It's Not Just X. It's Y
mail.cyberneticforests.com
mail.cyberneticforests.com
This is honestly both terrifying and well articulated.
High praise to the blog author.
And it should. If someone sends me an obvious copy-paste mass email it also has an impact on how serious I take that email to be meant as an actionable proposal targeted at me specifically.
If they half-ass their proposal which has to do with language I can reasonably infer they may also half-ass the real world implementation which in this case also had to do with language. If you're unable to describe your own idea on half a page of paper in your own words, maybe the students who are able to do so should be treated fair.
I don't care wheter or not they use LLMs, in fact do. But engage with the ideas and results and convince me you really care about them.
Recently a contractor presented a small internal tool they made for a very specific task. The UI text gave a direct impression to be entirely LLM-generated with no human edition. It makes the entire tool look vibe-coded, affecting how we judge it. Now I question whether they even reviewed the functionality correctness, which I wouldn't doubt if the UI didn't look generated.
If other approaches didn’t improve people’s writing you think LLM-shaming… will?
I personally avoid the em dash, but it has been used in writing for a long, long time.
Define "normal language?"
Most of the common tells stem from that desire to signal the gap between what it's writing about now and how you previously thought things worked. And they really fall flat when it's writing about some milquetoast truism with all the edges sanded off.
Rather than have "Z" be self-evidently interesting, the LLM need to tell us that it's not "A". Except no one thought anything was "A" in the first place, and the "Z" is barely a "B" let alone a "Z".
Or things are "quietly X," implying that there is some secret knowledge that other people have.
Barring that, the LLM will signpost arguments, telling you how interesting things are. "This is the critical part..." before launching into another banal non-observation. Why have only "Section Header" when you can have "Section Header (another idea I had) - omg I just have so much to say about this". Maybe a list of weak ideas, presented in multiple ways. Bulleted with emojis. As a rule of three. As choppy sentences in a paragraph.
All of these follow the same pattern - trying to follow the FORM of having something interesting to say that differs from consensus understanding, without having the intelligence or boldness to actually midwife a new concept into the world.
I think another tell with a long lifespan is that it will try to write about a thing it experienced, but having no real knowledge of the qualia, it'll end up only referring to it (not uncommonly using "that") and leaving all the substance as an exercise for the reader
If the system prompt is “flesh out template y with thought x” the form drives the generation, it feels compelled to use the whole template.
Of the system prompt is “refine thought x and then format it with the appropriate parts of template y” it becomes a simple transformation and formatting.
A lot of current gen llm pain appears to be being in the early days of understanding the nuance of system prompts.
If you present the idea as "an idea" rather than "my idea", it's pretty good at challenging, solidifying, and refining it. It still leans on the same forms in the writing part, but I think it can be tuned to not insist upon the idea's brilliance so breathlessly.
What you get is vacuous, choppy, wordy, and hyperbolic. I have found that adding secondary passes for tone and style improve the readability dramatically.
> In the end, shaming people for writing that gets flagged as AI can lead people to sidestep structures the model has learned from us: structures that are effective tools for argumentation. We take the tools of critical thinking out of the kit at the time we most need them.
Now we’re going to stop using effective rhetorical methods because they imply AI, even if we know we’re not using AI?
It reminds me of, as a teenager, asking my dad if he ever saw Led Zeppelin live. He hadn’t, because he didn’t really like fans of Led Zeppelin and didn’t want to be associated with them.
As an ashamed fan of certain bands I get this instinct but I also promised to myself when I heard this that I would do my best to not allow other people to influence how I thought about things I enjoyed.
On the same note I’m trying to be “braver” about things like em-dashes, though my personally style has always been to use them as I did in this comment- like this, which I guess distinguishes me, until an LLM picks that up too…
—
You're not using that, neither in the past from what I can tell, nor in this comment.
You're just using a hyphen/minus instead of a colon, that's not an llm-ism
- = hyphen
-- = n-dash
--- = m-dash
The above isn't slop. It's shit though!
That's really unfortunate though. It's like Michael Bolton from Office Space: "No way! Why should I change? He's the one who sucks."
Obviously there are exceptions; you can use certain constructions in a way that's still unmistakably human, or use them within a larger context of unmistakably human writing. But in general it makes me think about Orwell's argument against cliches:
> A newly invented metaphor assists thought by evoking a visual image, while on the other hand a metaphor which is technically ‘dead’ (e. g. iron resolution) has in effect reverted to being an ordinary word and can generally be used without loss of vividness. But in between these two classes there is a huge dump of worn-out metaphors which have lost all evocative power and are merely used because they save people the trouble of inventing phrases for themselves.
If LLM-isms give readers the impression that I'm too lazy to phrase things in my own words, even if I did in fact phrase things in my own words, then I take that as a sign that I should pick better words!
Granted, I've had a strong desire to write as distinctly and un-cliche-ish-ly as possible since long before ChatGPT's public launch, so I might not be as grumbly as other commenters who feel like this would force them to change how they write.
Possibly, but it's not a necessity. Click baiting (i.e. yt videos) has evolved to stable standards, that's at least my impression.
Also, products often only get improved until they are "good enough", not until they are "good". It happens, but then they just iterate towards the "how bad can I become" baseline from the other side.
AI companies generally are not in the "let's make the best AI possible" business but in the "let's make the most money" business. This just hasn't fully manifest because they get flooded with VC.
And you can pry my em dashes from my cold, dead hands.
The "AI Detection" tools employed by schools also regularly flag writing from those with Autism, ADHD, and non-native English speakers as being AI generated as well.
So, naturally, I can't stand the phrase "write like AI" when these things tend to come up because no, there are no humans that "write like AI" it's the models that have stolen the literary devices from us and now have poisoned them.
> While the detectors were “near-perfect” in evaluating essays written by U.S.-born eighth-graders, they classified more than half of TOEFL essays (61.22%) written by non-native English students as AI-generated (TOEFL is an acronym for the Test of English as a Foreign Language).
> It gets worse. According to the study, all seven AI detectors unanimously identified 18 of the 91 TOEFL student essays (19%) as AI-generated and a remarkable 89 of the 91 TOEFL essays (97%) were flagged by at least one of the detectors.
For better or worse (and pretty much for worse), these usages have become AI idioms. Language evolves over time, things that used to be harmless become offensive, certain terms end up taking on the complete opposite meaning than their original meaning, and we are watching certain language patterns and idioms become watermarks for AI and while it sucks, it doesn't make it false.
"We create a culture of self-censorship and AI-detector-pressured rewriting and paraphrasing as people strive to avoid these witch hunts. That is the opposite of protecting human expression. We should resist normalizing a trust in any machine's ability to determine matters of guilt. If using AI to write is, at its worst, an industrialization of the mind, then AI detection, at its worst, becomes a surveillance system for thought."
And, I'm sorry (I'm not), but I am not going to just roll over and shrug and say "welp, guess we all need to dumb our writing down to keep well-meaning idiots from screeching 'AI! AI! AI! WHOOP! WHOOP WHOOP WHOOP!' at us." That isn't the evolution of language. It's Idiocracy.
Now I'll have to find something else to overuse: maybe sentences structures around colons, or use of Japanese 「hook brackets」.
It's not X, it's Y, though? Couldn't be me.
No, fuck that. I'm not going to think twice about what I write just to avoid an AI checker, and I will delve into em dashes with gusto if that's what the writing calls for.
I'm not sacrificing the language simply to sound less like AI--that's absolutely a losing game.
And if anyone thinks my hand-crafted prose is AI-generated, they're free to look elsewhere. Right now AI detectors peg my pre-AI work as 30% AI-generated, and I'm certain that number will only increase as LLMs improve.
It depends on your environment, I guess. If you're a student writing an essay or a researcher writing a paper, it's in your best interest to avoid sounding like an LLM, which means going out of your way to avoid certain idioms, even if it means letting go of things you liked to write.
I used to love a spaced en dash (the British English equivalent of the American unspaced em dash), but I wouldn't risk it now.
I agree, but it’s worth noting that that has been done since long before LLMs. Fifteen years ago, I used to teach a graduate course on academic writing pedagogy. The students and I would read research papers on the teaching of academic writing; we also analyzed textbooks and course syllabuses to get an idea about what was actually being done in classrooms. While phrases like “critical thinking” did come up, the overall focus was clearly on language patterns: sentence and paragraph structure, the use of transition words, vocabulary for hedging and boosting (i.e., making assertions seem weaker or stronger), etc.
In a university context, it can be very difficult to evaluate student writing based on its content. In humanities-focused and creative writing, what the student decides to say can be seen as an extension of the student’s personality, identity, and individual experience; if a teacher evaluates the content, including the reasoning, it can seem that the teacher is evaluating the student as a person. And if the students are in the sciences, especially at the graduate level, the writing teacher often won’t even understand what the students write because it is too technical. Teaching and evaluating language patterns, not content, is often the only option.
What authors informed the course? What were the course’s main text(s)?
This was all before LLMs. The assumptions behind my class have since been upended, now that easily available chatbots can produce academic writing that, at least at the structural level, is perfectly fine.
I've also never used an AI detector, and probably never will.
In my experience:
1. The people who rely the most on AI writing don't like to admit it. I catch obvious AI hallucinations in my boss's "documentation," and he always insists it was his own human oversight, despite it being very obviously a mistake I've caught Claude (and importantly, no human coworker) making repeatedly
2. I don't trust a machine more than myself to judge writing
3. Obvious AI "tells" just make it clear i don't need to keep reading, not that i need some kind of validation. In some sense, i guess that might save me time? But i still have to have read enough to know what it is...
In general, i think the author makes great points about how _LLM "thinking" is just the reproduction of the language of reasoning_, that is not necessarily a replacement for actual reasoning. It'll take a lot more than that for me to believe an AI is "thinking" and not just giving statistically reasonable answers (reasonable or actionable though they may be)
I think of this as a dash (abstract function of punctuation) but not an em dash (concrete typographical form). I see the double hyphen as one of the possible representations of the dash¹ -- and a perfectly well-established one, to be sure² -- along with the em dash and the en dash, either of which can also be used to do the same job, depending on where you are and who you're writing for (the US uses em dashes of varying format (spacing) for that, while much of the UK uses en dashes with spaces to represent the same punctuation abstract). Are you instead conceptualizing the double hyphen as a representation of the em dash glyph?
If that makes sense then I'll leave it at that, but if it doesn't quite, then I have a somewhat longer hair-split I can send if you care for it.
1. A pseudo-glyph, maybe
2. Still acknowledged by Garner's as recently as 2003
Same has been applying to art for a while. Several artists who have an "AI-ish" style have been wrongfully crucified for using AI. And been forced to post videos of their process end to end, in order to prove that they aren't using AI. It's a thing for artists to post their new stuff with: "AI could never do this."
He used PERL for scraping and the same for generating "new" content with what he had scraped. He made a bunch of static websites with ads. The sites connected to each other. Sports, Celebrities, Movies, you name it. He had a formula.
Those sites were profitable enough that he could travel the world and have fun, basically. His mother collected the cheques from Google and cashed them into his account.
Now to get to the final point here, his secret was simply TYPO INJECTION to avoid Google's then embryotic duplicate content detection.
I don't know if it was written that way to show trust in the reader's intelligence, show disregard for reaching a wide audience, show a demonstration of skill, or was artifact of someone just thinking at that level.
In the interest of not occupying significant page/screen height with LLM output, example prompts+responses here: https://dpaste.com/H9DXKNYQH.txt
> I don't know if it was written that way to show trust in the reader's intelligence, show disregard for reaching a wide audience, show a demonstration of skill, or was artifact of someone just thinking at that level.
This is merely the end-state of industrialization, which is efficient and soulless.
This feels like an easy enough hypothesis to verify, for anyone in the business of training LLMs - does the not-X-but-Y rate increase after RLVR?
Also, the Greeks were worried about rhetoric and, in my opinion, rightly so. The skill to argue a point well is different than those that are needed to be correct. To become a skilled rhetoritician was viewed as dangerous (and right now AIs are only moderately good... though they are improving fast).
Example 1. Someone reached out to me on LinkedIn "from HN", when I replied to their initial message they just said "You look like an AI trawler" and disconnected.
Example 2. I read a popular non-English tech blog aggregator which also encourages comments and discussion. Since ~1.5 years ago every other comment is "thank you author, but if I wanted to read AI I would as AI" or some variation thereof.
Would it be some kind of "reverse psychosis"?
I think that the current models are still like over-achieving savants rather than true human level because the largest model is only 1/10th the complexity of the human brain. I've recently become fairly convinced that new hardware paradigms (like types of CIM) are about to move from research into real-world development and scaling. So I believe within a few years, the model sizes will increase by another 10 times.
Compared to upcoming 100 trillion parameter models, humans will obviously be _much_ dumber/slower than AI in all fields. Already with the 10T models, some LLMs beat 99.9% of humans in competitive programming.
The AI hatred from many may actually continue to increase, but in cases where the bottom line matters, we are rapidly approaching the point where writing or work product that looks like it is human-authored will be suspect just on that basis. In other words, for some people it will be the reverse -- "this work looks like it was created by a human" could be devastating for your businesses credibility at that point.
It's interesting why LLMs generate constructions like this more frequently than they presumably exist in the training set. I wonder if this is some sort of mode collapse caused by post training, and/or maybe because they are training on synthetic data so these things become self-perpetuating and self-amplifying (a feedback loop)?
The lesson for humans worried about being falsely identified as AI is just learn to write better! It doesn't matter where your repertoire of phrasing comes from (copying AI or not), but one of the basic rules of writing is not to repeat yourself unless you are doing so deliberately for a purpose. Go ahead and use "It's not just X. It's Y" if you want to, but if you use it multiple times in the same short piece of writing, then you may deserve to be called out for poor style, if not for being an AI.
If LLMs generated text based on training data frequency they'd likely be some of the most vulgar and hostile things ever created. The internet is full of insults, profanity, and low effort content. The repeated phrases are a side effect of reward optimization rather than some kind of model collapse.
Turing Test
https://xkcd.com/329/Anyway, yeah, people trusting AI to do a better job in reasoning than fellow humans, without justification, worries me. We have no formal theory of informal reasoning (that LLMs mimic), so we cannot verify it any better than with humans.
You have to trust someone, to ground your beliefs. Trusting AI is just trusting some other people (who trained it) by proxy. Once you realize it, you might as well try to trust people you know.
How is this different from humans? When I went to high school, my teachers extorted me too. Especially subjects like English and unlike Math, where evaluation is 100% subjective.
Until 2006, the national Finnish/Swedish (as a native language) exam at the end of high school in Finland consisted of two essays. One was based on materials that were provided, and another on a topic that was given. I believe there were a few options to choose from. Both essays were scored independently, and your score was the better of the two. If you had learned to write essays, it was effectively an intelligence test and a good predictor of future academic success across many fields.
Including CS, as my department found out. In particular, the ability to write essays was a better predictor of success in CS studies than the scores in mathematics and natural sciences. Probably because there has always been a large subset of CS students who are otherwise good at CS but can't handle anything resembling mathematics.
And that's why everyone on the receiving end of the AI slop deluge is so paranoid.
Sometimes it’s not just about the Ys but also the Qs.
- "No X, No Y, No Z." pattern
- "Here is X - it makes Y"
The worst and most obvious one is the constant over use of emoji ticks and crosses.
*actually a hyphen but it's functioning as an em dash.
This reminds me of another em dash+AI related topic: I've noticed LLMs have an extreme bias towards spaces around the dash while people can go either way with it.
A human writer would add that sparingly, for cases where there might be some ambiguity. But the bot added that to basically every single API endpoint.
I put to you, if you see a trope in AI writing it's because that trope appeared in the training corpus. Therefore, sure, being predjudice against it lets you catch some AI, but you'll also flag human outout. I think that may not be worth it in the end.
There are many cases out there of people proving that it does happen. I have personal text not published that do trigger all detectors between 70-100% AI depending on the tool. I wont be sharing these, as "providers" of these "tools" would simply add it to training data and continue to merrily overfit.
Bottom line, transformers regress to the mean like many other models, if you as a person produce output aligned with mean of corpus, you'll trigger detection. More importantly, evading detectors is trivial. Find a corpus of text from an author, get an llm to write a note on style, parlance, habits in writing etc. and then use that voice file to drive outputs from an llm. If the source text didnt register as AI, the new ai output also reliably avoids detection.
So my problem is, detection doesnt work, false positives are fact, so these tools at best offer harm.
A bigger problem¸ undermining your request for proof beyond that which many before me, and including me, burned in an effort to make folks see reason, is that even if i handed you proof on a silver platter you wouldnt understand it. If you could, you'd already understand, because the problem and the math are quite simple. Every example text i've offered in the past now registers 100% human, but many I kept to myself continue to show the same problem, and that never changed. So why would I waste my few remaining tools for sanity checking, when all i can expect from that endevour is losing a tool and shifting nothing in the conversation? No, best I keep that to myself for now.
I used one to help me plan a sales route, and it kept fucking it up. Every time I corrected it, it tried that hand wringing vizier sort of ass kissing. It's very off-putting, but I can see how someone struggling with social interaction could be sucked into that nonsense.
I went a few rounds with it and it kept pushing back, which was odd for sure, as I'm also used to being able to essentially just say "no, x = 2". There was a lot happening in that session (it was really someone else's and he just kind of lets his context windows fill up with all sorts of stuff), so I bet it would have sufficed to start a new one, but after a point I just wanted to see what would happen. It didn't concede until I sent it a PDF of some kind of white paper or tech spec sheet or something that included release dates from an authoritative source.
I may need to find that session later, because now I'm wondering if it was entirely the PDF that did it, or if it also helped to point out that the info cutoff could be a factor, and I don't remember whether I tried the latter first.
Also not sure off the top of my head what the situation was as far as model, custom instructions, or whatever else.
It is bad writing.