alt Hacker NewsVarious LLM Smells
Comments
- “(The) honest caveat:” (or “genuine caveat:”, both with the colon)
- “(The) honest answer:” (again, with colon)
- “The thing to internalize:”
- “The smoking gun:”
(really, sentences that start with “The <tag suggesting the next clause is the key point>:” are a strong tell, but those four are the most prolific)
- “load bearing” (when not talking about architecture)
- “blast radius” (when not talking about actual explosives, but rather the effect of an event/action)
- “smoke test” (esp. when “sanity check” is more apropos)
- Lists of three clauses/adjectives where the third is really just a combination of the first two
- Referring to the “shape” of things figuratively
- Social media posts that end with “Curious if anyone…”
- Stories or anecdotes using. “Oh. Oh.” (where the second “oh” is italicized)
Edit: Yes, some of those last ones are terms that we often use as devs...but I would argue about the actual frequency of their use. Plus, these tells live on in prose generated by the latest models.
The LLM writing sameness is bad. Use LLMs to help your writing! But don't include a word they generate, even just a vocabulary adjustment, in your own output. Have them critique structure and flow, spot overused words and passive constructions and dumb picks for topic sentences. It's great for that, and those are all objective improvements in your writing that won't mess up your style.
The LLM sameness in web design is good. Most sites shouldn't try to be idiosyncratic. The best design for a site with real utility is legibility, and LLMs are better at that than the median developer. Always laying out the same buttons? Always using the same type scales? Good! If it looks good to you, you weren't going to do better on your own, and you were very likely to do worse.
Don't forget about Contrastive Negation:
> Contrastive negation is a rhetorical structure that denies a specific idea in the first half of a sentence and asserts an alternative in the second half.
> It typically follows an "It’s not X, it’s Y" or "not just X, but Y" formula.
Wikipedia also has a great resource which covers many of the common LLM patterns: https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
No ___, no ____. Just _____
The interesting thing for me is that I do not feel like the writing of LLMs has improved very much lately stylistically. They have reached a "good" level some time ago but the newer models havn't brought such improvements that you would prefer them to an expert human writer.
Will be interesting if that holds in other areas when chasing super intelligence.
At this point, I want somebody's raw(ish) writing, with spelling errors and grammar mistakes and whatever, at least when it comes to most writing: blog posts, Slack messages, etc. LLMs are great for helping generate ideas, writing code, and maybe even cleaning up some writing, but doing the writing overall? Please don't. I want to hear what you have to say, not what the AI says, if it's something along those lines.
If you have Claude at work and are willing to point it to your emails, ask it to “read all my sent emails and create a skill to draft an email in my voice.” Even if you don’t want to use the skill, it’s fun to read the skill file it creates. It’s a bizarre feeling, asking Claude “who am I?”
I haven’t tried it with Slack messages because I’m a little scared to read what it says, haha. But the same concept surely applies.
There are a few people at work who are aggressively using Claude to write Slack messages. It’s easy for me to tell because one day they’re writing barely coherent English in multiple messages, and the next they’re sending perfectly coherent prose in a single message.
Scrolling down a LinkedIn feed is hilarious at the moment.
My favourite one today from today:
“The tax isn't the problem. The mindset is.”
I've been letting typos and punctuation errors slide in my communication so that its clear that its a human that drafted the message. Almost typing faster to encourage them.
> The "JetBrains Mono" font
Thought for sure we'd get a critique of Inter overuse. JetBrains Mono is a lovely font, though.
> Humans trust symmetry because it feels like intelligence made visible I hope this one didn't make it into the author's math blog. I don't understand what this means. I don't "trust" symmetry, I never think of symmetry as having to do with "intelligence". How did it even come up with that.
The LLM doesn't smell like authentic writing but it does a great job for fast and cheap words. We've gained something similar to fast food. Words made very cheap, very fast, easily digestible, but they have no emotion. In short stints it does have a place in the world.
I've found double and triple negations ("Not this. Not that. Not that either.") instead of just removing stuff to be one of the clearest tells.
<somebody> just <did something>.
And it changes everything.
You are right to push back.
> Late last year I started writing a math blog and decided to use LLMs to polish/enhance my writing.
Why would you do that in the first place? If you are just starting writing about math, I assume your goal would be to get better at it, and using LLMs is not how you get better at writing.
I personally wonder if the writing (and coding) smells are being leveraged for watermarking/steganography. Like, if it's quirky but intentional; they're biding the time, leaving breadcrumbs so they can more effectively sift new internet content, or perhaps tracing/de-anonymizing.
I can also see how it is not necessarily the case, since all models seem plagued by having unique cliche patterns. We all seem to experience the writing smells from the same models. I see 'cleanly' a lot less from GPT 5.5 vs <5.4
> :black_circle_for_record: Smoking gun.
> "belt and suspenders"
It's kind of interesting how genuinely hard it is to get models to deviate from basically all of these tropes. You can straight up tell it "I hate that card design, do something different, get creative!" and it'll do something either (a) ugly as sin (clearly just essentially a random walk through parameters) or (b) some same-y derivation of that card.
In coding, I've noticed a few tropes as well: everything is a "contract" or an "artifact" (clearly trained on like three decades of Java lol), everything is constantly "backwards-compatible" or "versioned" (even if working on a brand new greenfield project), and a few others.
It's funny, I've noticed my own writing has been getting a lot sloppier the past couple years just so that my writing doesn't get mistaken for AI. I actually used to use "--" pretty much all the time when I was young (granted I didn't go full em-dash, just the ascii version). It does make me kind of sad though, because writing in really short sentences (as one example) is actually a great way to make your writing easier to digest.
these stochastic models just replicate what they see in training data. the reddit speak, american english stuff (some through osmosis in pop culture), etc. makes sense when stepping back a bit. but some of these do happen to be from the people based on whom we train everything, and it is interesting when they get in the crossfire from the content perfumer.
also, how come jetbrain mono has become one of these tells is insane to me!
To be honest, I often feel that AI writes better than I do. Since AI drafts drafts quickly, my work speed has increased by about 80% compared to before.
I wonder if the tendency to write short punchy sentences stems from deliberate RL efforts to avoid repetitive, consistent writing? I seem to remember that a critique of early LLMs was that they would produce sentences whose construction was too homogeneous. Would be interesting to know the answer to this.
One I've noticed, and it might just be Claude, is putting "The" in front of titles and headers. Its documentation will have one section after another with headers like "The Architecture", "The Caveats", "The Fixes".
For designs, try https://impeccable.style/slop/ to detect AI slop patterns and improve your designs. I found it helpful for critiquing designs, developing a DESIGN.md guide, and iterating on UIs.
I don't think I've met anyone who uses the word "genuinely" as much as Claude does.
All of those are included in the bulk of the documents passing my work input these days. It is infuriating. Out of principle I maintain 100% me in all my writing but I don't know if it matters. Well maybe it does... an interviewee recently complimented me on the "nicest and most human resume" they saw recently. That felt good
Abusing the words "canonical" and "normalized".
New York is the Ljubljana of America
Does anyone else use the 'smells' as a sort of 'game' to ensure you properly go through a given LLM output (of any kind, be it document, presentation, code, etc) and 'make it your own' by eliminating them? I have a high bar for sharing content so I always do a rigorous pass to eliminate the em dashes, the contrastive negations, the 'quietly', and any other extraneous verbosity, and find that it helps me just really thoroughly polish it up.
Built DashBuster to bust --- :) https://chromewebstore.google.com/detail/dashbuster/pnfhimkh...
That's quietly the most important part.
Leave my consecutive short sentences alone. And jetbrains mono.
The color one is inaccurate since that's the tailwind color slate.
I wrote a skill for this! But I’m sure a lot of you have as well lol.
For those curious https://github.com/ryanthedev/oberskills/blob/main/commands/...
I think we should stop discriminating or the pursuit of detecting LLM-generated content. This is not helping to stop the slop, it causing a harm - making the genuine human-generated content being branded as something bad. It's like branding words like "segregation, black/white" as racist. It is hijacking the common a language patterns and make them unusable in common human expression.
Claude is currently and suddenly obsessed with describing anything a little bit involved as being “inside baseball”.
Last month it was quiet things being said out loud.
When you use the tools all the time it’s fun to watch these things pop up a week or so later as normies copy and paste the slop out on LinkedIn and their newsletters.
Thank you, these are all things I've noticed too.
Those cards, so familiar! Exactly what Opus produced for me.
Did Anthropic and/or OpenAI deliberately train their models to produce websites with a specific design language, or did these stylistic preferences emerge naturally as some kind of LLM-selected optimum?
SmeLLMs
Welcome to the future of fast-food software. Taste of deep frying and preservatives.
Quietly. Clean. Honest. Sharp take.
Moar! I want all the smells!
/endredditmode
I actually love folks documenting this. I'm all for LLMs producing rough drafts. But rough drafts are, as a rule, slop.
Can we have a Wikipedia entry for these? per model? I wouldn't mind using this page for that
That's an interesting difference between true intelligence and its LLM imitation. If you throw a pile of mediocre data at LLM, its level will drop to mediocrity, for it can't discern true from false, good from bad. However if you throw the same data at an expert human, he will ignore it, and maybe pick a few bits of novel knowledge, so overall his level will raise a bit.
"A is not B instead A is blah blah" instead of just saying "A" is a very common pattern have seen in Claude.
It is strange to read as the topic A has often not been introduced and introducing it by saying what it is not makes very little sense to a new reader.
One Python one I hate is that it adds crazy amounts of newlines for no real readability gains.
Instead of this:
def add_three_ints(x: int, y: int, z: int) -> int:
return x + y + z
def add_three_ints(
x: int,
y: int,
x: int,
):
return x + y + z
So the year is 2026 and we cannot point a LLM at, say, this HN thread, and give it the instructions: "I don't want to look like a dumbass, so don't make these obvious mistakes / don't use these obvious tells"!?
KPI cards, purple gradients
> The LLM generated writing obviously felt significantly better than my own writing.
A general pattern for LLMs is that they look really good at things you are bad at. What that means is that if you find yourself thinking of its output as significantly better than yours in a particular domain, there's a high chance that you are not equipped to judge that quality effectively.