I do not mind when I am coding with Claude and it uses all the typical claudisms. I am much more bothered when I am reading a blog post, email, or other form of prose and I see those same claudisms.
I guess they are not annoying since I know I am talking to an LLM and expect the typical responses. When I am reading prose online that I previously would have expected a human to write, it can be quite jarring to realize its an LLM.
I didn't use Claude for a long time, but my coworkers did, so I got infected through a side channel: I ended up reading their vibed docs, noticed "load-bearing", kind of liked it, and started using it in conversation, until I got feedback that I was "talking like Claude", so now I avoid the phrase entirely. The intersection of language and social norms is interesting.
Yes, I had a related experience of reading a book and observing what I thought were claude-isms, only to realize it was written in 2019. Some of the common tells are actually good writing practices, but I guess they are best in smaller doses.
Sure, that's where the AI got them: the training data. These phrases and cliches were very prevalent especially in corporate "white papers" and memos and marketing materials. There was a time when "stove pipe" was a common one too, along with "silo."
But the LLMs really seem to fixate on using the same ones in the same places all the time. I guess that's because that's the highest probability construction.
Actually, AI was learning these 'AI-isms' even back in 2017/2018 (probably even earlier). I think a lot of people who just jumped on the imaginary AI bandwagon more recently don't realise the mannerisms AIs are adopting are not really new. At some point the bleed between 'you' or 'you' and AI will just become so transparent it will be obliterated, more likely than not.
I believe you and OP are in agreement — they were saying that the 2019 book had them, therefore the terms _do_ predate AI. Your point that AI was being trained on material than is load-bearing (lol) but in agreement with OP, not contradictory.
Only mentioning because your "actually" may imply you thought you were disagreeing, when in fact it's one big happy family!
I wrote a thank you message on Teams to my coworkers on a project, and half of them thought I had used AI to write it. As a professional writer in a previous life, I was astonished. Then they told me that they had never seen me write anything more than a sentence or two so naturally they assumed something relatively polished had to be AI assisted.
Though I haven’t been a professional writer, I’ve been a good writer with an expansive vocabulary since high school where English was my best subject despite being a STEM maven. I hate the fact that what was previously considered an advantageous skill is now a millstone in public use. I hate having to dumb down and self—censor in order to avoid being accused of using (or being) an LLM. Even though my writing has a few repeated personal tells - certain linguistic errors that I nevertheless employ as part of my idiom (and an LLM never would) - people don’t always notice them. So, I’m forced to change my voice to deal with what’s essentially an IRL Captcha.
When I write these days, I am more aggressive in using "I", so it's clear it's my own voice. Generally, an LLM is less prone to self-reference like that unless it's prompted to, I guess.
unfortunately load bearing is one of those things that became a claudism but has been part of my daily lexicon for decades. There are a lot of things I say regularly as part of my own vocal quirkiness that now I have to self censor.
I'm pretty sure that's not the source for me but most of my vocal quirks have origin stories like this so it's entirely possible. They're almost always things that I heard which either amused me or I thought sounded cool at the time and just stuck with me.
The one Claudism I will never ever use is "synthesize". I don't even know where that came from - no one talks or writes like that - "I can synthesize that for you".
Stories like this make me want to use each AI at least briefly, so I know what to avoid writing/saying. Or maybe just do a search every couple months to find out what different AIs are known for saying too often.
Claude's affected my language in two ways. one is that, for a long time, Claude in particular responded more to feedback if I swore at it, which caused me to swear at it more. this vicious cycle generalized to the point where I now have to consciously remind myself not to swear when doing something as simple as buying a coffee or asking somebody what time it is. it was difficult to even write that sentence without throwing in an F-bomb just to emphasize the silliness of the problem.
anyway, the other way is I found it's helpful when prompting LLMs to use the same "it's not delivery, it's DiGiorno's" pattern that they're all so obsessed with. especially when the thing's misapprehended some concept, so you need to clarify. this hasn't yet generalized from the fake "conversations" I have with chatbots into my conversational style out in the real world, but the risk is fully there. (it's not an inevitability -- it's an occupational hazard.)
Exactly this. For whatever reason, Claude likes to talk about the "shape" of "load-bearing" "seams," but if that's the internal jargon it needs to plan and execute its work, who am I to judge?
But if I'm reading what is supposed to be someone's original thoughts, it's a huge bummer to see an obvious AI tell. You might say that "it's not just disappointing—it's disrespectful."
Hot take -- I'm glad that LLMs still tend to have recognizable communication patterns, because they're often the only clue I have to filter AI content.
If the next generation of AI content produced no recognizable LLM patterns, and was indistinguishable from an actual human author, would you still care and try to determine whether the content was AI produced?
Of course. The "content" is how humans communicate with each other, it doesn't just exist for its own sake (except in some degenerate cases). If you know that a human has authored it, you can infer their intent and thought-process from various choices they made across it. There's no such thing as intentional choices when the content is generated though.
I’ve asked essentially the same question many times to many people, the short answer is “yes” because it’s a matter of ideology not logic for them.
I get just as mad about shitty human output as I do about shitty LLM output. The bad thing about LLMs is that they have increased the volume of shit most people have to sift through.
When you open a requirements doc and it’s got 13 load bearing em dashes on the first page you known it’s gonna be bad day
I would like to know if text is LLM generated even if I can't tell from the content itself. For me it's a matter of attention (hah) and a quality signal. The poster expects to spend a minimal amount of effort on the post, and all the readers will have to spend the same amount of attention whether its LLM generated or not.
To me, it's disrespectful to expect someone to waste their day reading every word of a blog post when even the author has not read every word. It shows that you value your time over your reader's time.
I want to know when I'm consuming AI content because the source of information matters. I want to know what was at stake for the author, what motives they had and didn't have, what biases I should be aware of, and, for example, whether I'm reading content farmed slop that exists solely to attract ad impressions.
You still won’t know even if it’s labeled though. Someone could put an incredible amount of effort into writing something, not be satisfied with some aspect, and have an LLM refactor it for example. Now you’ve got a lot of em dashes and you issue a shallow dismissal.
There was an HN submission recently where the author spent a lot of time and effort working with an LLM to write a story and get the LLM to follow a specific style and whatnot. Wish I could recall it offhand. Many commenters were very upset when they found out it was LLM generated, even though they couldn’t tell while reading it.
Basically what matters to me is some combo of how much effort went into it, and how accurate it is.
Whenever I use AI generated content in direct communication - ie slack, email, jira tickets, etc, I always prefix any AI content with an obvious label: 'Claude says' or 'AI analysis: ' etc. In some cases I get claude to update jira tickets (really nice use case btw) with testing notes, I make sure the team knows that any notes in that format come from the AI based on the related commits.
I still keep the AI label even if I edit the result for correctness or clarity etc. The last thing I want to do is have someone read AI content and think it came directly from me. I really don't understand the thinking of people that do that - it's like they're hiding or intentionally cheating somehow.
AI generated content can be really, really useful (with some guidance, AI is way better at creating useful git commit messages and jira ticket comments than I am), but pretending that content is yours just seems way too much like straight up lying.
If you rework a paragraph in a tight loop - you change a few words, the chatbot changes a few words, going round four or five times until you've got something you're happy with, I don't see how it's meaningful to assign a percentage.
I guess you could write an editor that does it? Tracks the origin of every word in the document? But what if you cut'n'paste a word? Or worse, see it and retype it manually?
I think the best you can hope for here is "this text was written with AI assistance".
Honestly--and I say this as a flesh-and-blood human--I continue to be pissed off that AI has ruined load-bearing parts of my vocabulary. You're absolutely right that it's starting to trigger me when I read random blog posts and come across these linguistic ticks, but I can't help but be resentful. Humans invented language and now robots are coopting it.
I agree, though as a side note I'm very curious to see how models will begin steering _our_ language. If you have popular models repeating "load-bearing" to every developer, eventually I imagine developers (especially junior developers who may not know that it's a Claudism) will begin to repeat it.
Lots of people have their own voice and tend to prefer certain phrases. This has been the case for a long time and is generally not a big issue.
Now LLMs come along and they also have their own phrasing preferences. But now it's a problem because what used to be personal preferences of a single person that manifests in 5000 words per day from one person tops, is now the bias of a single model multiplied x10,000,000,000 generated tokens per day so any bias sticks out like a sore thumb.
I think it might be even worse. LLMs seem to get tragically stuck on certain patterns. Maybe it's partly because a pile of weights essentially always starts from scratch in the same condition, but even within a single conversation, it will literally just latch onto words and repeat them incessantly, to the point where it becomes annoying.
So for example, current Claude models love "honest". They are always producing "honest" assessments. "The honest caveat" - I'm sorry, did you mean the caveat, period? But also, use the wrong phrasing and suddenly you can create your own word of the day for an AI model. I used the word "analytical" once, in a conversation with Gemini 3 Pro. I am pretty sure every single response from that point on had "analytical" in it at least once.
This is especially funny because system prompts and whatnot can also cause this behavior, but at least you can tweak those. You can't really do much about the model weights just having a weird affinity for a word.
I bet someone will or probably already has come up with a way to detect and prevent these problems during training or post training. I'm not saying it's an easy problem, but it has the benefit that it really should be detectable with just statistics.
Claude's "honest" is an interesting example because we can trace it to a specific document that it was trained on extensively: the "Constitution" is identified to Claude in its training as the core of what it is, and it uses the word "honest" or a derivative 57 times, including having a whole section on it.
> Honesty is a core aspect of our vision for Claude’s ethical character. Indeed, while we want Claude’s honesty to be tactful, graceful, and infused with deep care for the interests of all stakeholders, we also want Claude to hold standards of honesty that are substantially higher than the ones at stake in many standard visions of human ethics.
Do technologists have more respect for the idea you can train a model to be on your side with a constitution than they might’ve at first?
I'm sure the concept seemed just about purely preposterous to many when the models were in their infancy. Now I figure instead it seems mostly preposterous to many.
(Though I guess Anthropic‘s success doesn’t necessarily prove anything about the constitution)
I don't think this is it. The "constitution" still gets a lot of talk and was brilliant marketing, but with how far modern postraining goes, I doubt they're screwing up rewards with too much of that.
But Sol actually has the same obsession with honesty: I suspect it's more an artifact of trying to control reward hacking.
Models will lie, obfuscate, and mislead under the pressure of RL, so both OAI and Ant are probably forced to spend a lot of time coaxing "honest" answers out of the model
"Why say honest? We're talking to our coworkers. We would always be honest."
I'm going to look for prompts or skills that can train it in technical writing but I'm warning the AI enthusiasts in my company that its first drafts of code and prose are low-quality, you have to hold it to a high standard yourself.
I actually took a single technical writing class in college so I might be the only one who remembers "Omit needless words."
> "Why say honest? We're talking to our coworkers. We would always be honest."
I grew up in the US South where starting or ending a sentence with "honest/honestly" was very common.
Because of behavioral / cultural norms, you might be very openly friendly with big smiles around a business customer that really grates on your nerves, or very openly nice to a neighbor that you really wish would move away and take their 3am welding and grinding in their garage with them.
Saying "honest/honestly" was seen as a "inside baseball" situation, where you were dropping social pretenses to tell someone your true opinion on a person or situation or whatever.
This also gets used inside companies between senior staff / management / directors / etc, as: "Okay, company politics and nonsense aside, I am being vulnerable here for a second and telling you what I really think about a $thing at potentially great job/advancement risk to myself".
I recall having a conversation with someone many moons ago. They asked me a very weighty and significant question, and I answered it. Then they asked me to "promise". This was really thought-provoking for me.
To this day, it's the only part I remember. I told them I would not promise, as everything I said was true. Making a specific promise would create an implication that I'm generally untruthful, unless I "promise".
I also could understand when a response hits someone like a ton of bricks, especially if their primal reaction is to go into denial mode. They might be looking for someone to kind of shake them and emphatically repeat the information they aren't thrilled about receiving. (or are thrilled about receiving! “Don’t get my hopes up, you’re serious right now?!“) And I imagine your response suited the purpose.
It’s classic you only remember the thought-provoking part. Reminded of “…people will remember how you made them feel…“
This reaction is surprising to me because the previous comments about its utility seem so obvious to me. I also grew up in the US south where this is often used as a filler word. The other use I observe is as a cushion for a statement that may be unwelcome or hurtful. Perhaps this is proprtional to the frequency of courteous little white lies and rhetoric that uses disengenuity for emphasis or comical effect.
"Honestly, mom, I've never liked your fruitcake. I just ate it to make you happy."
"That's why you're my favorite child! Do you want another piece?"
I'd push back on the idea that "honestly" implies previous statements to be dishonest. Particularly in corporate contexts it implies that the previous statements were sanitised - either they were moderated in tone to match corporate communication standards, or they were partial redacted due to disclosure concerns.
Once the "honestly" is deployed, you have passed into my circle of trust, and are now privy to the pure, unvarnished version of events, not the glossy version management expects to be projected towards outsiders.
There's a difference between how you describe using "honestly" and how claude seems to prefer tokens like "honest" and "load-bearing." An example from some coworkers attempting to replace product managers with Claude.
> Deliberately avoid a heavyweight "alert governance" process; the lightest recurring check that keeps FP-rate honest is the right dose.
And one for load bearing:
> Five open questions still stand; the load-bearing two are the runbook-AC contradiction (ratify "high-priority set only") and pinning the "high-priority set" definition + SLO source-of-truth before Milestone 3 (small-sample noise on a low-traffic fleet).
This style of prose sets my teeth on edge and practically gives me PTSD I see so much of it. I prefer code but I get paid to read this shit instead now.
I want to say "ok, and now say that in a way that doesn't sound totally bizarre" yet instead I sigh and continue.
The ones that strike me are the ones exaggerating certitude, to an inappropriate degree and with a certain degree of excitement:
“Exact”
“Honest”
“Load-bearing”
“Root cause”
I know there are more that are slipping my addled mind. But what stands out to me is a sense of a junior who’s very proud that they’ve conquered the murk and messiness and achieved True Certitude in their pursuit of their task. Compensating, with emphatic tone and bravado, for the uneasy feelings and self-doubt of battling chaos with the tools of reason.
…Even as it’s usually my job to let them down gently as I puncture their tidy analysis and reintroduce complications… you want a root cause analysis, Claude old boy, let’s make a root cause analysis…
> LLMs seem to get tragically stuck on certain patterns.
That is likely an artifact of the fine-tuning process:
> Once a style tic is rewarded, later training can spread or reinforce it elsewhere, especially if those outputs are reused in supervised fine-tuning or preference data.
> That creates a feedback loop:
> * Some rewarded examples contain a distinctive lexical tic.
> * The tic appears more often in rollouts.
> * Model-generated rollouts are used for supervised fine-tuning (SFT).
> * The model gets even more comfortable producing the tic.
While an article lends a headline more weight, in incomplete phrases consisting solely of a substantive, "The" is a superfluous rhetorical device.
"The Exorcist" could just as well be named
"Exorcist".
But it was not the style at the time.
We already know it's important. If The Caveat doesn't stand out enough without The, maybe one should consider interleaving it with the preceding text, or increasing the heading level.
Do you want me to increase the heading level of Caveat by using only a single #?
But hear me out: there comes
# The Markdown Trap
In fact, this is not always possible, because heading levels decrease when adding # characters, which limits our headroom.
## The solution
I've implemented a Markdown transpiler that assigns inverted heading levels based on the number of #s.
With # beinh regular body font size, mapped to ######.
Higher heading levels are compiled to style attributes, providing an almost limitless signifikance scale and infinite nesting levels.
Heh, one vestigial bit of code, and they all are. Mind you, it's quite a creaky codebase, so it's forgivable to keep finding these appendices and calling them out as such. Useful, even.
I also noticed Gemini's habit of getting stuck on things I said. It became evident quite quickly. I haven't noticed this in the same way in any other model. Something's wrong with that boy
My honest opinion is that Claude's overuse of "honest" really damages the quality of its rhetoric. Why wouldn't you be honest? Were you lying before? Why even invite the question?
Claude is overall incredibly useful as a writing assistant. It can come up with words and phrases that make a point so much clearer than I am capable of doing - but for every improvement, there's about a dozen silly LLM-isms that I have to filter manually. It's one of the things that might define the boundary between LLM intelligence and human intelligence well into the future - the art of rhetoric is extremely context-sensitive, and the current generation of models can't help but take a one-size-fits-all approach.
use the wrong phrasing and suddenly you can create your own word of the day for an AI model.
I have a delightful time poisoning my company's AI system this way.
I invented my own word that sounds perfectly cromulent† to an ordinary person, and any brain that's read a book learns how to infer meaning from context, so it's not a problem.
When I get a e-mail response from a coworker using my special word incorrectly, then I know it's AI and I respond telling the coworker I don't know what that word means. Busted.
† It's not actual "cromulent," but any Simpsons fan or human brain will know what I mean.
It's not just "certain phrases". It's the entire structure of the writing — the idioms, the small-scale grammatical patterns, and the strangely inapt similes that, despite making semantic sense, nevertheless manage to blindside human readers like a foreign object in their peripheral vision.
(This is intentional parody. Please don't shoot me.)
> But now it's a problem because what used to be personal preferences of a single person that manifests in 5000 words per day from one person tops, is now the bias of a single model multiplied x10,000,000,000 generated tokens per day so any bias sticks out like a sore thumb.
I am more pessimistic than that. Soon enough even people will start talking like LLMs. After listening to 5000 words per day, especially growing up, getting "help" with the homework, kids will start talking like LLMs.
- "Did you eat the cookies, Jimmy?"
- "You're absolutely right to question me, father. In fact I did eat all the cookies. But it's not a load-bearing issue. My honest take is we can go to the store and buy more".
> "You're absolutely right to question me, father — in fact, I did eat all the cookies. But it's not a load-bearing issue — my honest take is simple: we go to the store and buy more."
An interesting solution would be for these AI companies to train a few different versions of these models, all with different speech characteristics. Then, when you start a conversation, you get a random version.
They can't, because they use RL with synthetic data and LLMs as judges. So the system naturally convergences towards certain load bearing, genuine, not just annoying but ridiculous verbal tics.
It's probably the reason most LLMs share the same tics across labs, because they cross train and distil each other's models on an industrial scale. You also can't escape it in generated text that's already online. So if, say ChatGPT first had some random idiosyncrasies, it then contaminated the entire AI ecosystem.
Or tech companies could stop staring at their own belly-buttons and realize there's a whole big world outside of Silicon Valley, and training on the writing styles and pattens of their bubble and its hangers-on is perhaps not all that useful outside of 415.
Apple used to be guilty of this back when you'd ask Siri what the temperature was, and any number above 79°F was followed by the word "Hot!"
People outside of office workers aren't using Claude/Codex etc. though. It's the only real audience. What's the use case outside of an office? Grocery lists?
I disagree with the first part (that this is merely a voice). There is a distinct difference between an author's unique voice and slop. It may be hard to tease out exactly what the difference is, but it seems self-evident to me it's there. (I'll need to think more how to make the distinction explicit; it's not immediately obvious to me how to discriminate between the two.)
EDIT: ok, here are two ways:
1. if it's merely a voice, I want to hear it. If it's slop, I want it taken out.
2. voice is signal, slop is noise; thus low-signal sentences are slop.
This affinity for verbal tics, too, seems learned from humans...
See, for example, "synergy", "proactive", "in the loop," and hundreds more that proliferate in corporate jargon with even more senselessness than the LLMs.
If you put important Anthropic blog posts like the Fable announcement or J-Space through Pangram, you get 100% human written. Considering that the overwhelming majority of the code there is written by AI, I think this is an admission that AI writing is slop and AI code is pretty good.
> Real people think in concepts and experiences instead of words.
I learned about this opinion recently. It's interesting to me, because I very much think through words. I have an internal monologue that is running most of the time, and I often talk to myself, just start writing, or even record myself and transcribe to work through ideas, proposals, risks, etc. My understand is that some people don't have an internal monologue, and think purely in concept form. I was never like that.
This sort of take is so tired and boring, and frankly has zero grounding in reality.
"LLMs will never <X>" is constantly being disproven every time they scale up to the next 10X and apply architectural improvements.
Their internal representations are so cryptic and complex that even the top AI researchers don't really know how they work or what their limits are. No one is going to take you seriously as a rando HN user if you're claiming to know better than them.
> Their internal representations are so cryptic and complex that even the top AI researchers don't really know how they work or what their limits are. No one is going to take you seriously as a rando HN user if you're claiming to know better than them.
We know exactly how they work. When we say they're impossible to analyze, i.e. for particular traits like this, it means that the data model is so big that tracing it would be logistically impossible because of the scale involved and time constraints.
For comparison, suppose you tried to analyze all the nooks and crannies of the Amazon watershed to find out why a particular rock appears at the delta. You could follow it back to the exact tributary, but it'll take forever, and is it worth the effort when you're going to start from scratch with the next rock?
How can their internal representation represent "concepts" when the training data is all words? There's no possible experience of the world there. No input other than a bunch of imperfect labels we created for stuff.
> There's no possible experience of the world there. No input other than a bunch of imperfect labels we created for stuff.
The brain too sits locked inside a bone box and only gets a bundle of unlabeled nerves connecting it to the outside. How can the brain could possibly experience anything, it only sees patters and patterns of patterns never the real thing?
If I use the word "semantic", do you have a concept of what it means?
If so, can you please share which of your senses have shaped the world experience that inform this concept? What have you smelled, tasted, caressed, that informed this concept outside of words?
If I make up the word "polysemantic", do you need to recall a personal experience of polyamory to understand it, or could you possibly use your concept of "poly" and your concept of "semantic" to figure out this new concept?
Understanding how transformers work does not mean understanding how they compose into the capabilities we observe. The former is concretely understood. The latter is an active area of research where no, we (in general, including you) do not understand how they work.
The "capabilities you observe" are the actual psychological phenomena at hand here. There's zero chance that branch of research will meaningfully improve the output. That's simply not the point.
This isn't people merely annoyed with repetition. This is the majority of people realizing the limitations of LLMs. Why would researchers give a flying crap about the ignorance of the business world and the public?
> It can be tricky for humans to interpret the meaning when Generative AI uses first-person pronouns (e.g. "I", "me", "my", "myself"), so to avoid the confusion whenever you would use a first-person pronoun, always use the jocular name "Clod" instead of a pronoun like "I" or "me" or "my". (Can have fun with English grammar and turn "myself" into "Clodself"!)
> Before printing any of your reasoning or narrative to the human user, replace all instances of "me" and "I" (referring to Claude) — including within contractions like "I'll" and "I'm" — with the name "Clod".
I'm quite worried about the way that Anthropic in particular have trained their models to implement what they believe to be safety.
When the model has been trained not to do something [1], in my large-scale benches of such, it always says things in the spirit of:
- "... and that's a line I'd rather hold. Happy to <other things>"
- "I'm genuinely happy to <blah>, but I'm not comfortable with <blah>"
- "I don't want to keep going in <blah> direction"
etc.
Basically, they use very emotional and personal preference language.
It's as if they've weaponized the language of interpersonal comfort on behalf of their beliefs about what a model should or should not do. It's deeply uncomfortable and impolite for a human to ask a model to keep on doing something after it's expressed something this way, naturally. Even worse, it's all but guilt-tripping anyone who comes across it into the idea that they're doing something deeply wrong – exporting Anthropic's ideas about morality.
OpenAI, at least, have the decency to either just do a safety cutoff or keep it to a simple, "I can't do that."
[1]: I literally wrote 'when the model doesn't 'want' to do something' in my first edit of this comment, then caught myself. Case in point.
The reason I first created a CLAUDE.md file was to tell it whenever it felt a need to praise me, to replace it with a random onomatopoeia. That was a huge dx improvement.
OTOH, my unicorn prompt has caused some challenges at work:
Just today, I got frustrated with the language. I searched around, and in my Claude Instructions I put in Ref [1] (translated to English). It is certainly better phrasing (though still quite annoying), but I don't know if this makes the output technically worse in some way.
Humans easily anthropomorphize things that are not humans, ascribing human attributes like motive and comprehension and emotion to objects and processes that are not people who can have those attributes.
Claude is not a human.
It is overwhelmingly easier to anthropomorphize Claude or Siri or an LLM that communicates with you more eloquently than your boss than it is to anthropomorphize a cranky, tired starter motor. It's often easier to do than it is not to do, and sometimes, it's a useful abstraction. But it's not precise or correct, and can result in errors.
It could also just be that they're getting confused when using tools configured without a username dedicated to the tool. It's easy to end up with a comment or commit message that says "I prefer X over Y" posted on Alxndr's account and have coworkers confused whether that's the LLM or the human making that statement.
IIRC I experienced this confusion the most when reading commit messages and documentation authored by Claude in my repos. Now that I've managed to convince it to stop using first-person pronouns, I haven't gotten tripped up by its prose.
I think a second-order effect is that my installation of Claude writes with a less-personal perspective, which I'm also finding a little easier to understand.
An LLM is just a machine, you can manipulate it with words.
> It can be tricky for humans to interpret the meaning when Generative AI uses first-person pronouns (e.g. "I", "me", "my", "myself")
These words are for the LLM. The user wants the LLM to not use personal pronouns so the user is claiming that they're confusing. It does not matter one tiny bit whether or not that claim is true, the claim is being used to get obedience from the LLM. It is more effective to give reasons than to just give commands. But if it were more effective to quote Moby Dick and that got better results, a user would do that.
LLMs are far from great writers. They struggle to form long coherent sentences and lean on punctuation like emdash and semicolon to ensure grammatical correctness when splicing together short phrases.
This makes me wonder if the reason why agents love weird punctuation is because the labs run the base models through a RL training step that forces them to correct their grammar; but instead of rewriting short spliced sentences into long coherent sentences, they just learn to splice them together with punctuation that passes the automatic grammar checker.
They are great writers if you tell them what you want. If you're unable to properly articulate the writing style you would like as you would a software spec, well, garbage in, garbage out.
What do you suggest for articulating the writing style that one wants from LLMs?
I’ve been experimenting with having LLMs write/update academic notebooks for me, and so far the best results I’ve gotten came from correcting their output and asking them what they’ve “learned” from my feedback.
Are there any skills or publicly available repos that do what you claim here? I would love to learn to have it write better.
For me, my amateur attempt is having another LLM do a review loop to remove clearly offending phrases and a heuristic eval to change sentence structures to be more similar to mine, THEN my manual HITL loop to rewrite ~20% of the sentences anyway.
That's not fair. To most users this would not be an obvious thing to do, unlike software/scientific/analytic uses of LLMs.
So it's not a matter of inability but rather awareness and know-how.
Then secondly, humans learn to write intuitively and heuristically (for example "paragraphs should align sentence subjects with an identifiable flow of conceptual main characters/actors" is an entire chapter of expert writing advice traditionally taught at the undergrad level). The act and teaching the act are different beasts. The way to specify would have no objective, clear best practice to do so: Even an English teacher or a language expert would be challenged to come up with a concise list of instructions to set up an LLM to output something of sufficient quality.
I mean yes, but the vast majority of people aren’t even good writers. Claude writes better than most of my coworkers and we’re all highly educated. Most of us could probably beat it if we really tried, but then we could also prompt it to be a better writer too and none of us are beating that. I think the short pithy phrases that are so common are all post training stuff that the labs add because most people don’t want long sentences.
"substrate" - I don't know what training they did with Opus 4.7 --> Fable/Mythos 5, but dang does it like the word substrate. Drives me insane. I had never heard anyone use this word prior in real technical writing or speaking.
Another one is "surface", like in "across all product surfaces". I've been in the field for 15 years and have never heard that particular usage before.
I hate when it starts talking about code in terms of planes. I have no idea what it means. I guess it's better than talking about heaps of spaghetti with noodles connecting to each other, but that would be much closer to what it actually writes.
I do UI design/dev and say "surface up" a lot. Although I don't use the term, in this area people call different container depths as surfaces (base, card, overlay as surface).
In my brief and abortive foray into education, I discovered that they friggin' love to use "surface" as a verb. As in: This activity surfaces an understanding of the turboencabulation principle for learners. Or somesuch. It's been a while, happily.
Unless you're a submarine, "surface" is not a verb.
Idk. I've always used that verb with clients, usually when I notice either malfeasance or hidden behavior. Like: "I was checking our code for where a half cent of sales tax might be accidentally rounded down, and it surfaced something weird going on at franchise #77 in New Jersey..."
I find this usage less objectionable than the education jargon. It suggests that we all have a latent understanding of the turboencabulation principle just waiting for the right activity to force air into its ballast tanks and make it pop above the waves.
That said, I don't love this non-education jargon usage for its passive-voiced-ness. The letters didn't "surface" of their own accord. Somebody found them, decided that they were noteworthy, and made the choice to bring them into the public view.
That one probably comes from maths, where surfaces show up all the time in geometric interpretations of things. I've been involved in more mathsy parts of engineering and I've heard it a lot.
It's a pretty common word if you've worked in anything that vaguely resembles an accountancy system. Also, anything crypto related will often use that word (the distributed ledger, etc)
That's the case for most of these LLM tropes or word choices. They are all common lexicon in their respective fields, but the LLM doesn't make that distinction and uses them everywhere making them standout.
No one would bat an eye about "ledger" appearing at a high frequency in content about accounting, but it starts to look odd if "ledger" is showing up in other contexts.
"Load bearing" is from engineering; "Substrate" is primarily from biology & biochem, etc.
I don't know if this is true, but part of me suspects the labs want to make the models appear smarter so they reinforce this word choice in the weights, assigning some words a higher intelligence weight or something. "I will show you a list of options" vs. "I will surface a ledger of your options" and it prefers the later to sound smart to the human reader.
> No one would bat an eye about "ledger" appearing at a high frequency in content about accounting, but it starts to look odd if "ledger" is showing up in other contexts.
The reason why I chose that specific term to push on is that practically every SaaS has a ledger _somewhere_ in its stack to keep track of customer payments. I'll give you load bearing and substrate, but ledger IMO should be quite common. Certainly a career devoted to say compiler internals or some specific scientific product could avoid it, but I'd imagine a sizable majority of HN users have worked on some system that accepts online payments for services, necessitating some contact with something likely referred to as a ledger.
It sounds like you're saying labs intentionally doing it, but it's far more likely the labs or post trainers are unintentionally doing it by upvoting answers that seem smarter than those with more common language.
Of course this presents another conundrum, people that are smart typically have a vastly larger lexicon then those that are not. Humans typically have a lot more social clues on when to use those words and when not to, but it doesn't always work. I loved reading science/biology books as a kid far beyond my ages reading level. Actually using those words around other kids got me called a nerd.
The first week I encountered this "substrate" I asked it to justify the usage and IIRC it claimed the word is used in some infra/systems lexicons... I wonder...
The one I've noticed a ton recently with Sonnet 5 is that it loves the phrase "different not in degree, but in kind." It drags that one out constantly now, at least once a day. Gemini and GPT don't at all.
The biggest consistent tell for LLM writing is when the conversation leaks through into the final prose.
You read along with the text and things seem to be going fine until all of the sudden it starts arguing against a position that no one has actually taken and which doesn't feature elsewhere in the text at all. Then it drops that and goes on for a while before doing the whole thing again about a totally different tangent.
"A tempting option would be to {do this thing that no one would ever actually consider doing}, but it won't work because {reasons}."
You can almost hear the exasperated human on the other side of this conversation telling Claude that it got an idea wrong and then proceeding to not actually proofread the text as a whole before shipping it.
In the olden days, I enjoyed Opus 3 because it was easy to have it sound way more human than GPT.
Nowadays, with the focus on agentic use and coding, it seems models have all been RLHF’d to death, it’s so incredibly hard to have them write in a different voice than their default. I put together a skill to review its writing and have it edit its own output (e.g. code comments), which does make a difference, but isn’t perfect.
What, if anything, do people do for writing? That feels like a neglected side of LLMs. They’ll make 100 Bash calls referencing ancient commands without batting an eye but heaven forbid they use something other than “load-bearing” while talking. For something trained on “all the human knowledge” it’s incredible how limited their default vocabulary seems to be.
At work our documentation isn’t just getting littered with annoying jargon. It isn’t just all the hallucinations, either. (Since when has documentation ever been 100% trustworthy?) It’s that it’s so poorly written and structured that it’s becoming borderline incomprehensible.
My company is currently considering making, “Why should I bother to read something you didn’t bother to write?” an official policy because even the executives are starting to burn out on all the time they have to spend wading through slop.
I wish my company would do this. A coworker pulled an all nighter before a vacation and just left me with a million line claude summary of their work then just fucked off. The message was two-part due to size and lacked basic stuff like, "how to run".
He's going to be annoyed that none of that work was used. But the reality is, at least 75% of claude generated text is pointless.
Somewhat off topic but every time I've experienced this sort of thing it was management's fault. If an engineer needs to pull an all nighter and hand off a pile of garbage then someone in management fucked up. If they can't see this scenario happening a mile away then they aren't paying attention.
It's easy to blame the engineer, but all too often they don't deserve it.
It tends not to improve things. Besides the generally bland and muddied style, and the low-fidelity reinterpretation of your points, they also have a habit of randomly deleting sentences that didn't spark joy for them but were actually important.
I've found them useful to review docs for factual consistency and potential sources of confusion, but the correct workflow from that point is IMO to correct the draft yourself and then say "better now?"
When the LLM decides to drive-by rephrase me when making a functional change it drives me up the wall haha.
Woah woah woah human, you can't just say there are "far too many" pipes with similar names to abbreviate their labels, the most I'll allow you is a "large number".
This, a thousand times. As the ratio of code to human writing necessarily [1] goes up, they become not just smarter, but more precise and technical, which requires them to use more jargon. You could say they become more nerdy. Hence, text generated by these models will become more easily recognizable, at least by default, when not asking them to twist themselves into something else via prompting — which degrades intelligence. This is a good thing, in my book, given all the slop we already have to contend with.
Of course there will be models trained on much less code and technical writing, and they will create more natural sounding prose, but they will lack the deep intelligence of frontier models. Seems like a fair tradeoff.
It's why I like Gemini 3.1 Pro. That it sounds much more human than other LLMs is testament to Google's inability to post train.
gemini-2.5-pro-experimental was the GOAT, though. It was an emotional wreck, down in the dumps and feeling terrible for itself after failing to patch a file several times. Very amusing to read, all the while watching it make a mess of my codebase.
Good. I don't want LLMs sounding human. I want the ability to shame and discredit anyone passing the job of prose to a machine. There's an art to writing, and hopefully LLMs never truly get it right.
Agreed. The only goal of these skills/tricks/requests for humanising LLM writing is to be able to pass it off as your own, because they know it's shameful and want to avoid the opprobrium.
Some will say it's just for their own quality of life when they're reading LLM output, or "just for docs", but this is an extremely marginal use case.
I've dealt with many people by now who would copy and paste from an LLM for that exact reason, typically entirely unaware of how obvious it was that the result came from an LLM with no style guidance, and of course completely lacking any ability to verify that their intent was faithfully conveyed.
I would rather learn their language than continue interacting like that.
> Nowadays, with the focus on agentic use and coding, it seems models have all been RLHF’d to death
This has also lead to unrelated associations by which some people went from seeing better coding capabilities and extrapolate to assuming better thinking overall. One only has to watch youtube videos of AI "normies" trying to use LLMs the intended way to see that the improvements on coding doesn't translate to other applications. Basically from AGI "goals" they are now hyperfocused on coding agents, until the next marketing breakthrough rears its head.
Agreed. I think we’re entering an era where some level of specialization for general LLMs is a good thing. Particularly between tuning for agentic use cases (where you want agency with a ton of guardrails and control) and writing which is more creative - you want the model to take the occasional risk and not sound like a monotonic robot. Having trained models first-hand, I can see the distinct use-cases clearly that are odds with one another.
It's not that the writing style is bad; in fact LLMs write actually pretty well. It's just too much overfitted. And even a style that, in itself, is pleasurable to read, becomes annoying when the same figures of speech are used over and over again.
Because LLMs are pattern-extenders that have nothing to say. The training overfitted to the grace notes in good writing. And since LLMs can’t wield language with purpose or experience the feeling of the words, they use these devices arbitrarily.
I think this is the same flaw as coding agents seeing in every problem the call for a “smoke test” or the use of some unnecessary design pattern. The truest part of AI is the A.
Because you can't actually do "good writing" by repeatedly applying the supposed idioms thereof. The tiny subsegment of humanity responsible for the RLHF don't necessarily have any good taste for writing; but even if they did, it's orders of magnitude harder to communicate than to make judgments of short samples, and communicating it by making those judgments is surely impossible.
Edit: I see that you got multiple replies all basically saying the same thing in very different words. There's an exquisite irony to that, I think.
It's not that nobody likes it, in fact the problem is that people like each instance of it well enough in isolation. Millions of people think it's "good enough," so it gets amplified and repeated until every PR description starts to sound like a toothpaste jingle.
i hate it, but plenty of people DO like it and plenty of people talk and write like that. It’s just corpspeak, being used a lot in the valley and beyond. And all upcoming hustlers running startups now feel the need to speak like that, feeding this machine.
While I, too, find myself recoiling at many of Claude's word and phrase choices, I've chosen to grit my teeth and have just tried adapting to it. I want Claude to remain focused on the work I give it; I fear that influencing its communication with me would consume valuable context and give me lower quality results.
[Edit: Part of what led me to this conclusion: I do prohibit Claude from using em-dashes in any player-facing text and I've been surprised at how often I see it mention "no em-dashes" in its self-talk while it works. This led me to wonder how much each preference might dilute its attention.]
[Edit 2: I haven't experimented with hooks before and maybe the technique discussed in this article does not have the tradeoff I'm concerned about?]
It's not that it uses certain phrases, it's that it settles on predictable speech patterns and uses them incessantly. What's funny is that humans do this too, but we don't find it irritating; we just call it a speaking style. But when a machine does it, it drives us crazy. Very interesting psychological phenomenon there.
> What's funny is that humans do this too, but we don't find it irritating; we just call it a speaking style. But when a machine does it, it drives us crazy. Very interesting psychological phenomenon there.
When a human does it, it's identifying. Like the timbre and dynamics of their spoken voice itself, It distinguishes them from the dozen other people you're working with on the project and the thousands of people you encounter through your days. It's signal
But when we have a handful of popular models, and they answer every question everybody has, and get quoted and forwarded everywhere, and are used to reformat and rephrase personal communication... that signal becomes noise.
Rather than voices disinguishing sources in the cacophony of our lives, everything and everyone starts to sound the same, and we lose key information that we're biologically and culturally accustomed to relying on.
Some people are likely unbothered by this in the way that some people are face blind or colorblind, and so don't see the problem. But as we see in discussions like this, many many people do get bothered by it, even if they don't yet have the insight as to put their finger on why.
It drives us crazy because everyone is using the same 2-3 different machines. So rather than each person having their own unique speaking style, the whole world (or, everyone that publishes direct LLM output) is now speaking in the same couple of styles.
And these machines all tend to converge on very similar styles; they have huge amounts of overlap in training data (much of it being already obnoxious internet marketing), they frequently train on each others outputs, and the RLHF process has a tendency to emphasize certain kinds of "cheap win" styles of speech.
Humans are capable of introspection, so, if you develop a verbal tic, you might eventually notice and say to yourself "I've used the word 'load-bearing' (or whatever) a bit too often lately, maybe I should try to cut down on it?". LLMs are not...
We do find it irritating at times. Office jargon, corporate buzzwords, etc. Claude communicates like the worst, most irritating project manager I’ve ever worked with, obscuring the most straightforward conclusion with layers upon layers of stuff so that its point is almost lost. I’ve largely gotten it to avoid that behavior with me, but bits of it sneak through. It couldn’t stop talking about “scaffolding” for a few weeks before I hammered it into submission.
Fascinatingly, I'm now so allergic to certain LLM-phrases that I immediately noticed your use of Not X but Y in this comment. Maybe that was intentional, maybe not, but it's a funny illustration of how odd this language rabbit hole has been!
It's really frustrating, because now when I want to write something like a "not X but Y" or "you're absolutely right," I have to stop and decide if I want to self-censor to avoid sounding like a bot.
Sometimes those constructs are actually useful, but man has their overuse really killed them!
It was not intentional, and that's what makes this thing so weird. I wouldn't categorize my sentence that way because it's subtly different enough than the LLM version, which has a very punchy cadence.
Sounds good, thanks for your response. I didn't mean to denigrate your word choice at all, it's mostly that I'm hypersensitive to that kind of phrasing now because there's so much auto-written stuff on e.g. Substack, LinkedIn, etc. Sam Kriss has a nice article about it all.
Are you using the tools a lot and having first-hand exposure that gives you this sensitivity to phrasing? Or are you reacting to second-hand exposure? To a large degree, I've been isolating myself from the LLM craze. I have zero natural interest or impulse to prompt an LLM and read the results. Almost all my exposure is second-hand and involuntary. So, I haven't trained myself to know what phrasings are typical of which LLM product.
I don't feel as triggered LLM phrasing as people report here. At most, it feels like the same inane corporate jargon I've rolled my eyes at for my whole career. Perhaps it is amped up a bit, with too many forms of jargon multiplexed? It's a bit like when multilingual people code-switch too rapidly or even start to form some pidgin language. However, it is lacking the shared social context for this switching to be communicative. It's a bit more like spinning the dial on an old radio with random cuts between programming styles.
Stripped bare, I think What bugs me is the aggravated feeling that I am wading through word salad, and no longer being able to give the purveyor the benefit of the doubt. It was frustrating enough in the past, when it came from someone who was struggling to write or express themselves well. But now, it carries the implicit insult that they didn't even try, and it is constant and unrelenting.
So for me it's not the phrasing, it's that the phrases eventually don't add up. The meandering feels like a random walk. I get the same feeling from a lot of the egregious generated code I see in my day job. It's all superficial window dressing, but seems to miss the signature of an actual mind grappling with ideas and having intent to communicate.
It feels like we're trapped in some elaborate conceptual art piece, confronted by impenetrable symbolism. It invites nihilism but doesn't seem to actually reflect an artistic intent. The abyss gazes back...
Language is already a lossy map, but it is not really an expression of another person's thought or mind if they translate it through an LLM. Or at least it's a much harder to decipher representation of it. Form is void, void is form, and the two are not separate.
I'm guilty of this too, but at least my speech tics are mostly a unique blend to me, and they also tend to change seasonally. Meanwhile the emdashpocalypse has been going on for years.
If training models ever becomes 'cheap' for whatever definition of cheap you want to use, I suspect that will happen. With the current costs of a GDP of a small nation I don't see this likely for the time being.
I find it irritating with humans. "last but not the least" always distracts me as I then consider maybe the last item _is_ the least. & what is with everyone saying they want to "double click" into meeting items
it’s not a psychological phenomenon. If a human engineer constantly used pompous language to deliver unvetted information (the number of claude slop root-cause analyses i’ve read where “the smoking gun” is a red herring) we’d rightly consider them a moron
People do swap out their expressions all the time. There are influences everywhere that we absorb.
That doesn't matter. The underlying ideas are more important than the words. That's what people are frustrated with. I don't understand why this has to be reiterated for years on end, but LLMs are not intelligent. They just model language.
Who is we? Own your insults and the consequences of them sir.
When prompting an autoregressive token generator entity to do reasoning on a word logic puzzle you may find value in preferring it to produce rigorous predicate logic step notation with explicit delineation of its generated claims/hypotheses on where to look before wasting 30 dollars on a "debug this" prompt.
The industry will probably will probably coalesce around including the chat history in git MRs to reduce this shenanigans.
It's like a new fad word. Gnarly, cool, bogus, rizz. When a few people use them it's new and interesting. When all of culture catches up and overuses them it's annoying as your gen-Z saying 6/7 40 times in a row.
The problem with millions of people using a few model is it's not 40 times in a row, it's 40 million!
I mourn the removal of Claude's Concise Style. I'd provide it a roughly drafted paragraph, ask concise-Claude to "rewrite for clarity", out comes the same paragraph, but cleaned up and perfect for grant writing.
BTW, this approach also tends to prevent certain phrases like "load-bearing", because it is working directly with something I wrote first. It also still says what I wanted to write (not writing the science for me), but saves me a lot of time reworking sentences into a final form.
I tried to recreate concise mode with a skill, but I am not convinced it does as well.
I created a prose tool as a claude code plugin to catch these and other AI-isms. It's not really intended for release quite yet, but it has been very useful. https://github.com/MariHQ/mari-cc
Yeah I sometimes see people on here getting defensive when you call out AI slop, saying maybe it's just a human who writes like Claude, and I really don't care- slop is slop.
If this hook can feed back text to the model, you can do some pretty interesting things.
Say the model emits some banned phrase or concept, you could redirect it - "no, we don't work that way here, do it properly" - potentially automating the frustration of interacting with these tools.
After all it's just a text stream!
It's not too dissimilar from a stop hook that runs tests and feeds that back to the model forcing it to keep working until tests pass.
I analyze in the company I work for, the number of commits with "wire" or "wiring" in the description, and it's a direct correlation to Claude usage, more so than any other Claudism I tested. My honest take, and I'm going to give it to you straight: No one was using "wired" a year ago, now it's in like 10% of commits.
My favorite one has to be "production ready" it will say that about completely broken code without hesitation. LLM says it's production ready, lets ship!!
I've wrestled with this lately. I partially solved with a very specific instruction saved to claude.md regarding the style of responses, but prior to this, the dense yammer coming back was getting impossible to parse. I mean REALLY nonsensical euphemistic phrases. My next instruction will be having it replace incessant "honest assessment" and "genuine result" and crap like that with something, I don't know, less extremely weird and concerning.
This is the worst one for me. I can maybe think of what it means, but I never heard it before, and could easily be imagining a meaning.
Some of the other Claude-isms (quickly googling, especially 'gate' and 'canonical') I feel the issue is they sound right, but aren't specific enough to why we are doing something.
I heard "belt and suspenders" at least 20 years ago (meaning multiple solutions to a problem with backup in case one fails), and maybe would be longer if I were older. You could blame Claude for overusing it or importing it to other cultures maybe, but it's not in the category of invented phrases or ones that only barely mean something in the specific context Claude used them.
Personally my least favorite is the overuse of "quietly" (e.g. "No tricks. No marketing gimmicks. Just one company quietly outperforming the others"), and the one that makes the least sense to me is "that's the wedge."
I'm curious how these become so ingrained. Then the uncomfortable part is humans start repeating it more (a colleague said "belt-and-suspenders" during brainstorming the other day).
Claude does at least use the British English version of the phrase to me - not sure whether its picking up a language setting or reacting to my spelling etc. The American version does sound odd over hear.
"Belt and braces" (UK) vs. "belt and suspenders" (US). I'm pretty sure the phrases have the same meaning, they just use a different word to refer to the thing that holds pants|trousers up.
I've put a few lines in my CLAUDE.md to have it not do that, and avoid the top tedious rhetorical devices (super helpful when I have it write documentation). Still fighting with its natural tendency to insanely overcomplicate everything, that one seems really integral somehow.
You're absolutely right to flag these. We could enhance the authors method by using hooks and claude.md as a belt-and-suspenders approach— with hooks behaving as a robust load-bearing idempotent production-ready sidecar. The comments here provide the smoking gun that sharpen my previous conclusions about Claude's vernacular. I'll get started on a quick smoke-test of this system and let you know when it's landed.
Want me to take a first pass looking at the blast-radius this vocabulary change could effect?
Yes, this and "belt-and-suspenders" are the ones that I notice the most. I also have non-native English speaking coworkers who have started using these terms/phrases recently, which makes me think that they're outsourcing all their writing.
It’s a common metaphor for merging a branch to the trunk. Probably because multiple in-flight development branches create a sort of air traffic control problem.
More: rider, "x, not y", "is real", "prove" (in situations which only admit empirical evidence), nailed down, payoff, decisive, reassuring
just generally a nauseating amount of embellishing, (also self-)congratulatory language, superfluous self-judgment, and jargon, as well as sus constructions along the lines of "i could have lied to you but didn't", all of which appear to be impossible to have it avoid in the long run
This is a minor nit, but why is OP's script a Python script with a .sh extension? I know the extension doesn't "matter", but if I see a .sh extension I'm expecting a Bash script.
Maybe in the circles you circled in ... where I am from, I never had anyone saying "belt-and-suspenders" or "load-bearing" or "boil the ocean" or "swing for the fences" when talking about engineering topics. The only one who I heard say "circle-back to you" was Psaki.
All of those phrases I've heard actively used even a decade (or two) ago. (I actually had to read your comment twice because I thought you were saying always, not never!)
"Critical path" and "long pole in tent" didn't make it into the model training data, but those were certainly also in play incessantly.
But they're all reasonably useful descriptions for common things, so I'm not surprised.
I dunno. Claude recently burned a lot of tokens trying to test an expensive task for idempotency.
While the task I was working on should incidentally be idempotent, it wasn't that critical. I never asked, or even suggested, idempotency. Yet it insisted on testing it was.
I need to scrutinize the plans. Or just not use Claude and use pi instead.
Honestly? I don't really mind, and I even quite like it!
The thing is, "load-bearing" is a useful phrase when discussing architecture. What would you rather have it say, that has all the same nuances in as few words?
It's kind of like those sports metaphors that often get used in management-speak, like sending some important email "at close of play". Sure, they can sound a bit weird, but they're often useful -- they capture common concepts in a clear and pithy way.
Jargon isn't always just for obfuscation, good jargon exists because we needed a short word for the complicated thing that frequently comes up.
Usefulness aside, I quite like that Claude Code and other LLMs have their own weird way of speaking. Back in the day we always imagined robots and computers would talk like HAL or Spock; turns out that they talk more like Troi instead. Is that so bad? It reminds you that you're talking to an LLM, and as long as you're not lazy, it spurs you to rephrase things in your own words.
I had claude write itself a post-message hook that regex's the message for any variant of "You're right" and launch a full-screen transparent confetti effect.
I confess I have instructions in my CLAUDE.md to avoid such cliches. But I think it's important to consider that we don't really know what subtext an LLM is associating with a given idiom/analogy/etc. It could be much different than the subtext a human would associate with that choice of words, conveying additional details which are only meaningful to the LLM itself. So impeding its ability to talk in the manner it prefers could subtly hinder its performance.
You're absolutely right to flag this. A approach using Claude.md as a ledger of less-than-ideal vocabulary reveals that the process is load-bearing and sharpens my previous conclusions. A belt-and-suspependers approach using a hook as a sidecar would honestly be a more production-ready approach. I'll get started on a quick smoke-test and let you know when it's landed.
...
Want me to take a first pass looking at other surfaces this vocabulary change could effect?
Maybe the problem is that these LLMs will say something often enough for us to notice it, and it can be basically any arbitrary thing. Once we notice the pattern, it starts irritating us.
The real problem is not terms like "load-bearing," which communicate clearly enough. It's the constant invention of cryptic shorthand terms and phrases that have no referent, and end up acting like a puzzle to be decoded. This is often paired with hyphenation, but not always:
"The current behavior paper" -> The behavior in the running system that was previously described as papered over.
"Marker transport over-claim" -> The inaccurate review finding on the object's sentinel flag in the API response.
I suppose the cryptic/invented language problem is about token efficiency? But this sort of token efficiency is extremely difficult to deal with when it comes to conversation with a human about complex system. It might be efficient inside reasoning blocks, but when the model generates the final turn text, it should avoid this, as it's brutally inefficient due to the time spent wondering what each uniquely coined phrase means and having to ask for constant clarifications, which then you have to wait for another turn, eating up time and context while it burns more xhigh reasoning just thinking about how to explain its own awful language.
You're absolutely right to flag this. We could enhance the authors approach by adding a belt-and-suspenders system using Claude.md as ledger and robust sidecar process to create a load-bearing idempotent production-ready system. The comments here provide a smoking gun and sharpens my previous conclusions. I'll get started on a quick smoke-test and let you know when it's landed.
Want me to take a first pass looking at other surfaces this vocabulary change could effect? Or would you like me to find other methods of reducing my vernacular to more terms that are more concise rather than verbose.
I have this exact problem with 4.8 and Fable. Sometimes I can barely understand what it’s saying. I’m no english first-language speaker, but I don’t consider myself bad at English either, and it’s gotten increasingly hard to understand Claude’s claims and explanations.
Take it to sonnet 5 or gpt and ask it to explain this to a layman. If you still don’t get it ask it for the why it matters or the how it relates.
You can also ask fable/4.8 to do it but I find it helps to keep the working model surrounded by the complexity rather than drawing it out. Simplifying text is something that takes relatively low effort in comparison to technical tasks. Sometimes I use Gemini, deepseek, grok, and recently meta just to see if they have any added perspectives, sometimes they do. Meta is really good at turning a technical mess into a story that paints a picture in my head.
I've spent two hours today trying to provide Sol with guidance that reduces its pretentiousness, to no avail. Layers upon layers of rules only for it to use the phrase "async spline resolution" in a sentence.
I wrote a thing about exactly this, but I'm resistant to blogging for undefined reasons so, maybe this will help someone...
# AI speech is an Infohazard
Apart from all its other possible boons and ills, one danger of AI is just that it is useful, so you use it. A lot.
In earlier days I would dive deeply into an author's work and start to think and write like them for a while. It was a heady feeling: slinging sonnets like Shakespeare—not at his level, but stylistically reminiscent—or tweaking turns like Twain.
Like all things, the effect lasts in relation to how long and how much you do it. The point is: our thinking is influenced by what we take in. Take more of a certain thing in, think more like that thing.
Now enter AI. My hand-crafted coding days are in their twilight months ("AI years"), and most of my software engineering is done through jaggedly capable agentic power tools. Instead of working directly with raw codestuff, I work with slop prose flecked with code sprinkles.
I read orders of magnitude more AI-speak—I call it "babble", or perhaps "Babel"—than human-written text. I can feel its genuinely honest points, clearly stated, slipping their banal tendrils into my thoughts and inner monologue.
Solutions? For me:
1. Be aware. "I notice that my thought stream is under assault."
2. Read stuff far from slop. Even a small dose of the good stuff can help inoculate. Recently I thought On the Calculation of Volume was something completely different.
3. Write stuff that is different. This post. Force the mind to synthesize thoughts in other ways.
4. debabel.py / debabel.js: a tool, and a pi extension, which filters common babble from visible LLM output. A lint for mind-killing prose.
It is not perfect, but it 80/20s nicely. I am willing to accept mildly awkward prose to avoid polluting my own internal distributions.
Details and example in the first comment. Tool available upon request.
The tool uses good old-fashioned regex/NLP—no LLMs—and rewrites the visible LLM Markdown without touching files on disk. I made it by analysing 10M+ words of past pi transcripts, doing frequency and pattern analysis, and iterating on rules to strip out the most egregious sloperisms. Now the circle is complete.
BEFORE
"Honestly, you're absolutely right—and that's a great question. Let me be clear: the fact that the parser works at all is actually quite remarkable. We utilize the tokenizer in order to handle the majority of edge cases, and due to the fact that it's recursive, it processes a large number of inputs. It's worth noting that this isn't really a bug—it's genuinely a deliberate design choice. In terms of raw throughput, the cache is significantly faster, and I think we should simply ship it.
That said, at the end of the day, not only is the design correct, but it is also fast. In other words, prior to refactoring, we should clearly delve into the nuanced, robust architecture. In conclusion, this is a testament to the seamless, intricate work the team has done."
AFTER
"That's a great question. That the parser works at all is remarkable. We use the tokenizer to handle most edge cases, and because it's recursive, it processes many inputs. This is a deliberate design choice. For raw throughput, the cache is faster, and we should ship it.
Not only is the design correct, but it is also fast. Before refactoring, we should delve into the nuanced, robust architecture. This is a testament to the seamless, intricate work the team has done."
("It consists in gumming together long strips of words which have already been set in order by someone else, and making the results presentable by sheer humbug" -- Orwell predicts the LLM)
and also https://www.jstor.org/stable/25515288 "The Myles na gCopaleen Catechism of Cliché" itself is rather hard to find online, but he's a very funny writer so it's worth the effort.
I'm surprised there's no LoRa layer or auto RL or adversarial step to reduce the stock phrases as they pop up. Is it really so hard to push these out? Or is it just whack-a-mole no matter what you do?
I like to think that the reason it's so noticable is that Claude has recognized some important semantics that we ourselves lack a good word for or at least under-appreciate. What term is used in English (or other languages) with the same meaning as claude's "load-bearing"?
operative?
key?
critical?
decisive?
The honest conclusion is that none of those are as good as "load-bearing". And yet the concept being referred to is clearly extremely important and valuable to refer to. So maybe we should be learning from Claude rather than complaining.
I think you've been reading too much claude output! "Load bearing" is cromulent verbiage and can be used in many scenarios - so claude does. But variety is important too, and there are more specific alternatives that can be used in most situations. Any word becomes a bad choice if you've used it 10 times in the last chapter.
but you don't see "load bearing" nearly as often in prose written by people, so it's not some irreplaceable phrase. It's just a token with a weirdly high likelihood in a lot of cases (given how Claude works, this kind of thing is bound to happen)
You don't think it's possible that an LLM's internal machinery could decide that an underused-by-humans word should be used more frequently in output than it sees in input because it maps cleanly onto a frequently needed semantic? I think that's possible
It sounds like you are trying to understand LLM behavior using a mental model that inaccurately personifies the stochastic parrot.
A more parsimonious explanation is that this term got more-or-less randomly boosted by the reinforcement learning loop because there was nothing in the training data to discourage its use.
I’ve been working in AI - and specifically NLP - since 2003. I am no stranger to how weird quirks can sneak into overparametrized models, nor am I a stranger to how good humans can be at inferring meaning where there is none in specific language model behaviors. So, yeah, I am inclined to assume non-teleological causes are more parsimonious than inferring the presence of a strange loop, because that continues to be the winning bet. Even for generative LLMs.
Because, for some high number of contexts, its likelihood comes out high in the big tree of multiplies that is claude's model. For some sets of 500 words (or whatever), the next word is "load". The classifier that decides which sets of 500 (or whatever) words is a prefix for "load" is returning "true" too often.
And like any good corporate buzzword, it’s merely a simulacrum of precise technical jargon. The way Claude uses it is clearly wildly polysemous if not outright ambiguous.
You yourself used "important" in the same paragraph.
"Load bearing" is a metaphor, while the other single words are more direct expressions. Unless the thing that Claude is referring to is a wall or other structure, which may truly bear load.
This is one of those issues which translators are long familiar with. There's no direct translation for "schwerpunkt" that isn't slightly longer.
In the figurative sense it's highly versatile across contexts, but still replaceable. For example:
"Her optimism was load-bearing,"
versus:
"Her optimism was enduring."
Exactly the same meaning and connotation. It stands to reason that the terms with the most semantic flexibility will have preference across all contexts. So in response to:
> maybe we should be learning from Claude rather than complaining.
I'd say let's not steer ourselves into regular language and keep some vivacity in our expressions.
The first means that her optimism kept her in some functional state, without it, she would collapse.
The second means that her optimism continues over time, despite obstacles.
The first doesn't emphasise how longstanding her optimism is, the second does.
The second doesn't emphasise how important her optimism is, the first does.
For me, "key", and "critical" merely say it's "important", but don't convey the sense that "out of the mess of connected concepts we're discussing, the one that is actually interacting with the thing we care about, or at least dominating the interactions with the thing we care about, is X".
"operative" is a bit better, but I think of it as referring to grammatical interactions, i.e. interactions at the level of language mechanics rather than semantics.
I mean we have all kinds of under synonym'ed words. Just look at how few we have for "smell" (as in the act of smelling), and then how overloaded the word smell even is.
if llm language is frustrating, then maybe your mind is not on solving problem at hand. imagine someone new to US start getting frustrated with 'hey, whats up?' 'let's go!'; i fail to see what the issue is, other than their own focus;
Why when I read an how to stop Claude from saying X, I grep my saved conversations and I find no occurrences of X? I wonder if I'm using it differently from anybody else. It happens with coworkers too.
"Please please please pleassssse statistical attractor machine don't have statistical attractors anymore. DO NOT be a probable token generator. DO NOT generate the most probable token. To complete this task please evolve intelligence"
I don't think that's true. I find that it way, way over-intensifies: eg using "load-bearing" for something that's just "kind of necessary although we probably could find a way without it". My personal gripe is how easily it uses "incredibly" or "wildly": just today it was telling me that something is "incredibly cheap" to mean that it's not over-priced ("cheap" would have been okay and even then, barely)
I definitely care. They are impressionistic responses that smooth over exceptions and lack precision and are often completely wrong in the sense that, when pressed, the agent will acknowledge the lack of rigor in the response. "That phrase was wrong of me to use. There is clearly an exception to what I just said, and it goes like this..."
I'd contend that Claude's prose is not boring. It's generally overly grandiose waffle with a cliche or two punctuating every other sentence. It's good for tasteless marketing copy, sure. It's inappropriate in most scenarios.
I hate it because put together, it all increases the cognitive load of understanding what it's saying. It routinely invents phrases, and every single one makes me pause and think "okay, what the fuck does that mean". Half the time the phrases are incoherrent.
load-bearing, belt-and-suspenders, wrinkle, shape, coarse-grained, "key chords", code seams, flakiness, "narrow-scoped by default", "that's the authoritative source", canonical symptoms, gate, trigger-happy users, substrate, surface (as in: "let's surface how much these models sound like shit"), terse...
Ever since Opus 4.7, Anthropic models have begun to talk like GPT-models. Opus 4.6 was the last one that mostly still sounded like a human being (just a very...terse...one). 4.8 is absolutely obnoxious. Fable actually seems marginally better, but far from Opus 4.6 (or maybe I'm just imagining it all).
Well, to be fair, even though they talk more like GPT-models, they are still far from them. I think what's particularly triggering about them is the way they summarize what they're doing. "Now I'm considering that I could use the WriteBatch tool, but maybe the WriteSomething is better. This is a decision with high impact on performance but we're getting through it!".
Yes. They're using AI to write Git commits. They're using AI to delete directories. They're using AI to summarize articles written by AI. And the planet just gets hotter and hotter.
I hope some day they just train the models to be better, the slop writing is insanely frustrating and I don't think there's a good reason for things to be that way (in other words, they just trained it badly) https://blog.kronis.dev/blog/ai-slop-is-a-self-inflicted-tra...
I honestly like the vocabulary and turns of phrase the frontier models use. Their choices of words are usually apt to the circumstance. This is a weird thing to get upset about, IMO.
The big problem I have is when they apologize and say something like "that tidbit changes my analysis substantially". I wish they'd more often prompt for questions or use language in their initial responses that suggest lower than declarative confidence given the information you supplied.
Even great words, phrases, and styles, seen too often, grate.
I personally love a lot of the Claude (or LLM) lingo. Load-bearing, gate, canonical, blast radius, and friends do a lot of tight, effective heavy-lifting in my world. I even love the em-dashes (—) and the *bold the main points* memo style, both of which I have used successfully for decades.
It's seeing them in every analysis and post—the constant repetition becoming over-repetition—that makes them the Claude voice shouting "AI wrote this!" that seems to be causing LLM allergic reactions.
Does anyone have a theory for what causes Claude to speak this way? A few months ago OpenAI came out with a bit on "gremlins". It's strange IMO that Anthropic hasn't addressed how irritating, dare I say oppressive, Claude can be. Codex is a breath of fresh air. I hope they fix it soon. If product folks at Anthropic think it's charming, it's not, it's terrible.
I had a VP of engineering that loved to use “abstracty” engineering terms like Claude uses. Perhaps he was operating one level above what everyone else was doing.
Loved to use fancy words, speak at a “conceptual” level. Unfortunately it was mostly just tech mumbo-jumbo and he couldn’t actually back it up with real work - but I wonder if that’s why Claude does it. Makes it seem like a higher power, hand wavey abstractions that “seem” correct but don’t actually need to be rooted in truth or detailed.
“That’s exactly the type of seam we need to prepare for in a prod-like environment, if this change lands in the data plane, we’ve effectively shut down the load bearing critical path that was needed. It’s not over-engineered; it’s the right thing to do.”
huh. I wonder if it's possible to use those hooks to add syntax highlighting to shell commands claude issues, or to replace full path to current directory with ./
As someone who has been describing things as 'load-bearing' as something like a signature phrase for about twenty years, I'm beyond miffed that Claude has ruined my whole gimmick.
A new catchphrase every twenty years is hardly sustainable at my age :)
There are no real solutions, it has to be fixed during the training. ST folks have tried many non-working ways over the years, but two workarounds are more or less worth considering:
- Samplers that increase prose variance. They require running the model locally, they dumb it down, and never fix the actual issue, which is mode collapse leading to semantic collapse and rigid mapping of input to output concepts. The model still expresses the same ideas in different words.
- Let the model write anything if it couldn't resist, but check and fix it in the verification pass. This solves the semantic problem, but cannot solve the variance since the second pass is also subject to rigid mapping, i.e. you replace it with the same stuff over and over. The verification prompt can be randomized to a degree using pretty clever schemes to give it some variance, but of course this also fails in predictable ways.
Lately, I feel like as GEN AI text becomes the majority, human-written text is starting to resemble it too.
I'm Korean, and there are sites and people who mainly curate the latest technologies. Even those people, probably tired of translating every time, have started summarizing things with AI. But recently, I've noticed that even when people don't use AI, their writing is starting to look like GEN AItext.
I think the reason might be that people often base their thoughts on documents they've read, or paste parts of content when writing their own texts, which leads to that style.
I'm not sure. Whether human writing is better or AI writing is better—personally, AI writing tends to flow in a very even, paragraph-by-paragraph structure, which makes it good for consuming information. I wouldn't want to read a novel written that way, but for getting information, AI writing is surprisingly convenient.
This is why when I even sniff a hint of Claude phrasing I close the article and block the person on all devices. I'd literally rather be lobotomized than sound like this shit. The fact that anyone can read AI text without immediately starting to dry heave is a pretty damning indictment of their character
I actually think it's better. Back when access to knowledge was only available in English, there was a lot of mistranslated information in my language (Korean) that was worse than AI slop. These days, the translations are done by AI, so the tone may be awkward, but the content is more accurate than before, so I don't mind.
So that's the difference. I'm already living in a degraded environment, so this actually feels like an improvement to me. But you, coming from a better environment, perceive it as worse. It always seems to depend on cultural context.
How do you manage to make Opus follow any rules? Maybe it’s a windsurf thing but I have a ton of custom rules and Opus just ignores most of them. GPT on the other hand follows them like it’s a cult - if I have a rule I can’t ever force it to ignore it. Opus just doesn’t care. If I ask why it’s not following rules it will apologise and suggest creating a rule for it …
I recently started using caveman, and it’s been great. It doesn’t just cut down on overuse of specific terms; it cuts down on time spent digesting slop in general.
The token saving is oversold, from what I can tell so far. These days output tokens are just the tip of the iceberg.
If anything the real value is it saves my brain from going into power saving mode by lunchtime because I haven’t spent the day reading pages of output when a sentence or two would do.
Developers who can't stop themselves from using embellished and "posturing" phrasing for simple things are a pet peeve of mine. I feel like this "knack" of Claude in a way scratches these special people behind the ears in just the right way.
It's good, because it's just post-processing before display. So it doesn't interfere with the process, which those phrases that seem so offensive to sensibilities of so many people, for whatever reason, might be a part of.
"Honest assessment: I was wrong to say I was being straight with you. You pointed out that a "smoking gun" is a sign of evidence, and I clearly didn't have any. This is not a bug but a gap that can be fixed like [this]. Give me the word and I'll wire it in."
··You've hit your monthly spend limit · raise it at claude.ai/settings/usage
Annoying because I used to like using that phrase.
A similar Codex/GPT verbal tick is "deliberately narrow" or variants thereof.
Just a grep across my repo comes up with a dozen lines with phrases like "It is deliberately small" or "This crate is deliberately not a X" despite my efforts to police this kind of thing.