be better a different way, etc. Is the final work product any different than if I wrote it from scratch without absolutely any LLM use? Personally I would say no, if I do enough editing to be perfectly satisfied with it then it's the exact same output in the end.
I think it's hard to generalize because people have different approaches to writing, but to me, there's a marked difference between iteratively polishing my own text and editing for someone else (and that includes LLM output produced by my own prompts).
In the second case, these sentences were never a "voice in your head", and you never invested any effort into shaping them, so it feels more like janitorial work / drive-by nitpicking than craftsmanship.
I offended a coworker by asking him to not send me AI written things because they’re wrong in subtle ways and overly verbose. I’m glad someone else shares my opinion.
Any form of thoughtful writing is usually meaningful when it's deeply human. Having said that, I've found it helpful to use AI as a sounding board and planner before actually getting into the writing. For me, the hardest part of writing a piece is always in getting the actual writing started, and AI has helped me unblock that a few times.
I don't find AI helpful for truly constructive or creative prose, it devolves to the median output very quickly and it's almost painful to hear in my inner voice when I read out an AI-ified idea. The "not this but that" formulation is especially bad or cloying attempt at pretending to refine an idea or lightly correct something to start from a position of authority, very annoying.
That said, I have a lot of situations at work where I am asked to simplify something I am an expert in or convey something for a different audience, particularly as I prepare for presentations and I do find it helpful in helping me step down my writing or work. YMMV.
I also skip words a lot and can miss that even after 2-3 editorial passes and it's very helpful at that vs. say normal spellcheck.
One way I use AI to help me write that I've not seen anyone else talk about is this: I'll say "give me multiple versions of saying xyz". I'll almost never use any of what it outputs verbatim, but seeing it said in different ways helps me clarify how I'd prefer to approach writing it.
When I was a graduate 20 years ago, the partner would sit with a print out of every word I wrote and a red pen marking every change required, as the partner knew anything that was written in his name had to be exactly what he would have said.
Same needs to go for LLMs, leverage them for productivity both make sure you are owning their output.
Every time there are some articles like this, some people replied with: "Yeah I agree, but I use LLMs 'this way' and my way doesn't suffer from the problem mentioned." I'm not saying these people are wrong but it's interesting that people always think their way of using LLMs is the best (and other people are using LLMs in the wrong way).
Welcome to the internet :)
I’ve found Gemini and Codex can do pretty well by just telling them to follow my writing style, be brief, be less robotic, be professional, be semi-professional, use less technical jargon, no em-dashes, etc. depending on the task. But, I only use that for short summaries of things I am deeply familiar with to send a slack message about or similar. I also edit before sending. It takes maybe a couple of prompts to get 90% of the way there.
Claude on the other hand is atrocious at sounding human, it’s like it’s deliberately following LLM -speak stereotypes no matter how you tell it to write. I finally told it to stop using “Claudeisms,” which gave better though not great results.
That's a lot of qualifiers for one title
Irony an article written by AI, wants to say do not use AI for writing. Following sentence looks like written by an AI, human do not use double dash, it seems AI instead of emdash wrote double dash;
“I think you should almost never use AI to write -- that is, to do the thing you’re doing when you type words on a page -- whether for a blog post, a research report, a memo,”
This absolutely doesn't look like AI writing and this is a pretty weird signal to use.
I have a much better theory: as it happens, Substack automatically converts "--" to em-dash as you type. I'm guessing the author knew that and wanted to use this feature, but composed the article in another editor and then copied-and-pasted the whole thing. Because of how Substack implements the feature, the substitution doesn't actually kick in on Ctrl-V, so all these "--" remained in place.
Haha, I do actually use double dash, as I write in org-mode. I then convert that org-mode text to Markdown and copy and paste the Markdown into the Substack editor, which preserves the double dash. In case it needs to be said, the article is 100% human-written.
In some of your other writing can see use of emdash probably thats also intentional, emacs might have very special key combo for it to use unicode emdash instead of ASCII “-“. I commented about double dash (“--“) because because thats what LLM does when asked not to use emdash for writing.
“Another promising approach is to conduct on-site inspections regularly—say, every six months— to make sure chips are not just installed to fool inspectors and then removed and smuggled, as has happened before. However, conducting on-site inspections at this scale would likely require additional budget for export enforcement, or for US authorities to implement a third-party export auditor program.”
Yes, for that newsletter, where I write in my day job as a policy researcher, we use em dashes. For my personal writing (example: https://www.lesswrong.com/users/erich_grunewald), I use double dashes.
Double dash has long been the way to clearly indicate in plain ASCII text an em dash as distinct from a simple hyphen. So I think the author here is here emphatically reclaiming the use of the em dash while pointedly not using the Unicode character so beloved by LLMs.
I think about this a lot. I started my teenage-to-young-adult life in the literary world as a poet who loved programming, and made a living as a ghostwriter. At some point, I fell in love with mathematics and wound up in ML in research/engineering for the last 8 years or so. So, I've thought a lot about writing and ML and their intersection.
I think one of the underappreciated things about writing "substantive" work is that the work you see at the end isn't the first attempt. And I don't mean the first draft of the piece. I mean that almost always, writers iterate on the same topic many times, either with complete published pieces or abandoned drafts or even just conversations and sessions of unproductive daydreaming. It's a cliche that your best work typically also comes out fastest, but it's not because of divine inspiration, it's because you've whittled the big idea you actually care about down so much in your mind that you instinctively know exactly how to write it.
My experience has been that for people who don't work this way or don't write a lot, LLMs can give them this incredible feeling of leaping straight from inkling to "substantive" writing. And because they haven't built up those muscles or "taste", they don't immediately recognize that it's imprecise and hard to follow.
That's not to say they're not brilliant in their own right, just that they haven't spent a lot of time on this particular thing. Sort of like a very gifted programmer who doesn't have a ton of experience yet (speaking as someone who is gifted at nothing and frequently has to do things they're inexperienced at).
So I don't think the problem is that LLMs just write bad. It's that LLMs are so wonderfully powerful that they allow you to confidently leap forward to create something that is a little beyond your experience.
And that's why I have different reactions to heavily AI generated writing. When it feels like marketing at scale, it grosses me out. But when I feel like it's just someone who is excited to write an idea and maybe doesn't have a lot of experience doing it, I'm not judgemental. My hope is that it makes them more excited about writing, and that trying to make their next piece better will lead them inevitably to start thinking about where the last piece fell short. And if there's some slop along the way, eh, I'm not compelled to read it.
This advice is too coarse. I believe LLMs have made me a better writer of technical reports. I have a keen ear for language, and I can usually tell when a sentence or a paragraph is orbiting around a point rather than landing on it firmly. Some writers insist that unclear prose is dispositive proof of unclear ideas -- a sign the writer has not thought things through. My experience is different: I often have a clear idea in mind that I cannot quite find the right idiom to express. Sometimes I can find it if I keep revising, but staring at a sentence too long can entangle you in a particular linguistic frame that can be hard to escape.
LLMs are helpful because their idiom-list is vast and they are indefatigable and infinitely patient. You can keep iterating on a sentence and it will keep giving you fresh takes. Eventually, whether through careful steering or brute-force iteration, it will come up with something that has the right resonance. Like a word on the tip of your tongue, you recognize it when you hear it.
For me, the end result of this writing-process is often a document where much of the language first came from the agent, but the voice is recognizably mine, and I feel a clear sense of authorship. It is still my work because of the microscopic attention I paid to every word. The marble is the agent's, but the chisel and mallet are in my hands.
I don't want to sound offending, but if that is your own writing style, it makes sense why AI writing resonates with you.
> Sometimes I can find it if I keep revising, but staring at a sentence too long can entangle you in a particular linguistic frame that can be hard to escape.
This expresses quite well an experience that I have sometimes had. Your first words can trap you in a box that you know isn't right, but you can't figure out how to escape.
You say the voice is clearly yours, but that's an illusion. I can recognize your comment as AI written.
I understand this is a confusing time, and we all feel vulnerable to being duped. But please be a bit more circumspect with these confident accusations, because they can be genuinely hurtful. You are factually wrong here: I don't use LLMs to write or revise personal content.
I thought it was an interesting comment and I suffer from the same issue when trying to get a clear idea from my head into writing. I sacrificed some Pangram credits to reply to the parent comment, fwiw.
I can see why you might think that, but Pangram says otherwise.
https://www.pangram.com/history/3d55b442-d151-4490-9c71-28d0...
> I think so because (1) the writing process is an essential part of the thinking process, (2) AI writing is vague and wrong in hard-to-notice ways, and (3) writing with AI (and not labeling it as such) is rude and misleading. I’ll explain these points in more detail below, but first, a few throat clearings.
1. Agree writing process is essential, but just getting your thoughts down is the start.
2. Human writing is very often vague and wrong in hard to notice ways, but I would say it's more likely to be wrong in easy to notice ways, which is ... better?
3. I think it's rude to not use the best tools to convey the message most clearly and most respectfully of my time. I often use AI to help me write emails and nearly every time it provides more concise well structured versions of what I have to say. 99% of the time I'm just trying to convey a message clearly, that's all. When I write my draft there is often words I can delete or phrases that I can shorten. It's hard to spot but when you have AI point them out it's obvious. This is the role of an editor (check out Stephen King book On Writing which makes this point). Unfortunately I don't have an editor, but I have AI that helps trim the fat, get to the point and save my readers time.
If these tools are available to you and you don't use them, then I think that would be disrespectful. I don't take any solace thinking how much time and effort went into what someone wrote me. If I could save them time and have AI write or edit, and save me time reading it since it would be more concise and better structured, it's a win-win.
Just a couple of days ago, a customer rep at a company we partner with dropped a long doc that had basic reasoning errors once you got into the individual sections.
I would have much preferred if the rep had written a shorter document and reached out about things they were uncertain about. Instead thet wasted my and my team's time as we first tried to make sense of the document on good faith before realizing the problem.
I don't care if people use LLMs per se, but if this is functionally the result, then in this aspect of my work things would go much better if people did not use them. They are a great interface for talking to a machine, but due to the laziness they breed, they are a terrible interface for talking to other humans unless approached with a great deal of discipline.
Yes, you hit on the main problem with LLMs for writing. It doesn't cost you anything. It can do the reasoning for you and creating long messaging with dodgy reasoning.
I think you should write down your thoughts, tell it to make the points you spelled out and be concise and respectful of the reader's time and attention. When I do this, it works very well. But I agree if you just have it respond and reason it tends to often produces bad responses.
I think your point on “cutting down on words”, so to speak, is relative to personal writing style. I find that AI tends to write far more words than I personally do, which is sometimes useful, but in the opposite direction that it works for you.
PROMPT: make this more concise while conveying the same message
I think “cutting down on words” depends on personal writing style. AI tends to write more than I would, so I usually have the opposite problem.
I use AI to write all the time. It can write better than I can. This article is wayyy to pretentious.
This resonated with me - I’ve been avoiding things that affect my word choice since mobile messaging apps started introducing “next logical word choice” features for the exact reasons outlined here. If writing approximates thought, I don’t want some third party process constantly butting in and derailing that train, even if I’m just texting someone.
But as a thought experiment, replace “writing” in this article with “coding”, and ask yourself why the same points wouldn’t hold? Maybe it’s because I enjoyed writing essays in college, before I wrote software as a job, but I’ve never seen much difference between writing code and writing natural language. They’re both a means to an end, some lossy way of encoding thought. One is machine executable and one isn’t (arguable maybe, at this point) but I think they both suffer from the same pathologies outlined here when AI takes over, eg, “writing is thought”, “AI writing can be subtly wrong” etc.
Yet the preponderance of hot takes I see these days are along the lines of “code review is dead”, “writing code by hand is dead”. Perhaps people largely don’t see code as communication? It bothers me to imagine a world where the vast majority of textual communication produced becomes machine generated, whether that’s natural language or code.
LLM’s can be soooo derailing. Better ask a few friends to give feedback on your writing. Yes, it will take two weeks before you get any, but it is waaay more valuable than any LLM and absolutely worth the time (also gives you some fresh eyes ;) ). Dont forget to give them a gift, copy, credit.
"I think"
You do not.
i like to write and then workshop with the ai. i find that i have to push back on tendencies it (has less of these days) has to rewrite my work in massive ways. it also can help with writers block sometimes, or parts of a paragraph i'm not that interested in.
This is just some sort of Luddism all over again. Sorry to be blunt, but nobody cares about your proofreading, because your text will most likely be read by another AI anyway. Your boss doesn't care about what's more convenient for you either. If you generate X amount of value for the company and your colleague generates 3X, you'll soon find yourself out on the street. It's like saying you don't like Googling things and instead prefer spending the whole day at the library reading through articles and books. LLMs are incredibly powerful tools, and everyone's task right now is to learn how to maximize their productivity by leveraging their advantages and mitigating their flaws, rather than urging people not to use AI. The genie is out of the bottle, and the world will never be the same.
LLMs are incredibly powerful and useful tools, but offloading your writing to them oftentimes means you’re placing an undue burden on whoever you’re sending the generated output to.
To take it to the extreme, a “3x worker” is by definition relative, and someone malicious could achieve that by overloading a colleague with generated text that they themselves hadn’t bothered to read in full. You can stay at your baseline but be “3x” if you’re sufficiently able to sabotage your coworkers.
I’ve been on the receiving end of needing to review “documentation” that was generated almost entirely by LLM. It wasn’t anything that was done with malicious intent, but it’s still exhausting, and you need to exercise an even greater degree of caution because of the potential subtle mistakes or inaccuracies that a real human would never make.
I agree that sending someone a text you haven't read and understood is a sign of poor skills, much like forwarding someone else's message that has nothing to do with the issue being discussed. But we have to acknowledge that the market for speechwriters and corporate copywriters has existed for centuries. Surely you don't think a president is being dishonest when delivering a speech that someone else wrote for them?
>Surely you don't think a __president is being dishonest__ when delivering a speech that someone else wrote for them?
Oh, but I would!
Unfortunately the sad truth is that any good writing on the internet these days is going to be drowned by a tidal wave of SEO optimized slop. The other unfortunate thing is that LLMs write better than most people I know, so the number of people using LLMs to write, even when the purpose is to inform or persuade, will only grow.
I know I'm an outlier but i truly just can't imagine using an LLM for absolutely anything outside programming itself. I'm something of an expert in programming so i can spot the hallucinations and nonsense. I'm not an expert in lots of other things where i wouldn't be able to spot the problems. That and the writing is so grating and awful, the docs are so verbose yet vague. Reading the responses i explicitly asked it to generate is bad enough, i could never ask someone else to do it with something they could just generate themselves.
Define "use AI to write".
If you're talking about the general plan, structuring paragraphs, organizing your thoughts then I totally agree. Claude, in particular, is very hard to read.
But if English is a second language for you and you still get confused about what is the proper preposition or phrasal verb then you need AI. I suggest Grammarly.
I disagree, but think that there's a specific way to use AI to write. I use AI to get critiques on my writing, but never want it to actually produce text for me. I experimented with this by making https://www.writelucid.cc, and I really like it, I use it for all my writing nowadays.
Getting comments on specific sentences/passages and being able to debate them has been great for my thinking, and I've learned a lot about what works and what doesn't work while writing with it.
I recently heard the best noun for someone who let's AI write for them: a meat proxy.
Apart from the extremely good points made in the article; there's also whether you care that someone reads what you (or rather, the AI) has written. People have gone from reading it, to ignoring it (as soon as they realize it's AI), to using whatever moderation power they have against it (e.g. downvoting).
My 75 y/o mother wrote a kids' book about the animals on my brother's farm, and she was singing praises about how much AI was helping her. I reminded her that I would rather have her grandchildren reading _her_ words and not some clanker's words. I don't think I am alone in valuing the human soul.
This has to be the most toxic topic ever
People were self-conscious about their writing (quality, speed) before LLMs. AI seems like a way out of that weakness. That's just the nature of a tool: it seems to extend our ability. People are going to use it if they don't have a better option.
The only cure: be the person people can turn to instead of the machine, whether that's helping them write or helping them become better writers. And, yeah, you'll have to do that cheaper than the AI (netted against the social cost of being the guy/gal who got caught slopping).
So, I think I agree with all of this, but want to use this post to host my take on AI writing, a take I don't think I can get across on the big thread here on my post about AI writing:
https://news.ycombinator.com/item?id=49747070
I think people are stuck in the idea that the way you "use AI to write" is to let it generate alternative words for your thoughts. That's a terrible way to write. I advocate for a rule: "any word an LLM suggests to you is disqualified, even if it's better than the word you've chosen". Nobody is vigilant enough to keep LLM word selection from bleaching out their personal style.
But LLMs can do things for language (natch) that deterministic programs can't. Those things are helpful and you should consider taking advantage of them. Here's a short list of ways an LLM can potentially improve your writing:
* It can instantly spot overused words and turns of phrase, or, better still, worthless filler and throat-clearing like "just" and "very" and "it's important to note".
* It can rescue active verbs trapped inside nouns, where sentences are wrapped in zombie verbs like "make" or "reach" or "start" or "have", carbonite-frozen verbs like "decision" or "agreement" or "distortion".
* It can match the subjects of your sentences with actual characters in the action of your story or argument, flagging all the times you accidentally nail the subject of a sentence down to some random part of the scenery. It can check to see whether the new detail each sentence adds (if it adds any at all --- something else it can check) is in the stress position of the sentence.
* It can check your transitions and flag places where the openings of sentences and paragraphs are abrupt. For that matter: it can check the topic sentences of paragraphs and the flow from graf to graf.
* It can check for passive voice, but also note the (many) instances where passive is the right call for what you're trying to say.
The model never gets tired. It generally never forgets the rules. It can instantaneously diagram out a sentence and work out an accurate model of the semantics of your writing. It's doing things you cannot do with a grammar checker.
You're not going to want to act on everything an LLM flags. LLMs have an idiosyncratic sense of style (I keep saying they write every sentence like it's the headline of a magazine article). Some of these quirks of writing will be part of your voice, and you'll need to keep them.
You're not going to want the LLM to give you alternate words and sentences. That way lies Velveeta. And many times, the LLM will flag a usage issue and your response won't be to act on the suggestion, but just to rewrite the sentence or paragraph entirely --- or, better yet: just delete it, which is an awesome feeling.
This is using an LLM as a copyeditor. I've worked with professional copyeditors, and the feeling of working with an LLM copyeditor is comparable, except that the LLM is much more thorough, and I don't feel bad about ignoring it when we disagree. This style of working doesn't allow the model to infect your writing; it's just making you better informed about the words you're choosing. I think more people should try it.
I suspect that between the advice in this article and your own...people fear that thinking is the only thing they have left to do in order to write.
I find it so weird that people use AI to write things. Writing is one of the easiest things you can do. Just put one word after the other.
Writing is pretty hard for a lot of people, maybe especially so if they are more non-verbal thinkers, and then doubly so again if one must write not in one's native language.
Depends on the writing. Writing a blog post or something is challenging because you want to have an engaging style. But I see people at work using LLMs to fill out tickets, which is the easiest thing in the world. It's just a plain description of things, it requires no skill at writing at all. It baffles me that people are using LLMs for something like that, which should take less than a minute of your time to fill out and will be more pleasant for the recipient than the slop the model produces.
Here I would maybe argue that the simplicity makes these kinds of tasks tedious (like doing taxes), so it also makes a lot of sense for people to want to just throw AI at it. Tedious, easy rote tasks can be much more unpleasant than engaging ones.
I think people massively over-rely on AI and I really worry about the consequences of this. But there is nothing baffling at all about the basic appeal, IMO.
It depends on what you are trying to convey and how.
I don't think I have to speculate why comments like yours get downvoted so much.
Anyway, I completely agree. Writing isn't something to be delegated.
The problem is not new with LLMs. Businesses have been delegating writing to idiots who don't really care for the entire history of business. Authors and musicians have traded their souls for inauthentic crowd-pleasing results for about as long.
No, the problem with LLMs is that people are all using the same generic models. This only further shows that the future of LLMs is locally trained and locally run. Personal LLMs on a personal computer. Business LLMs on-prem. We will continue see them flourish in ways that generate absolutely no money whatsoever on their own, but do add marginal value when combined with a heavy dose of human creativity.
All the clueless old farts that are still alive by this point will continue to lecture the rest of us with more condescending gee whiz "whaddyaknow" and "whodathunkit" nonsense. As if they've even had their finger on the pulse since the early 2000s.
> I don't think I have to speculate why comments like yours get downvoted so much.
I'm genuinely not sure what you're implying here. Can you expand?
No, I won't. That's why I said that.
I'm even more confused now. I guess this was an oblique way of accusing people of astroturfing?
Hmm, when do you suppose HN started (ab)using the downvotes reddit-style anyway? I don't recall seeing hardly any downvotes before 2023.
Who cares? AI gives me lots of text and it's not like it's ever going to go away. Any time in my life in the future when I need text, I can just have AI generate it. Thinking is solved.