For me, the biggest AI writing tell (other than the blatantly obvious ones) is an unnatural consistency in style, whatever style that may be. It's most apparent in longer pieces, and I'm not sure I can really pin down exactly what it is. But human writers seem to lack the ability to keep a 100% consistent voice and lapse into different registers at different times. LLMs don't have this natural rhythm, which makes for an exhausting reading experience.
The best example of AI writing I've seen so far was caused by my six year old narrating a prompt ( to me to type in ) to Gemini in story mode. The results were so unexpected and wild I couldn't stop reading it.
Strangely I have yet to get such a compelling result with my own prompts. I think for myself it is tainted with the expectation of what I really wanted and would have written had I taken the time to write the words of the story instead of the prompt.
This is a situation where the work to write the prompt is equivalent to the work to just write the story.
1. The default LLM behavior (at least what I’ve used as a consumer with ChatGPT Claude etc.) is to be excessively verbose, presumably because costs are tied to usage and therefore the assumption is that more text = better.
I’ve spent over a decade working as a copywriter, and IME the most important part of writing is the edit - what to cut out.
So I think it’s probably possible that an AI could write stuff we’d want to read, the default behavior of the AIs most people are using works against it.
2. A lot of the writing that actually gets read today is either a description of a lived experience and/or involves slang. Neither of those things are interesting if done by an AI - I don’t care about the imagined experience of an LLM.
I find ai generated deep code wikis very valuable. They provide clear walk path to read the code. Reading code raw is always painful, trying to trace the right start points, especially with lots of legacy code.
One really valuable thing i'm seeing in open source though is everything is being localized. Most before it was just not available. In a way, that's really good because it helps to bring the chinese and english speaking communities.
The paradox is that we love reading our own AI generated writing and hate reading anyone else's AI generated writing.
On a recent weeklong trip to the Philippines, I generated over a 500 page novel's worth of content from AI around various aspects of Filipino history, culture, social dynamics etc. and actually went over it at least 3 times to fully absorb the material.
But if someone handed me even a 3000 word essay on the Philippines clearly written by AI, I would not be able to get to the end of it.
I wonder if there is a separation of story/structure and drafting. As a movie sometime has separate story and screenplay credits, could a human architect a structure that is then drafted in an acceptable way by Claude. Has anyone found good examples of Claude drafted articles?
Who is “we”? Head over to Reddit and you will see that plenty of people do not notice even the most obvious AI-generated engagement bait and happily spend their time talking to it. Even the people that post about how awful AI is will chat about that very subject to a spam bot without realising.
The average person is not good at spotting AI-generated content. They accept it and want to read it just as long as they don’t realise it’s not real.
Probably not to the quality of the New Yorker fiction section yet but this has already occurred here on HN. People frequently read and comment on things that are LLM based. And Reddit’s usual outrage and repetitive subreddits are almost entirely Dear Abbys written by LLM.
I think there’s lots of stuff where the LLM is translating information that is poorly represented into readable English where it’s fine. Though that does require someone with taste but not ability and usually someone who has taste and cares about their taste would have eventually cultivated their ability.
When I was using free version of AI for my everyday task that were to machinable writing but now I am using paid version AI for my tasks these are very feels that humans have written.
Most positive AI thing shared on HN in a while, and here’s why.
Even after handing the agent a validator and iterating the validation, even after the validator get really good at telling the model what to do and what not to do, there’s still an emptiness.
This gives the author strong ground to rightly assert that both now and after new models render 2026 model writing’s emptiness old news, there’ll remain reasons to write, to communicate, to share understanding.
It’s nice to read something not proclaiming doom and not lamenting career death.
No matter how smart our tools become, the human condition still has plenty ahead.
This is a great thread! I’m currently working on a project that edits AI, slop. And it is slop, however, underneath the puzzling similes and over used metaphors is a story. I’ve discovered that most times the story is worth unearthing if it is you story. And that’s the key. You cannot simply tell Claude or any other AI to write me a great mystery or romance novel. But if you use your imagination, outline, and prompt well you’ll have a solid working draft. It’ll be shit, but at least it something worth working with.
The other thing I’ve discovered is the people using AI to write and do not edit are the people who don’t do one thing…Read. So they don’t know the a paragraph becomes cringe after a bizarre metaphor. What is does isn’t wrong, but that doesn’t make it right. I agree with Mr. Kang, but I believe the future of the publishing industry will include human edited AI stories. We just haven’t gotten there yet.
Yeah, it can. I used AI and just won a national IBPA Benjamin Franklin award for it. I have a new series that's getting good reviews from readers as well as the industry, PW BookLife and Kirkus. I didn't use AI to generate a story, I used it to interrogate the self through sci-fi and memoir.
26 comments
[ 3.2 ms ] story [ 50.5 ms ] threadthey both suffer from the same lack of dimension and intent.
For opinion pieces, I'd rather the human work through an opinion and read about that personal journey!
Strangely I have yet to get such a compelling result with my own prompts. I think for myself it is tainted with the expectation of what I really wanted and would have written had I taken the time to write the words of the story instead of the prompt.
This is a situation where the work to write the prompt is equivalent to the work to just write the story.
1. The default LLM behavior (at least what I’ve used as a consumer with ChatGPT Claude etc.) is to be excessively verbose, presumably because costs are tied to usage and therefore the assumption is that more text = better.
I’ve spent over a decade working as a copywriter, and IME the most important part of writing is the edit - what to cut out.
So I think it’s probably possible that an AI could write stuff we’d want to read, the default behavior of the AIs most people are using works against it.
2. A lot of the writing that actually gets read today is either a description of a lived experience and/or involves slang. Neither of those things are interesting if done by an AI - I don’t care about the imagined experience of an LLM.
I find ai generated deep code wikis very valuable. They provide clear walk path to read the code. Reading code raw is always painful, trying to trace the right start points, especially with lots of legacy code.
https://deepwiki.com/ArroyoSystems/arroyo.
One really valuable thing i'm seeing in open source though is everything is being localized. Most before it was just not available. In a way, that's really good because it helps to bring the chinese and english speaking communities.
On a recent weeklong trip to the Philippines, I generated over a 500 page novel's worth of content from AI around various aspects of Filipino history, culture, social dynamics etc. and actually went over it at least 3 times to fully absorb the material.
But if someone handed me even a 3000 word essay on the Philippines clearly written by AI, I would not be able to get to the end of it.
The average person is not good at spotting AI-generated content. They accept it and want to read it just as long as they don’t realise it’s not real.
I think there’s lots of stuff where the LLM is translating information that is poorly represented into readable English where it’s fine. Though that does require someone with taste but not ability and usually someone who has taste and cares about their taste would have eventually cultivated their ability.
Even after handing the agent a validator and iterating the validation, even after the validator get really good at telling the model what to do and what not to do, there’s still an emptiness.
This gives the author strong ground to rightly assert that both now and after new models render 2026 model writing’s emptiness old news, there’ll remain reasons to write, to communicate, to share understanding.
It’s nice to read something not proclaiming doom and not lamenting career death.
No matter how smart our tools become, the human condition still has plenty ahead.
The other thing I’ve discovered is the people using AI to write and do not edit are the people who don’t do one thing…Read. So they don’t know the a paragraph becomes cringe after a bizarre metaphor. What is does isn’t wrong, but that doesn’t make it right. I agree with Mr. Kang, but I believe the future of the publishing industry will include human edited AI stories. We just haven’t gotten there yet.