Why are comments turned off on the blog posts itself? I was planning to make 200 thoughtful comments with the cheapest AI I could find to prove a point, but the lack of an option to comment shows you already know the point: something has to gate participation in the conversation. Human authorship is a fantastic gate, which keeps me from pulling my stunt here, where that gate is firmly in place. You could achieve almost the same goals by requiring blog comments to hash to a number ending in X 0s, but requiring human writing is a strict improvement on this. A true devotee of “human writing is not fundamentally better than ai” is signing themselves up to never read a block of text written with actual effort again without some proof of effort mechanism.
Yeah, in a small way I agree with the post, but agree with you as well. For me: human authorship is a pretty solid gate, because the vast majority of people are not doing the work described in the post’s hypothetical “argue through 6 drafts and iterations, wholesale rearrangement, critically considering the content, adjusting further”.
I don’t exactly care about whether or not the actual bits on my screen came directly from a human’s fingers on a keyboard or from an LLM, but I do care about whether or not the material is good, well-written, appropriately concise (which varies both on the content and the audience), fact-checked, and I’m sure a bunch more factors. Some humans fail at these criteria as well.
Human authorship is an imperfect proxy for quality, but saying it’s imperfect doesn’t mean that it’s worthless… quite the opposite, I’d argue. Even if there are flaws in the writing, it at least shows that the author tried, which I do personally put value on. Copy & pasting Claude output without even fact-checking… that’s inherently low effort and if that’s what I wanted to read, I’d just ask Claude myself.
I use LLMs in my workflow to varying degrees depending on the task, but their output never makes it into the wild without serious scrutiny. Sometimes their code is better than what I’d have written, which is great. Sometimes it completely misses the point and needs significant re-steering several times, to the point where it’s easier if I write it myself. For writing, I’ll happily have an LLM harness put together a document explaining a dataflow through a complex system; I ask it to include source filenames and line numbers and can do my own validation before accepting it as truth.
Authors aren't obligated to let strangers publish content on their site, any more than you're obligated to let me exercise my free speech by spraypainting messages on your house. You're not being gatekept from the conversation. You can participate in the conversation by publishing on your own site, just like the author has.
The fact this is AI-generated aside - I guess that's to be expected - it's a clever position to take if you completely ignore the realities of LLM-produced writing and imagine some fantasy world where AI blogs bring us profound wisdom and are unfairly targeted by pitchfork-wielding mobs.
Although there are some niche uses where LLMs assist or enhance the creative process, the vast majority of machine writing exists specifically because it takes zero effort and allows you to spam human cognition at an unprecedented scale. There is no redeeming quality to 99%+ of AI-generated LinkedIn posts, AI-generated books on Amazon, and so on.
The style-based heuristics we previously had at our disposal to filter zero-effort content no longer work here, so detecting LLM text is the next fallback.
>I don’t care whether you wrote every word with your bare hands. I care whether the thought is yours, whether you exercised judgment, whether the result deserves my attention and whether you’ll stand behind it once it leaves your screen.
Alright, compare the following degrees of autonomy:
- I wrote it, made sure everything's correct, let the LLM proofread it/translate it to English, and checked it afterwards
- I had brief scattered notes with no coherent idea, then threw it at an LLM to make it an article in whatever way it feels best because I don't care
- I had a three word prompt "Write about X", then a deep research agent spent ungodly amount of tokens and search and scraping requests, and put up an article (at least some grounding and it can depend on the harness quality)
- I had a three word prompt "Write about X", then a model did it in a single call with no real-life grounding at all
You're looking at my clearly generated article. Can you tell what is mine, what is collected from the web or another source by the agent, and what is sourced from the model's own knowledge? Sourcing is exactly what the generated text obfuscates.
The article is quick to shoehorn LLMs in as the last of a long list of established tools.
LLMs are fundamentally different from other writing tools because they attempt to construct meaning in a way that is not done by the author. A main purpose of writing is to convey meaning, so LLM writing defeats that purpose.
Just send me the prompt if you’re going to write with AI. No point in trying to convince me it’s great, the marginal cost for me to re-run the prompt is zero, just send the prompt - it’s like unzipping a file before you send it to me, it doesn’t help, it just decreases information density.
I don't know, just generated a 15 page paper by discussing it intensely with the AI for about 2 weeks. By now, pretty much every sentence in that paper must be AI generated, but at the same time, every sentence now says exactly what I want it to say. You'd definitely not want the prompt.
"Just sent me the prompt" is also such a hive-mind thing to say these days. It seems you definitely don't need the AI to think for you, HN does it already for you.
> A person can type 3,000 words by hand and have no thought pass through their head beyond the need to finish before lunch.
This isn’t the common case. How many of the thoughts produced while working on the words have much novelty or high quality may be an open question, but it’s rare to type or handwrite 3000 words and not have any thoughts.
> Another can dictate an idea, argue with a language model for an hour, reject six drafts, rearrange the whole piece, rewrite half of it, and publish something they understand down to the final comma
Also probably not the common case. And so far my experience is that models make decent to good editors/critics, but average originators (though with remarkable breadth and sometimes outlier originations).
> A person can type 3,000 words by hand and have no thought pass through their head beyond the need to finish before lunch.
I get this is tongue in cheek, but it's so wrong that it needs pointing out. Even if that is a thought in their head, they clearly have to have others in order to produce the text.
But in general, checking that a human wrote it is economic as much as anything else: attention is currency limited by time and energy, AI effort is abundant and therefore low value, human effort is less abundant and therefore more valued. Pyrite can look pretty but it's not what people want.
Even if I agreed with the author (which I don’t), from a purely selfish perspective why would you want the idea you’ve ostensibly worked hard on communicated in a way that says “low effort”?
I’ll gladly use AI to write if it were good at writing as much as how it is good at coding. But it’s just a terrible writer, and currently it’s easier to type things out myself compared to yelling at Claude.
Presumably the author believes checking food for contamination, and vehicle safety standards are also daft. Insofar as LLMs are a tool, they are not a tool that reliably produces desired output. There was just a thread on this. [1] Even setting aside any moral questions of using them, LLMs create plausible sounding BS. We're not just talking about fire, we're talking about methanol fire.
It most probably wasn’t though, because that would mean doing only the tedious part of writing.
Another thing though. How many times does it happen to you to start writing after having an idea, but scrapping it because you realize it’s wrong/bad/unhelpful. To me, quite often. But if you prompt an LLM, there is no time to have that realization.
I once had a scriptures class where I had to hit word count on an essay and replaced every instance of “Jesus” with “Jesus the only begotten Son of God who died on the cross so that we may be saved from our sins”.
So, yes, humans can also expand text in a mechanical way. I just don’t want to read expanded text. So really the question is “is this text expanded?”. Pre-LLM very little was. Post-LLM much is and has certain distinct flavours that are detectable.
But I’m happy to say that I also don’t like human expanded text and the content in OP certainly reads expanded. It’s like inlining functions everywhere and then writing unrelated comments around them. Not interesting.
After all, if I just want to read text there are reams of content marketing blog posts and SEO keyword concatenations that I could go read. Or to put it in a different way: SaaS dropshipping shemale singles near you this one mother found a puppy buy crypto toys cheap car loans zero down furniture sale foreclosure mesothelioma lawyers car Trump USAID ethereum get rich quick
Hmm. Expanded text. You have just explained why I hate the writing style of the New Yorker.
And when you put it that way, it makes me think of steganography. That's a form of "expanding" the text, and it seems like it might be a useful way to think about the problem. Somewhere hidden in that wall of text is an actual message... but I don't want to have to go find it.
I think this ignores that we're being spammed with written content.
The craft of writing is part of the interesting part. The words you choose are important,same as brush strokes in a painting. Offloading the act of writing to llms takes away that whole part.
I think it depends on the context too. Dale Carnegie (in his famous Win Friends.. book) mentioned how it's annoying to receive letters that say Dictated But Not Read. So people don't want to read (for example) a twitter reply or email that's written on their behalf rather than their own thoughts
When it comes to something like a LinkedIn post however who cares. Everyone uses AI cause the content itself is mercenary anyway right. Thought leadership.
But I have been using Sonnet to write some Youtube narrations and I think I've discovered a kind of value that is not 'AI slop'. I already posted one of them -- about the giant pink hologram in Blade Runner 2049 -- and it's been getting thousands of views and a dozens of likes when clipped into shorts. [https://www.youtube.com/watch?v=sK-TsLQkIos]
The text is like: "The neon is pink. Her hair is blue. Her face is light. She smiles because she was made to smile, and somewhere beneath the awareness of that fact is the older, simpler, more animal truth: a face is smiling at you. Someone said your name. Someone said “you look lonely” and meant it as an invitation rather than a wound."
To me that is a kind of LLM writing that's "value above replacement". Like if I lost that Sonnet 4.6 script I wouldn't be able to get it again it was just a particular permutation of the neural network manifold that emerged from our discussion about the movie and the style I wanted etc.
> The text is like: "The neon is pink. Her hair is blue. Her face is light. She smiles because she was made to smile, and somewhere beneath the awareness of that fact is the older, simpler, more animal truth: a face is smiling at you. Someone said your name. Someone said “you look lonely” and meant it as an invitation rather than a wound."
That sounds cool. It's also a style of SciFi writing--clipped sentences that juxtapose images and let you put together meaning. I wonder how long this sequence can go before the LLM messes up because it has limited understanding (if you will) of the underlying meaning of the images.
Nah, AI writing is bad because it breaks the implicit trust between the reader and the author. It's implicitly assumed when you read something that a person write it. It's not merely a tool. the writing is worse, and people feel like they are being ripped off when they see it and it's undisclosed.
How would you feel about an article if you knew the author circulated it to respected friends for review? That's another way to use LLMs. I do it all the time. LLMs are great sounding boards.
Similarly how would you feel about writing that the author passed over to an editor?
This is my thought, exactly. Reading has always been a contract between the person who wrote the text and the person who’s reading it: “these are my thoughts, recorded to share with you.” When reading LLM-generated text that purports to be a real person’s that contract is broken. Verification that even just the basic ideas are the author’s is impossible. It feels cheap and deceitful.
I simply don’t like AI writing because its prose is terrible. It gives me the ick. No other justification is needed for my dislike. Same with AI generated code for the most part. Takes a lot of work to make AI generated anything palatable, which is fine - we don’t really want to deprecate effort or skill do we?
I dunno man. If you had something worth saying I kinda feel like you would figure out a way to say it with your fingers and a keyboard.
There is SO MUCH STUFF that you could read. You could read all day every day and not keep up with the human-written output in just a single sub-niche of writing within a niche sector. That being the case, you must use some heuristics to determine what should be read. Whether the entity publishing a piece felt it was worth human time to write that piece seems like a fine heuristic.
The privilege argument is so bad! We aren't doing this any more! The tyranny of edge cases is over. It ended some time in 2023. I am not going to feel obligated to read every piece of slop out there because 0.01% of authors have lost the use of their hands. Stephen Hawking somehow managed to write without AI.
Low grade engagement bait that reads like a mediocre attempt at whitewashing slop. If the author's premise holds then we're saying that there's no daylight between the works of an individual who has taken a literal lifetime to hone a skillset and any random dilletante armed with the appropriate AI tools.
Why are comments turned off on the blog posts itself? I was planning to make 200 thoughtful comments with the cheapest AI I could find to prove a point, but the lack of an option to comment shows you already know the point: something has to gate participation in the conversation. Human authorship is a fantastic gate, which keeps me from pulling my stunt here, where that gate is firmly in place. You could achieve almost the same goals by requiring blog comments to hash to a number ending in X 0s, but requiring human writing is a strict improvement on this. A true devotee of “human writing is not fundamentally better than ai” is signing themselves up to never read a block of text written with actual effort again without some proof of effort mechanism.
I guarantee you that people are using AI to write posts here.
I do think that there is a gate against poorly written comments.
Yeah, in a small way I agree with the post, but agree with you as well. For me: human authorship is a pretty solid gate, because the vast majority of people are not doing the work described in the post’s hypothetical “argue through 6 drafts and iterations, wholesale rearrangement, critically considering the content, adjusting further”.
I don’t exactly care about whether or not the actual bits on my screen came directly from a human’s fingers on a keyboard or from an LLM, but I do care about whether or not the material is good, well-written, appropriately concise (which varies both on the content and the audience), fact-checked, and I’m sure a bunch more factors. Some humans fail at these criteria as well.
Human authorship is an imperfect proxy for quality, but saying it’s imperfect doesn’t mean that it’s worthless… quite the opposite, I’d argue. Even if there are flaws in the writing, it at least shows that the author tried, which I do personally put value on. Copy & pasting Claude output without even fact-checking… that’s inherently low effort and if that’s what I wanted to read, I’d just ask Claude myself.
I use LLMs in my workflow to varying degrees depending on the task, but their output never makes it into the wild without serious scrutiny. Sometimes their code is better than what I’d have written, which is great. Sometimes it completely misses the point and needs significant re-steering several times, to the point where it’s easier if I write it myself. For writing, I’ll happily have an LLM harness put together a document explaining a dataflow through a complex system; I ask it to include source filenames and line numbers and can do my own validation before accepting it as truth.
Idea for a captcha
Are there infinitely many primes p such that p − 1 is a perfect square? In other words: Are there infinitely many primes of the form n2 + 1?
Answer: [ ]
( ) I don't know
Authors aren't obligated to let strangers publish content on their site, any more than you're obligated to let me exercise my free speech by spraypainting messages on your house. You're not being gatekept from the conversation. You can participate in the conversation by publishing on your own site, just like the author has.
The fact this is AI-generated aside - I guess that's to be expected - it's a clever position to take if you completely ignore the realities of LLM-produced writing and imagine some fantasy world where AI blogs bring us profound wisdom and are unfairly targeted by pitchfork-wielding mobs.
Although there are some niche uses where LLMs assist or enhance the creative process, the vast majority of machine writing exists specifically because it takes zero effort and allows you to spam human cognition at an unprecedented scale. There is no redeeming quality to 99%+ of AI-generated LinkedIn posts, AI-generated books on Amazon, and so on.
The style-based heuristics we previously had at our disposal to filter zero-effort content no longer work here, so detecting LLM text is the next fallback.
>I don’t care whether you wrote every word with your bare hands. I care whether the thought is yours, whether you exercised judgment, whether the result deserves my attention and whether you’ll stand behind it once it leaves your screen.
Alright, compare the following degrees of autonomy:
- I wrote it, made sure everything's correct, let the LLM proofread it/translate it to English, and checked it afterwards
- I had brief scattered notes with no coherent idea, then threw it at an LLM to make it an article in whatever way it feels best because I don't care
- I had a three word prompt "Write about X", then a deep research agent spent ungodly amount of tokens and search and scraping requests, and put up an article (at least some grounding and it can depend on the harness quality)
- I had a three word prompt "Write about X", then a model did it in a single call with no real-life grounding at all
You're looking at my clearly generated article. Can you tell what is mine, what is collected from the web or another source by the agent, and what is sourced from the model's own knowledge? Sourcing is exactly what the generated text obfuscates.
The difference between the first and the rest is usually blindingly obvious: LLMs are really terrible at crafting their own coherent narrative.
The article is quick to shoehorn LLMs in as the last of a long list of established tools.
LLMs are fundamentally different from other writing tools because they attempt to construct meaning in a way that is not done by the author. A main purpose of writing is to convey meaning, so LLM writing defeats that purpose.
Just send me the prompt if you’re going to write with AI. No point in trying to convince me it’s great, the marginal cost for me to re-run the prompt is zero, just send the prompt - it’s like unzipping a file before you send it to me, it doesn’t help, it just decreases information density.
I don't know, just generated a 15 page paper by discussing it intensely with the AI for about 2 weeks. By now, pretty much every sentence in that paper must be AI generated, but at the same time, every sentence now says exactly what I want it to say. You'd definitely not want the prompt.
"Just sent me the prompt" is also such a hive-mind thing to say these days. It seems you definitely don't need the AI to think for you, HN does it already for you.
> A person can type 3,000 words by hand and have no thought pass through their head beyond the need to finish before lunch.
This isn’t the common case. How many of the thoughts produced while working on the words have much novelty or high quality may be an open question, but it’s rare to type or handwrite 3000 words and not have any thoughts.
> Another can dictate an idea, argue with a language model for an hour, reject six drafts, rearrange the whole piece, rewrite half of it, and publish something they understand down to the final comma
Also probably not the common case. And so far my experience is that models make decent to good editors/critics, but average originators (though with remarkable breadth and sometimes outlier originations).
> A person can type 3,000 words by hand and have no thought pass through their head beyond the need to finish before lunch.
I get this is tongue in cheek, but it's so wrong that it needs pointing out. Even if that is a thought in their head, they clearly have to have others in order to produce the text.
But in general, checking that a human wrote it is economic as much as anything else: attention is currency limited by time and energy, AI effort is abundant and therefore low value, human effort is less abundant and therefore more valued. Pyrite can look pretty but it's not what people want.
> A person can type 3,000 words by hand and have no thought pass through their head beyond the need to finish before lunch.
But when they do that, you can tell.
Even if I agreed with the author (which I don’t), from a purely selfish perspective why would you want the idea you’ve ostensibly worked hard on communicated in a way that says “low effort”?
I’ll gladly use AI to write if it were good at writing as much as how it is good at coding. But it’s just a terrible writer, and currently it’s easier to type things out myself compared to yelling at Claude.
Presumably the author believes checking food for contamination, and vehicle safety standards are also daft. Insofar as LLMs are a tool, they are not a tool that reliably produces desired output. There was just a thread on this. [1] Even setting aside any moral questions of using them, LLMs create plausible sounding BS. We're not just talking about fire, we're talking about methanol fire.
1. https://news.ycombinator.com/item?id=49042354
I think we're just in an transition point.
When I get materials that are clearly AI generated, I look at them as if there was less effort put into it.
But that is irrational. If it was verified for accuracy, and edited for conciseness, and addresses the task at hand, does it matter?
It most probably wasn’t though, because that would mean doing only the tedious part of writing.
Another thing though. How many times does it happen to you to start writing after having an idea, but scrapping it because you realize it’s wrong/bad/unhelpful. To me, quite often. But if you prompt an LLM, there is no time to have that realization.
I once had a scriptures class where I had to hit word count on an essay and replaced every instance of “Jesus” with “Jesus the only begotten Son of God who died on the cross so that we may be saved from our sins”.
So, yes, humans can also expand text in a mechanical way. I just don’t want to read expanded text. So really the question is “is this text expanded?”. Pre-LLM very little was. Post-LLM much is and has certain distinct flavours that are detectable.
But I’m happy to say that I also don’t like human expanded text and the content in OP certainly reads expanded. It’s like inlining functions everywhere and then writing unrelated comments around them. Not interesting.
After all, if I just want to read text there are reams of content marketing blog posts and SEO keyword concatenations that I could go read. Or to put it in a different way: SaaS dropshipping shemale singles near you this one mother found a puppy buy crypto toys cheap car loans zero down furniture sale foreclosure mesothelioma lawyers car Trump USAID ethereum get rich quick
Hmm. Expanded text. You have just explained why I hate the writing style of the New Yorker.
And when you put it that way, it makes me think of steganography. That's a form of "expanding" the text, and it seems like it might be a useful way to think about the problem. Somewhere hidden in that wall of text is an actual message... but I don't want to have to go find it.
I think this ignores that we're being spammed with written content.
The craft of writing is part of the interesting part. The words you choose are important,same as brush strokes in a painting. Offloading the act of writing to llms takes away that whole part.
I think it depends on the context too. Dale Carnegie (in his famous Win Friends.. book) mentioned how it's annoying to receive letters that say Dictated But Not Read. So people don't want to read (for example) a twitter reply or email that's written on their behalf rather than their own thoughts
When it comes to something like a LinkedIn post however who cares. Everyone uses AI cause the content itself is mercenary anyway right. Thought leadership.
But I have been using Sonnet to write some Youtube narrations and I think I've discovered a kind of value that is not 'AI slop'. I already posted one of them -- about the giant pink hologram in Blade Runner 2049 -- and it's been getting thousands of views and a dozens of likes when clipped into shorts. [https://www.youtube.com/watch?v=sK-TsLQkIos]
The text is like: "The neon is pink. Her hair is blue. Her face is light. She smiles because she was made to smile, and somewhere beneath the awareness of that fact is the older, simpler, more animal truth: a face is smiling at you. Someone said your name. Someone said “you look lonely” and meant it as an invitation rather than a wound."
To me that is a kind of LLM writing that's "value above replacement". Like if I lost that Sonnet 4.6 script I wouldn't be able to get it again it was just a particular permutation of the neural network manifold that emerged from our discussion about the movie and the style I wanted etc.
> The text is like: "The neon is pink. Her hair is blue. Her face is light. She smiles because she was made to smile, and somewhere beneath the awareness of that fact is the older, simpler, more animal truth: a face is smiling at you. Someone said your name. Someone said “you look lonely” and meant it as an invitation rather than a wound."
That sounds cool. It's also a style of SciFi writing--clipped sentences that juxtapose images and let you put together meaning. I wonder how long this sequence can go before the LLM messes up because it has limited understanding (if you will) of the underlying meaning of the images.
Nah, AI writing is bad because it breaks the implicit trust between the reader and the author. It's implicitly assumed when you read something that a person write it. It's not merely a tool. the writing is worse, and people feel like they are being ripped off when they see it and it's undisclosed.
How would you feel about an article if you knew the author circulated it to respected friends for review? That's another way to use LLMs. I do it all the time. LLMs are great sounding boards.
Similarly how would you feel about writing that the author passed over to an editor?
This is my thought, exactly. Reading has always been a contract between the person who wrote the text and the person who’s reading it: “these are my thoughts, recorded to share with you.” When reading LLM-generated text that purports to be a real person’s that contract is broken. Verification that even just the basic ideas are the author’s is impossible. It feels cheap and deceitful.
I simply don’t like AI writing because its prose is terrible. It gives me the ick. No other justification is needed for my dislike. Same with AI generated code for the most part. Takes a lot of work to make AI generated anything palatable, which is fine - we don’t really want to deprecate effort or skill do we?
I dunno man. If you had something worth saying I kinda feel like you would figure out a way to say it with your fingers and a keyboard.
There is SO MUCH STUFF that you could read. You could read all day every day and not keep up with the human-written output in just a single sub-niche of writing within a niche sector. That being the case, you must use some heuristics to determine what should be read. Whether the entity publishing a piece felt it was worth human time to write that piece seems like a fine heuristic.
The privilege argument is so bad! We aren't doing this any more! The tyranny of edge cases is over. It ended some time in 2023. I am not going to feel obligated to read every piece of slop out there because 0.01% of authors have lost the use of their hands. Stephen Hawking somehow managed to write without AI.
Low grade engagement bait that reads like a mediocre attempt at whitewashing slop. If the author's premise holds then we're saying that there's no daylight between the works of an individual who has taken a literal lifetime to hone a skillset and any random dilletante armed with the appropriate AI tools.