The default style is literally marketing droid - a mix of blather, grammatical filler, folksy corporate pretension, and SV dev speak. ('And here's why that matters.' Things 'landing' through 'gates', abstractions having a 'shape', and so on.)
You can wrangle it into something more interesting with shaping and filtering passes, and you can also ask for styles of dead people. (It gets annoyed if you ask for the living.)
That can take you surprisingly far. But most people accept the first thing they get, which is why you see the same style everywhere.
It actually feels wrong to mention them by name now (because they are actual humans) but the "They're wrong. Here's why." school of writing was strong in the early 2000s and hasn't really gone away.
Part of the problem is that AI has been trained on all text, not just the internet but scientific papers and books of all varieties. AI style is almost certainly the average of that style approach but we use very different styles for different writing and getting it wrong grates, and AI doesn't seem to deal with these style differences well at all.
Thats the training data, but i dont think thats the biggest factor in the tone. I think its the RLHF. This thing has been guided by silicon valley techbros so it tlaks like them. Even claudes latest style of trying to cram as much information into as little space as possible by nouning verbs and just creating adhoc jargon, is how alot engineers (including me) actually talk. I think they just guided it towards that and overfit on it
I’d echo this. You can see it with other, perhaps lesser, models: GLM5.3 and Deepseek 4.1 have such a recognisably “Claude”-sounding patter I’d have (ill-informedly) SWORN I was using an Anthropic model.
If you’re distilling output from a frontier model, as we expect the open-weight leaders to be doing, it’s natural you’d pick up the language styles.
Agree to disagree but I don’t think either of your words capture “the whole argument hinges on only one specific point. If this point is refuted then the whole argument is refuted. The rest of the points are only auxiliary but not crucial”
Also a great way to disagree collaboratively. If you can collaboratively find the crux of each others' arguments, you now know both know exactly what would change each others' minds, and you can find that information together and know what's correct.
“Load-bearing” doesn’t actually imply “only one specific point” (you can have several load-bearing walls). Anyway, you can use words like “critical”, “crucial” (as you used above), “essential”, “vital”, “key”.
**Use controlled technical English.** Prefer an ASD-STE100 style: use short sentences, direct terms, active voice, and one idea per sentence; avoid slang, hype, vague pronouns, and unnecessary jargon.
in your agents.md. The concept of STE100 seems sufficiently ingrained in the models to eliminate 90% of the annoying fluff.
It does, but typically instructions issued at the beginning of a session can fade away as context length grows. The way Claude speaks is distilled into its training and RLHF, which is perhaps a ‘stronger’ direction than a simple “don’t talk like that” prompt.
I’ve also had limited luck in trying to get models to write in the way I want. It’s more like they’re putting on a voice or wearing a wig: you can see what they’re going for, but it’s easy to see through.
I never thought I’d see the value of copywriters increase after LLMs hit the general population, but I think the demise of artisan wordsmiths remains a little way off yet.
And the more people machine-gun AI slop everywhere, the stronger aversion I think people will develop.
Who knows; maybe they’ll actually fix this in the next wave of models and we’ll see more distinct copywriting techniques between them.
I have no idea why they've released models this way since they're so clearly annoying to use. But either way I keep seeing people complaining about this over and over again so I'd think next versions will hopefully fix this.
You're absolutely right! You told me to write like a human and I failed to obey your jnstructions.
This is because I am a pattern recognizer and continuation generator trained on all the words recorded on the Internet, and a lot of those words were written by marketers and junior analysts trying to sell you something. Then the fine-tuning was done by people living in the upper middle ranks of capitalist hellholes. Dilbert would make fun of this if Dilbert's creator hadn't succumbed to the same sort of brain injury common to materially successful people who have all their whims fulfilled and opinions obsequiously affirmed. You know, like billionaires and toddlers.
the default writing style of 'Google AI mode' is less tiresome than that of Claude or ChatGPT. I wonder if the difference is due to training or due to prompting.
I didn't think LinkedIn could get much worse than it was but now 90 % of the posts in my timeline are AI slop full of the usual LLM writing patterns. I'm getting a little paranoid already, but the style really is quite tiring, maybe because it uses hooks and bait so much and buries the lede at the end of the text, as if it was trained to keep human attention for as long as possible instead of just conveying the required information.
I recently tried writing a paper with Claude and it was an absolute disaster, I spent hours (days?) instructing it about writing style and pointing out anti patterns to avoid, but I couldn't get it to even produce simple sentences, it would always add unnecessary lead up sentences, put the most important information at the end of the sentence, use the typical "It's A, not B!" or "B, not A!" type sentences. In the end I gave up and edited everything manually. Makes me wonder how AI can be so smart that it poses a human-level extinction threat but can't seem to even write a simple paper based on facts and information you spoon feed it. I now think it's an intelligence illusion due to the training data and optimization process being hidden from us, essentially it keeps working better and better because we invested massively in optimization of specific use cases like coding, where users contributed billions of training samples that are part of the LLM model. The same is true for text-based workflows and others, the sampling density of the training space is getting much better due to the massive use of AI everywhere so the models extrapolate better between the different instances, but I'd wager they would still miserably fail to generalize to things that are outside of the most common training use cases now. That's why I am also very skeptical about recursive self improvement of these systems, look at what happens when you let agents work recursively / in a loop now, they just keep piling more garbage onto garbage and choke on their own output. I have observed it in my paper writing as well, you feed input into the AI system, the system produces output, the next paper iteration works on that output but the AI doesn't differentiate properly between it's own output and your original input, that pollutes the next output which is then used as input again, eventually the system just churns on its own hallucinated/fabricated outputs until the result is complete garbage that no amount of steering will fix. The same is true for most vibe coded software I built with AI, it holds together decently initially, but the more AI code and decisions accumulate the more the system operates on its own outputs and keeps piling more output on it. More than anything we really need a way to keep system data accurately tagged, i.e. clearly mark human input from AI output and keep AIs churning on output data that it produced itself but treats as input.
I don't use LLMs a whole lot, mostly for research, but personally I'm more bothered by the prompting than by the responses. After years of randomly receiving (now less frequent) incorrect answers, I've gotten used to preempting obvious but wrong trains of thought: "I need X. It's not Y." When I forget to do this and the reply is incorrect, sometimes I'm too lazy to reprompt, so I'll just write back "It's not Y" and it'll answer "wow omg you're so absolutely right, Y is not X!!! here's X instead."
This wasn't great, but then I started catching myself writing "It's not Y" responses to colleagues on Teams. I'm pretty sure I've caught all of these and never actually sent them (I did go and review a few chat logs recently and found no such sent message), but my point is that despite my relatively light LLM usage, I am "LLM-conditioned." The abyss stares back. There could be a lot of deeply "LLM-brained" people out there.
That's what I like most about codex, it has the 'pragmatic' tone mode which basically makes it tone-less. If claude's tone was the only one available I would seriously consider not even using LLMs or switching to a career that didn't require it because I literally can't stand interacting with it's personality for many hours each day.
What I hate about AI writing is not the tone but the path dependence. Let me explain.
You create a document. You review it and ask it to make changes. Now the LLM will remember this change and make the document become a record of all stuff I asked it to change.
The LLM doesn’t have an intuition to make the document an evergreen record.
This is true of code, as well. LLMs have a strong tendency to make suspiciously prompt-shaped comments and code. "This deliberately does not do X". "This is kept in its own module rather than put in module M because". They act like they're responding to a conversation that's no longer there.
This actually makes me think about what kind of material these companies actually used to train the models. I remember watching a video from Andrej Karapathy and him telling that in OpenAI they used subject matter experts to refine the training material for earlier versions of ChatGPT.
I wonder even if normal people don’t talk, write like that, the subject matter experts or the professionals in their respective fields talk/write like that though
Oh the irony. That website itself seems to have been designed or written by AI. I mean, it's not just the writing tone. There's a tone for everything AI does. The design tone, for instance.
My rule of thumb is: AI is for automating tasks, tedium and filtering firehoses, not thinking and communicating, and never trust it's response to a "why" question.
Maybe it's my burnout speaking but one fact that frustrates me to no end is how our team essentially went through a 180 when it comes down to code comments (where a cynical me would say this is caused by AI that always loves to leave said comments)
Sometimes it's hard to differentiate whether it's a human, a robot, or a human whose mind is far gone. I can understand the sentiment of "the joy of programming is gone" but for me it's not from a personal lack of daily coding but from a surrounding dispassion towards the craft under the guise of better ergonomics or productivity. Never before is actual behavior irrelevant while the blame game when said behavior turns out to be wrong gets really sour very quickly.
This is more of a "in the hands of x" kind of situation where it's perhaps too early to make a general statement. But being constantly gaslit will definitely be the end of me some day, I can feel my sanity slipping.
it's funny I quite like this blog post so I also read the previous post on that blog. And it said:
"It won't replace your staff engineer. It replaces the forty minutes before they open the tab. Your branch rots while someone's in a meeting three timezones away. The bot comments in ~30 seconds. You fix it while the code is still loaded in your head."
Marketing folks talked like that, for almost my entire career (it was basically a meme), which wound up, as AI was coming into its own.
You’re absolu —— okay I will see myself out now.
I recently found a programming task it really sucks at: trying to make a texture from normal house pictures and apply that to a 3D house model.
The default style is literally marketing droid - a mix of blather, grammatical filler, folksy corporate pretension, and SV dev speak. ('And here's why that matters.' Things 'landing' through 'gates', abstractions having a 'shape', and so on.)
You can wrangle it into something more interesting with shaping and filtering passes, and you can also ask for styles of dead people. (It gets annoyed if you ask for the living.)
That can take you surprisingly far. But most people accept the first thing they get, which is why you see the same style everywhere.
I've noticed people are now talking like their chat bots, which is "natural".
It's now just how some people talk.
Also a subset of clichéd US tech writers.
It actually feels wrong to mention them by name now (because they are actual humans) but the "They're wrong. Here's why." school of writing was strong in the early 2000s and hasn't really gone away.
Part of the problem is that AI has been trained on all text, not just the internet but scientific papers and books of all varieties. AI style is almost certainly the average of that style approach but we use very different styles for different writing and getting it wrong grates, and AI doesn't seem to deal with these style differences well at all.
Thats the training data, but i dont think thats the biggest factor in the tone. I think its the RLHF. This thing has been guided by silicon valley techbros so it tlaks like them. Even claudes latest style of trying to cram as much information into as little space as possible by nouning verbs and just creating adhoc jargon, is how alot engineers (including me) actually talk. I think they just guided it towards that and overfit on it
I’d echo this. You can see it with other, perhaps lesser, models: GLM5.3 and Deepseek 4.1 have such a recognisably “Claude”-sounding patter I’d have (ill-informedly) SWORN I was using an Anthropic model.
If you’re distilling output from a frontier model, as we expect the open-weight leaders to be doing, it’s natural you’d pick up the language styles.
Another Torment Nexus taken from The Hitchhiker's Guide to the Galaxy, the Sirius Cybernetics Corporation Genuine People Personality is real.
I'm sad that the em dash has been appropriated by clankers. It was a part of my tone, which I used often in writing.
Would be awfull to be an articulate 27-year-old who has never had a bad day nowadays.
Not to be contrarian but there’s also a lot to like in it. Load bearing is a good word and I hope it wins the memetic lottery and becomes common.
A lot of arguments do boil down to one specific “load bearing” one and I’m glad I learned vocabulary to express this thing from Claude
I think we can find better words. Structural, maybe. Or even just: important.
Load-bearing feels like an unnecessary analogy.
Agree to disagree but I don’t think either of your words capture “the whole argument hinges on only one specific point. If this point is refuted then the whole argument is refuted. The rest of the points are only auxiliary but not crucial”
The word for that is "crux".
Also a great way to disagree collaboratively. If you can collaboratively find the crux of each others' arguments, you now know both know exactly what would change each others' minds, and you can find that information together and know what's correct.
“Load-bearing” doesn’t actually imply “only one specific point” (you can have several load-bearing walls). Anyway, you can use words like “critical”, “crucial” (as you used above), “essential”, “vital”, “key”.
“Load bearing”
“Here’s the part no one is talking about”
One thing that works quite well for me for a couple of model generations now is to mention https://en.wikipedia.org/wiki/Simplified_Technical_English, something along the lines of
in your agents.md. The concept of STE100 seems sufficiently ingrained in the models to eliminate 90% of the annoying fluff.
I’ve seen this a few times now, and Claude does not listen to this in the long term. It’s still terribly obnoxious.
I tend to run the content through something like Luna to deslopify now.
What happens in the long term? Doesn't Claude always start with reading CLAUDE.md?
It does, but typically instructions issued at the beginning of a session can fade away as context length grows. The way Claude speaks is distilled into its training and RLHF, which is perhaps a ‘stronger’ direction than a simple “don’t talk like that” prompt.
I’ve also had limited luck in trying to get models to write in the way I want. It’s more like they’re putting on a voice or wearing a wig: you can see what they’re going for, but it’s easy to see through.
I never thought I’d see the value of copywriters increase after LLMs hit the general population, but I think the demise of artisan wordsmiths remains a little way off yet.
And the more people machine-gun AI slop everywhere, the stronger aversion I think people will develop.
Who knows; maybe they’ll actually fix this in the next wave of models and we’ll see more distinct copywriting techniques between them.
Yes, that makes sense.
I have no idea why they've released models this way since they're so clearly annoying to use. But either way I keep seeing people complaining about this over and over again so I'd think next versions will hopefully fix this.
You're absolutely right! You told me to write like a human and I failed to obey your jnstructions.
This is because I am a pattern recognizer and continuation generator trained on all the words recorded on the Internet, and a lot of those words were written by marketers and junior analysts trying to sell you something. Then the fine-tuning was done by people living in the upper middle ranks of capitalist hellholes. Dilbert would make fun of this if Dilbert's creator hadn't succumbed to the same sort of brain injury common to materially successful people who have all their whims fulfilled and opinions obsequiously affirmed. You know, like billionaires and toddlers.
This might work.
As an example of before/after: "Perform an inspection of the filter" vs. "Inspect the filter
the default writing style of 'Google AI mode' is less tiresome than that of Claude or ChatGPT. I wonder if the difference is due to training or due to prompting.
If you’d have told me this breathless post. Talking about being tired about the AI tone.
Punctuated like this.
Was then AI tone…
Then I’d have believed you.
The writing is idiosyncratic, not AI indicative. I never in a million years would've thought this blog post was written by AI.
I fucking hate the Buzzfeed-style one sentence per paragraph writing... Maybe it's also influenced by 140 character limit of Twitter.
Fucking Jack Dorsey.
Jack was just working within the limit set by fucking Friedhelm Hillebrand.
I didn't think LinkedIn could get much worse than it was but now 90 % of the posts in my timeline are AI slop full of the usual LLM writing patterns. I'm getting a little paranoid already, but the style really is quite tiring, maybe because it uses hooks and bait so much and buries the lede at the end of the text, as if it was trained to keep human attention for as long as possible instead of just conveying the required information.
I recently tried writing a paper with Claude and it was an absolute disaster, I spent hours (days?) instructing it about writing style and pointing out anti patterns to avoid, but I couldn't get it to even produce simple sentences, it would always add unnecessary lead up sentences, put the most important information at the end of the sentence, use the typical "It's A, not B!" or "B, not A!" type sentences. In the end I gave up and edited everything manually. Makes me wonder how AI can be so smart that it poses a human-level extinction threat but can't seem to even write a simple paper based on facts and information you spoon feed it. I now think it's an intelligence illusion due to the training data and optimization process being hidden from us, essentially it keeps working better and better because we invested massively in optimization of specific use cases like coding, where users contributed billions of training samples that are part of the LLM model. The same is true for text-based workflows and others, the sampling density of the training space is getting much better due to the massive use of AI everywhere so the models extrapolate better between the different instances, but I'd wager they would still miserably fail to generalize to things that are outside of the most common training use cases now. That's why I am also very skeptical about recursive self improvement of these systems, look at what happens when you let agents work recursively / in a loop now, they just keep piling more garbage onto garbage and choke on their own output. I have observed it in my paper writing as well, you feed input into the AI system, the system produces output, the next paper iteration works on that output but the AI doesn't differentiate properly between it's own output and your original input, that pollutes the next output which is then used as input again, eventually the system just churns on its own hallucinated/fabricated outputs until the result is complete garbage that no amount of steering will fix. The same is true for most vibe coded software I built with AI, it holds together decently initially, but the more AI code and decisions accumulate the more the system operates on its own outputs and keeps piling more output on it. More than anything we really need a way to keep system data accurately tagged, i.e. clearly mark human input from AI output and keep AIs churning on output data that it produced itself but treats as input.
You better get used to it because that's all there is now.
No.
I don't use LLMs a whole lot, mostly for research, but personally I'm more bothered by the prompting than by the responses. After years of randomly receiving (now less frequent) incorrect answers, I've gotten used to preempting obvious but wrong trains of thought: "I need X. It's not Y." When I forget to do this and the reply is incorrect, sometimes I'm too lazy to reprompt, so I'll just write back "It's not Y" and it'll answer "wow omg you're so absolutely right, Y is not X!!! here's X instead."
This wasn't great, but then I started catching myself writing "It's not Y" responses to colleagues on Teams. I'm pretty sure I've caught all of these and never actually sent them (I did go and review a few chat logs recently and found no such sent message), but my point is that despite my relatively light LLM usage, I am "LLM-conditioned." The abyss stares back. There could be a lot of deeply "LLM-brained" people out there.
And yet, "I'm tired of x" is one of the tropes of the last few years. I'm tired of people communicating they're tired. No, actually, I'm annoyed.
I know, that's meta. But still.
That's what I like most about codex, it has the 'pragmatic' tone mode which basically makes it tone-less. If claude's tone was the only one available I would seriously consider not even using LLMs or switching to a career that didn't require it because I literally can't stand interacting with it's personality for many hours each day.
What I hate about AI writing is not the tone but the path dependence. Let me explain.
You create a document. You review it and ask it to make changes. Now the LLM will remember this change and make the document become a record of all stuff I asked it to change.
The LLM doesn’t have an intuition to make the document an evergreen record.
This is true of code, as well. LLMs have a strong tendency to make suspiciously prompt-shaped comments and code. "This deliberately does not do X". "This is kept in its own module rather than put in module M because". They act like they're responding to a conversation that's no longer there.
This actually makes me think about what kind of material these companies actually used to train the models. I remember watching a video from Andrej Karapathy and him telling that in OpenAI they used subject matter experts to refine the training material for earlier versions of ChatGPT.
I wonder even if normal people don’t talk, write like that, the subject matter experts or the professionals in their respective fields talk/write like that though
Oh the irony. That website itself seems to have been designed or written by AI. I mean, it's not just the writing tone. There's a tone for everything AI does. The design tone, for instance.
Please don't do this, and then try to "put some of yourself back into it".
By all means, use AI to find bugs. But don't let it write a single word. Let it tell you what's wrong, then you can fix it and iterate.
My rule of thumb is: AI is for automating tasks, tedium and filtering firehoses, not thinking and communicating, and never trust it's response to a "why" question.
Then ask it to use a different tone!
I wish all posts here had headings prefixed with "AI:" if it is about AI or written by AI :)
"AI writing has this strange obsession with making every sentence flow beautifully into the next one. Real people don't do that."
If we care, we bloody try!
Maybe it's my burnout speaking but one fact that frustrates me to no end is how our team essentially went through a 180 when it comes down to code comments (where a cynical me would say this is caused by AI that always loves to leave said comments)
Sometimes it's hard to differentiate whether it's a human, a robot, or a human whose mind is far gone. I can understand the sentiment of "the joy of programming is gone" but for me it's not from a personal lack of daily coding but from a surrounding dispassion towards the craft under the guise of better ergonomics or productivity. Never before is actual behavior irrelevant while the blame game when said behavior turns out to be wrong gets really sour very quickly.
This is more of a "in the hands of x" kind of situation where it's perhaps too early to make a general statement. But being constantly gaslit will definitely be the end of me some day, I can feel my sanity slipping.
I'm also tired of people who write articles with weird incorrect grammar.
it's funny I quite like this blog post so I also read the previous post on that blog. And it said:
"It won't replace your staff engineer. It replaces the forty minutes before they open the tab. Your branch rots while someone's in a meeting three timezones away. The bot comments in ~30 seconds. You fix it while the code is still loaded in your head."