Stock Claude writing is extremely obvious to me, it always uses the same sentence structure, it loves to verb nouns, every sentence has MAX IMPACT, etc. I find it really offputting nowadays, just because it's so overused.
Surprised to find no mention of "load-bearing" and "gated" particularly for Claude.
In particular Opus 5 is scaling new heights of incomprehensibility with excessive metaphor-making, especially when it uses an existing word in the prompt or the code comments as a novel metaphor in an unrelated context.
At some point, and I think the article hints at it, the real difference will be substance. No matter how it changes its writing style, a lot of problems with it are more about the fact that it usually says nothing. Even if it wrote like your favorite author, if there is no substance, there is no reward as a reader (maybe even as a writer).
Overall, i think its kind of useless to think of ai writing as bad because of ai writing quirks but rather because they imply low effort and a lack of interest even by the writer. That cannot be fixed no matter how they change their syntax and style.
For example, I recently spent 2 weeks researching Bruce Lee to write an article on his image and how society compressed it into this warrior monk guy. I could put the idea as a prompt and put in my exact writing style but its output will just be a reiteration of already common knowledge instead of the actual efforts I made researching all of it + having my own idea of the topic. I use AI a lot even as a non-technical person mostly but I find it lacking when it comes to being able to form subjective opinions and unique creative ideas. Ironically, after having done so much research, I don't really have any problem typing away the article and I would rather do it by hand than use an LLM. But I also love doing all of it and if you just want to create "content", low effort is not a hinderence as long as it hits some external unrelated marketing metric.
"The point which the amount of AI generated data on the internet surpasses human generated content, to the extent that it is impossible to distinguish between human and AI generated content.”
The term is derived from the Turing test + Kessler syndrome
This is distinct from the dead Internet theory (2) where humans can actually distinguish if a response is a bot.
I like how this article pokes fun by not only using the tropes it describes sarcastically, but then adding subtext that AI never would. It also uses alliteration, grouping synonyms that use similar sounds, another playful literary device that LLMs don't get.
I have a little script (also visible in my comment history) which (respectfully) scrapes the algolia side of HN, and dumps the entire comment history of mortal users. As noted, it will break on legends, ie Thomas's, certain Books, etc, but successfully dumped my entire history from 2026 - 2016.
And yes, I can see a slight change in my own style, where I am using more "not" statements (negative propositions). While there are some folks with higher integrity and greater immunity to this infection, I am seeing that I am not entirely immune.
However, I have probably generated more dialog with AI over two years than in most of my years altogether. I have well over 7GB of content, of course, much of it the interlocutor -- but similar to how some adopt expressions and accents when imbued in various environments, alas, I have with AI, to whatever extent has bypassed my efforts to resist. The bright side, if it can be viewed such, is that I have developed an involuntary sensitivity to AI writing, often to the point of paranoia and false positives.
It's not about em dashes. It's not about particular turns of phrase. It's about getting the whole picture.
Seriously the biggest tell for AI writing is "eighth grade essay syndrome". You know how an eighth grader writes an essay and they don't have enough original thought to fill up the word count so they pad it out with impressive sounding bloviation and dramatic language, often repeating key points or rephrasing them in a more long-winded way? AI does that on its default settings, probably because it's been trained on a lot of posts from StalkedIn douchebags who never matured beyond that level of writing. It's a reliable sniff test even when the AI has been prompted to avoid the usual obvious tells. Except if you're reading StalkedIn posts. Then it's just how people write.
> With every update, our study shows, AI writing is becoming more similar to human prose.
Any possible vice-versa: There's a meta problem with AI writing detection. I have found my (non-AI-assisted) prose possibly enhanced or modified by my learnings and observations at how LLMs write. In some sense they have made me better at structuring my thoughts and how to communicate them.
What happened to unreadable sociological journals since AI, the kinds where people used to make (manual) generators that could imitate their vacuous fluff. For ai, it should be effortless to flood their submissions with literal nonsense.
Can't find a link to the pre-AI generators... they are ALL ai paper generators now.
I'm fine with reading AI edited content as long as the author is credible and trustworthy. Unless a reliable author keeps getting exposed... Because there's so much GENAI content on websites these days that I don't really have a choice
One way to tell is if the tone is hostile, cynical, or otherwise not full of that giddy optimistic tone associated with what I’ll call “California Business English.”
The mainstream AI companies are loathe to output anything too controversial, so if someone is writing like Burroughs (in content, not merely style) they are less likely to be AI. That is possible to bypass with local models, but the vast majority of writers aren’t going to do that (yet.)
IMO we’re about to see an explosion of slang and offensive language in writing culture.
30 comments
[ 0.24 ms ] story [ 14.3 ms ] threadIn particular Opus 5 is scaling new heights of incomprehensibility with excessive metaphor-making, especially when it uses an existing word in the prompt or the code comments as a novel metaphor in an unrelated context.
Overall, i think its kind of useless to think of ai writing as bad because of ai writing quirks but rather because they imply low effort and a lack of interest even by the writer. That cannot be fixed no matter how they change their syntax and style.
For example, I recently spent 2 weeks researching Bruce Lee to write an article on his image and how society compressed it into this warrior monk guy. I could put the idea as a prompt and put in my exact writing style but its output will just be a reiteration of already common knowledge instead of the actual efforts I made researching all of it + having my own idea of the topic. I use AI a lot even as a non-technical person mostly but I find it lacking when it comes to being able to form subjective opinions and unique creative ideas. Ironically, after having done so much research, I don't really have any problem typing away the article and I would rather do it by hand than use an LLM. But I also love doing all of it and if you just want to create "content", low effort is not a hinderence as long as it hits some external unrelated marketing metric.
"The point which the amount of AI generated data on the internet surpasses human generated content, to the extent that it is impossible to distinguish between human and AI generated content.”
The term is derived from the Turing test + Kessler syndrome
This is distinct from the dead Internet theory (2) where humans can actually distinguish if a response is a bot.
(1) https://kemendo.com/blog/turing-syndrome.html
(2) https://en.wikipedia.org/wiki/Dead_Internet_theory
I've put in some effort to try to help agentic tools avoid these problems: https://www.joshka.net/practice/rules/documentation/docs-avo....
It is one of the biggest turn offs and most of the times I end up giving it again to AI to summarise that into a format which is more readable.
And yes, I can see a slight change in my own style, where I am using more "not" statements (negative propositions). While there are some folks with higher integrity and greater immunity to this infection, I am seeing that I am not entirely immune.
However, I have probably generated more dialog with AI over two years than in most of my years altogether. I have well over 7GB of content, of course, much of it the interlocutor -- but similar to how some adopt expressions and accents when imbued in various environments, alas, I have with AI, to whatever extent has bypassed my efforts to resist. The bright side, if it can be viewed such, is that I have developed an involuntary sensitivity to AI writing, often to the point of paranoia and false positives.
Seriously the biggest tell for AI writing is "eighth grade essay syndrome". You know how an eighth grader writes an essay and they don't have enough original thought to fill up the word count so they pad it out with impressive sounding bloviation and dramatic language, often repeating key points or rephrasing them in a more long-winded way? AI does that on its default settings, probably because it's been trained on a lot of posts from StalkedIn douchebags who never matured beyond that level of writing. It's a reliable sniff test even when the AI has been prompted to avoid the usual obvious tells. Except if you're reading StalkedIn posts. Then it's just how people write.
Any possible vice-versa: There's a meta problem with AI writing detection. I have found my (non-AI-assisted) prose possibly enhanced or modified by my learnings and observations at how LLMs write. In some sense they have made me better at structuring my thoughts and how to communicate them.
Can't find a link to the pre-AI generators... they are ALL ai paper generators now.
The mainstream AI companies are loathe to output anything too controversial, so if someone is writing like Burroughs (in content, not merely style) they are less likely to be AI. That is possible to bypass with local models, but the vast majority of writers aren’t going to do that (yet.)
IMO we’re about to see an explosion of slang and offensive language in writing culture.
"...this issue is unusually well defined..."
"...this document is unusually well structured..."
"...the reviewers came back and were unusually well aligned..."
AI is regression to the mean. That is, being the next great writer in the starboard tail of the distribution will not be the output.
If that's OK, then fine.
What's more telling is that this external machine is not the internal growth needed to become that great writer.
"There is no royal road to geometry", and Euclid yourself if you think that AI is that road.