Art 18 ↔ Art 30 — Content promotes critical autonomy against AI patterns (Article 18) through adversarial 'cat and mouse' framing that potentially subordinates balanced protection of all rights including AI development rights (Article 30).

mvkel 2026-03-07 22:43 UTC link

Weirdly, LLMs seem to break with these instructions. They simply ignore them, almost as if the pretraining/RL weights are so heavy, no amount of system prompting can override it

carleverett 2026-03-07 22:56 UTC link

"The "It's not X -- it's Y" pattern, often with an em dash. The single most commonly identified AI writing tell. Man I f*cking hate it. AI uses this to create false profundity by framing everything as a surprising reframe. One in a piece can be effective; ten in a blog post is a genuine insult to the reader. Before LLMs, people simply did not write like this at scale."

This one hit home... the first time I ever saw Claude do it I really liked it. It's amazing how quickly it became the #1 most aggravating thing it does just through sheer overuse. And of course now it's rampant in writing everywhere.

CharlesW 2026-03-07 23:16 UTC link

A substantive human-written resource: https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing

bitwize 2026-03-07 23:26 UTC link

You know how no one ever wrote their own software and then generative AI came along and suddenly we could have app meals home-cooked by barefoot developers? (The use of such cottagecore terminology for a process that requires being an ongoing client of a hundred-gigabuck, planet-burning megacorporation rubs me in many wrong ways.)

If AI finally gets rid of the thing that drove me nuts for years: "leverage" as a verb mean roughly "to use"—when no human intervention seems to work, then I shall be over-the-moon happy. I once worked at a place where this particular word was lever—er, used all the damn time and I'd never encountered something so NPC-ish. I felt like I was on The Twilight Zone. I could've told you way back then that you sounded like a bot doing that, now people might actually believe me and thank god.

I will stick by the em dashes however. And I might just start using arrows too. Compose - > → right arrow. Not even difficult.

Jordan-117 2026-03-07 23:29 UTC link

Wikipedia also has an exhaustive guide, though it's not fun finding tropes you use yourself (I'm very guilty of the false range "from X to Y" thing):

https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing

Another one that seems impossible for LLMs to avoid: breaking article into a title and a subtitle, separated by a colon. Even if you explicitly tell it not to, it'll do it.

joshvm 2026-03-07 23:32 UTC link

No mention of Claude/ChatGPT's favourite new word genuine and friends? They also like using real and honest when giving advice. Far as I can tell this is a new-ish change.

> Honestly? We should address X first. It's a genuine issue and we've found a real bug here.

Honorable mention: "no <thing you told me not to do>". I guess this helps reassure adherence to the prompt? I see that one all the time in vibe coded PRs.

capnrefsmmat 2026-03-07 23:46 UTC link

I work on research studying LLM writing styles, so I am going to have to steal this. I've seen plenty of lists of LLM style features, but this is the first one I noticed that mentions "tapestry", which we found is GPT-4o's second-most-overused word (after "camaraderie", for some reason).[1] We used a set of grammatical features in our initial style comparisons (like present participles, which GPT-4o loved so much that they were a pretty accurate classifier on their own), but it shouldn't be too hard to pattern-match some of these other features and quantify them.

If anyone who works on LLMs is reading, a question: When we've tried base models (no instruction tuning/RLHF, just text completion), they show far fewer stylistic anomalies like this. So it's not that the training data is weird. It's something in instruction-tuning that's doing it. Do you ask the human raters to evaluate style? Is there a rubric? Why is the instruction tuning pushing such a noticeable style shift?

[1] https://www.pnas.org/doi/10.1073/pnas.2422455122, preprint at https://arxiv.org/abs/2410.16107. Working on extending this to more recent models and other grammatical features now

tiahura 2026-03-08 00:10 UTC link

Many of these are standard fare in legal writing.

Negative parallelism is a staple of briefs. "This case is not about free speech. It is about fraud." It does real work when you're contesting the other side's framing.

Tricolons and anaphora are used as persuasion techniques for closing arguments and appellate briefs.

Short punchy fragments help in persuasive briefs where judges are skimming. "The statute is unambiguous."

As with the em dash - let's not throw the baby out with the bath water.

awakeasleep 2026-03-08 01:15 UTC link

If this bugs you, open chatGPT personality settings, choose “efficient” base style, and turn off the enthusiasm and warmth sliders

It makes a tremendous difference. Almost everything on this list is the emotional fluff ChatGPT injects to simulate a personality.

newAccount2025 2026-03-08 03:51 UTC link

Great list. Invented Concept Labels is the one I think I get most frustrated by. When exploring new areas, I’ll read its paragraphs of acronyms and weird words and think I just don’t know some term of art, and as soon as I ask for a definition it’s like, “I just made that up, that’s not a formal term, blah blah blah.”

matusp 2026-03-08 06:25 UTC link

I tried using Gemini for some light historical research. It could not stop using tech metaphors. Lords were the CEOs of their time, pope was the most important influencer, vassal uprisings were job interviews, etc. The metaphors were almost comically useless and imprecise, and Gemini kept using them even when I explicitly asked it to not do that.

FiniteIntegral 2026-03-08 08:18 UTC link

A subtle tell for generated text is just how damn flat it is to read. Not that technical documentation require some form of grand prose, but how unspecific the text can truly get. Reading a high school persuasive essay can have more detail, and those are often just written for a grade.

I can understand someone needing help with writing but getting an agent to do the job for you feels like a personal defeat.

mosselman 2026-03-08 09:39 UTC link

I feel like the audience of the file is more for me the reader rather than the LLM.

> Add this file to your AI assistant's system prompt or context to help it avoid common AI writing patterns.

So if I put this into my LLM's conversation it is like I am instructing it to put this into its AI assistant's system prompt, so the AI assistant's AI assistant.

The alternative is to say:

"Here is a list of common AI tropes for you to avoid"

All tropes are described for me to understand what that AIs do wrong:

> Overuse of "quietly" and similar adverbs to convey subtle importance or understated power.

But this in fact instructs the assistant to start overusing the word 'quietly' rather than stop overusing it.

This is then counteracted a bit with the 'avoid the following...' but this means the file is full of contradictions.

Instead you'd need to say:

"Don't overuse 'quietly', use ... instead"

So while this is a great idea and list, I feel the execution is muddled by the explanation of what it is. I'd separate the presentation to us the user of assistants and the intended consumer, the actual assistants.

I've had claude rewrite it and put it in this gist:

https://gist.github.com/abuisman/05c766310cae4725914cd414639...

adrianh 2026-03-08 10:24 UTC link

This contains a lot of advice about good writing in general. Ironically I’d recommend it to humans as well as AIs.

Peritract 2026-03-08 21:03 UTC link

I find this repellent; why not, instead of trying to push unwelcome generated prose below the radar, stop trying to waste everyone's time? People don't object to these patterns because they hate lists of three; they object to them in this context because of what they signal about the content.

If using AI to write is nothing to be ashamed of, then you shouldn't feel the need to hide it. If it is something to be ashamed of, then you should stop doing it. If someone objects to you poisoning a well, the correct response is not to use a more subtle poison.

dwaltrip 2026-03-08 21:23 UTC link

AI writing sucks because it doesn't have a voice. It's not trying to say anything. Human writers are interesting because they offer a unique perspective from their lived experience.

It also struggles to maintain deep coherence. This is all probably related. It might be very hard or impossible to have deep coherence without human-like goals, memory, or sense of self.

firefoxd 2026-03-08 21:29 UTC link

If you are serious about sharing written ideas, I suggest you avoid using this type of prompts at all cost. I've worked with LLMs to write on my blog and they are pretty good at first glance [0]. But do it a few time and you'll notice that those tropes are the least of your problems. Not only all your articles will sound the same, but you'll see that same voice on other blogs, news articles, white paper, etc. It's as if they were all written by Mo Samuels. Readers are often here for the author's voice, not just the content of the text.

I often hear this here: "if you don't bother writing, why should I bother reading?" In fact, save us some time and just share the prompt.

[0]: https://idiallo.com/blog/why-we-hate-llm-articles

runako 2026-03-08 23:29 UTC link

These things are always so misguided, and this was no exception. The only way to have a piece of writing not flagged as AI is to write poorly. Ignore grammar, misspell words, etc. Don't follow basic guidance on composition. Generally write in such a way that you would merit no better than a C on a high school writing task.

I'll give some examples. Some will be from this list of "AI writing tropes" and some will be from prominent human-written (prior to 2020) sources. Guess which is which (answer at the bottom).

- "Let's explore this idea further."

- "workload creep"

- "Navigating the complex landscape of "

- "Let's delve into the details"

And I'm not going to get into how silly this is as a so-called LLM trope: "Every bullet point or list item starts with a bolded phrase or sentence." I remember reading paperbacks published before the first PC that used this style.

Fractal summaries is literally how composition is taught to students. Avoiding that style will make the writing more likely to sound less like a person wrote it.

I would suggest the author upgrade this to a modern version of Strunk & White and go on a mission to sell that. Call it Anti-Corpspeak or whatever. But don't pretend that these formulations only arrived in bulk in the last 2-3 years.

ANSWER KEY: these are all obviously prominent in text published before LLMs hit, as well as in the tropes doc. They are no more signs of LLM-generated text than is the practice of using nouns, verbs, and adjectives to convey ideas.

rdiddly 2026-03-09 15:09 UTC link

I have a suspicion that saddling a chat context with all this instruction would paradoxically produce worse results due to being overconstrained. But I haven't tested this. It's just that some of these are legitimate writing techniques that are simply overused. Is every single one of them always and automatically bullshit?

Also whoever claims "no human writes like this" hasn't been to LinkedIn... though the humanity of those writers might be debatable. But all the vapidity, all the pointless chatter to fill up time and space, it learned that from us.

I wouldn't have delegated this to an AI. Human for human, human for AI.

Kye 2026-03-09 15:44 UTC link

>> "Consider this list: https://tropes.fyi/tropes-md"

>> "How would you organize these LLM quirks, ontologically speaking? I have this notion that the better path is to identify what kinds of things are emerging and prompt to do those things better; accept it as something LLMs are going to do and treat it as something to improve on instead of something to eliminate."

The output is a bit better on blind prompting with applying the results. Here's the gist:

1. Compression artifacts — the model encoding structure implicitly

2. Attention-economy mimicry — the model trained on engagement-optimized writing

3. False epistemic confidence — the model performing knowledge it doesn't have

4. Affective prosthetics — the model simulating emotional register it can't inhabit

5. Mechanical coherence substitutes — the model managing the problem of continuity

Spot corrections are too spotty. Going higher levels with these kinds of problems seems to work better.

RandomWorker 2026-03-07 22:56 UTC link

It's a beauty. We can easily detect the issues with Youtubers that generate scripts from this tool. I've noticed these tropes, after 30 seconds, remove, block, and do not recommend any further. I hope to train the algorithm to detect AI scripts and stop recommending me those videos. It's honestly turned me off from YouTube so much, or I find myself going to my "subscribed" tab and going to content creators that still believe in the craft.

nh23423fefe 2026-03-07 23:19 UTC link

Weird to care about a harmless construction along with punctuation.

bitwize 2026-03-07 23:42 UTC link

If you sound like a car ad from Road & Track, I'm going to flag you as bot.

"No rough handling. No struggles to accelerate. Just pure performance. The new Toyota GT. It's not just a car—it's a revolution."

Most of the tropes listed on this page give text a more "car ad" (or sometimes "movie trailer") quality. I wonder if magazine scans and press releases unduly weighted the training set.

networked 2026-03-08 00:06 UTC link

You may be interested in my links on AI's writing style: https://dbohdan.com/ai-writing-style. I've just added your preprint and tropes.fyi. It has "hydrogen jukeboxes: on the crammed poetics of 'creative writing' LLMs" by nostalgebraist (https://www.tumblr.com/nostalgebraist/778041178124926976/hyd...), which features an example with "tapestry".

> Why is the instruction tuning pushing such a noticeable style shift?

Gwern Branwen has been covering this: https://gwern.net/doc/reinforcement-learning/preference-lear....

djoldman 2026-03-08 00:09 UTC link

The RLHF is what creates these anomalies. See delve from kenya and nigeria.

Interestingly, because perplexity is the optimization objective, the pretrained models should reflect the least surprising outputs of all.

pinum 2026-03-08 00:26 UTC link

Similarly, "X that actually works"

duskwuff 2026-03-08 01:03 UTC link

IIRC, it's well documented that negative instructions tend to be ineffective - possibly through some sort of LLM analogue to the "pink elephant paradox", or simply because the language models are unable to recognize clichés until they've already been generated.

malfist 2026-03-08 01:42 UTC link

Thats the thing about AI writing though. Those tropes are things humans do too. But like once or twice in an article. Not every single freaking paragraph

esperent 2026-03-08 02:16 UTC link

I assume it'll work more as a review pass rather than expecting good results outright. For all kinds of things like this where I feel like I'm fighting the LLM, doing the initial work then auditing it seems to be the best approach (the other one is writing all kinds of tests, LLMs including Opus 4.6 love to fudge tests just as much as they love telling you how insightful you are).

o_____________o 2026-03-08 02:52 UTC link

glenstein 2026-03-08 03:00 UTC link

There are some subreddits where this trope is completely out of control. For better or worse I follow the NBA subreddit and in the comment sections the number of people who throw in honestly as a qualifier is like way more than you would assume from natural conversation.

crabmusket 2026-03-08 03:30 UTC link

> (The use of such cottagecore terminology for a process that requires being an ongoing client of a hundred-gigabuck, planet-burning megacorporation rubs me in many wrong ways.)

I hadn't noticed this - great point. To be fair the "home cooked meal" metaphor comes from 2020, predating genAI coding[1]. But even then, CPUs themselves are so normalised that we just kind of... forget how vertiginously complex the entire supply chain is.

[1] https://www.robinsloan.com/notes/home-cooked-app/

wisemang 2026-03-08 03:59 UTC link

I’ve noticed the honestly thing for sure.

But I feel like I’ve noticed an uptick in people using the adverb “genuinely” in what I genuinely believe to not be AI generated comments, articles, etc. Maybe it’s just me, I got similar vibes about the word efficacy a few years ago, before the ascent of GenAI (but after the pandemic — again, maybe just me).

red_hare 2026-03-08 06:47 UTC link

I wonder if it has to do with how meaning is tied to the tokens. c+amara+derie (using the official gpt-5 tokenizer).

There's also just that weird thing where they're obsessed with emoji which I've always assumed is because they're the only logograms in english and therefore have a lot of weight per byte.

zahlman 2026-03-08 06:49 UTC link

I would say that the constant attempt to create false profundity (as you call it), itself, is more of a tell than any of the rhetorical constructs used to do it.

lucumo 2026-03-08 07:07 UTC link

I think that's Gemini trying to personalize the answer specifically for you. It really leans heavily into that to the point of being galling.

You can give it additional instructions in the settings, but you have to be careful with that too. I've put my tech stack and code preferences in there to get better code examples. A while later I asked it about binary executable formats and it started ending every answer with "but the JVM and v8 take care of that for you."

Which is both funny in an "I, Robot" kind of way, and irritating. So I told it to ignore my tech stack. I have a master's in CS and can handle a bit of technical detail.

Turns out, Gemini learned sarcasm. Every following answer in that thread got a paragraph that started with something like "But for your master brain, this means..."

pixelmelt 2026-03-08 07:35 UTC link

I found a new one in claude recently with "Fair enough, ..."

grey-area 2026-03-08 09:08 UTC link

They can work well when sparingly used and well thought-out, unfortunately LLM use is more on a par with:

‘It’s not mashed potato. Its potatoes lovingly mixed to perfection with butter and milk which quietly dominate the carrots beside them.’

The words are in the right order, th grammar is ok, but the subject is so banal as to undermine the melodramatic style chosen and they often insert several per paragraph.

grey-area 2026-03-08 09:10 UTC link

I wonder if th style shift has anything to do with training for conversation (ie. tuning models to respond well in a chat situation)?

orbital-decay 2026-03-08 10:21 UTC link

I have nothing to contribute but speculation based on my intuition, but IMO RLHF (or rather human preference modeling in general, including the post-training dataset formatting) is a relatively small factor in this, RL-induced mode collapse is much bigger one. Take a look at the original DeepSeek R1 Zero, the point of which was to train a model with very little human preference, because they've been on a budget and human preference doesn't scale. It's pretty unhinged in its writing, like the base model, but unlike the base model it converges onto stable writing patterns, and the output diversity is as non-existent as in models with carefully engineered "personalities" like Claude. Ask it to name a random city and look at the logits, and you'll still see a pretty narrow distribution. At the same time some models with RLHF (e.g. the old RedPajama) have more diverse outputs.

Collapsed mode makes the models truncate entire token trajectories, repeat themselves, and indirectly it does something MUCH deeper, they converge on almost 1:1 input-to-output concept mapping (instead of one-to-many, like in base models). Same lack of variety can be seen in diffusion models, GANs, VAEs and any other model regardless of the type and receiving human preference.

Moreover, these patterns are generational. Old ones get replaced with new ones, and the list in the OP is going to be obsolete in a year. This is what already happened to previous models several times, from what I can tell. Supposedly this is because they scrape the web polluted by previous gen models.

cainxinth 2026-03-08 11:35 UTC link

It amounts to telling it: “Stop doing that thing you can’t stop doing.”

nimonian 2026-03-08 12:36 UTC link

I completely agree. This is a good list, but a poor prompt.

Also, I sometimes find a sort of Streisand effect: when you tell the LLM to avoid something is starts doing it more. Like, if you say "don't use delve" it contains the words "use delve" which, amongst a larger context, seems to get picked up.

I have more success telling the LLM to write in the style of a particular author I like. It seems to activate different linguistic patterns and feel less generic.

Then, I make an "editor agent" comb through, looking for tropes and rewording them. Their sole focus is eliminating the tropes, which seems to work better.

Terretta 2026-03-08 15:56 UTC link

The source doc and the gist name dozens of specific bad patterns by label (“Negative Parallelism,” “Gerund Fragment Litany,” etc.) and repeat examples of them.

An LLM guide would do better to avoid every one of those labels and examples, since the whole point is not to prime the pattern.

Instead each instruction should describe the positive shape of good writing – what a well-constructed sentence, paragraph, or piece actually looks like.

Following this line, here is Claude rewriting OP:

https://gist.github.com/abuisman/05c766310cae4725914cd414639...

// This post’s typography and Oxford commas by human hands.

Terretta 2026-03-08 16:01 UTC link

Particularly to LinkedInfluencers.

Terretta 2026-03-08 16:24 UTC link

> now people might actually believe me

“Leverage” feeling “impacted” much?

TacticalCoder 2026-03-08 21:42 UTC link

> It could not stop using tech metaphors. Lords were the CEOs of their time, pope was the most important influencer, vassal uprisings were job interviews, etc.

That happens all the time if the previous discussion was about the other subject you don't want (tech in this case): LLMs (not just Gemini) go out of their way to reconcile the two topics.

As an example at some point I asked about the little shrooms people (the tiny people people do hallucinate all mostly the same when eating a particular mushrooms) to a LLM and forgot to begin a separate discussion and asked... About the root "-trinsic" in "intrinsic" and "extrinsic" and the city of "trinsic" in the Ultima game. Oh man... The LLM went wild. I totally forgot I asked about the little shrooms people hallucination but the LLM didn't forget and went totally nuts.

I think you'll get better result if you launch a new discussion and specify "Context: history" or "Context: cooking". Once it goes off the rail, asking it to "not do that" ain't really working: by that point it's just gone, solid gone.

Lerc 2026-03-08 21:44 UTC link

I have seen people suggest that the problem is that LLMs let you express any of your ideas, but the number of people with ideas worth expressing is limited.

In a sense I think this is accurate, but not inevitable. I think there is a lack of creative thinking, but it has come from a world that doesn't value it and suppresses difference.

There is a brilliant line in Treehouse of Horrors IV where Principle Skinner says "Now I've gotten word that a child is using his imagination, and I've come to put a stop to it." Which is just the perfect comment on the modern education system.

Models trained on the lack of diversity will push one way, but I think it will also avenues for expression that didnb't exist before. The balance will come from how we react and support what we would like to have happen

winwang 2026-03-08 21:47 UTC link

I don't think lived experience matters too much to me. In some sense, AI has very unique "lived" experience, which is what creates the voice it uses ("doesn't have a voice" seems like an impossibility to me by definition).

I find AI very "human-esque", and its "self-reported" phenomenology is very entertaining to me, at least.

I also think AI writing might feel trashy also because most human writing is trashy.

fudged71 2026-03-08 21:49 UTC link

My favorite:

"And honestly? That's rare"

mobrienv 2026-03-08 22:31 UTC link

Treating the act of refining text as a confession of shame misses the point of how writing works. Whether a draft begins as a model output, a dictation, or a scribbled note, the final responsibility belongs to the person who hits publish.

Improving prose to remove predictable patterns is the work of an editor. This process ensures the content is worth reading and respects the audience's time.

Comparing a software tool to "poisoning a well" turns a debate over style into a moral crisis that doesn’t fit the situation. If the information is accurate and the writing is clear, the water in the well is fine, regardless of the pump used to get it there. If the water tastes good, complaining about the plumbing is just a distraction.

Editorial Channel

What the content says

Structural Channel

What the site does