Surely these leading tells will be trained out of models pretty soon, given how well known and overused they are. And it might make the writing slightly worse in a way. But it is quite annoying how often this type of construction is used in everything at the moment.
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.
I like that these AI idioms exist. They're like watermarks for text. It's worth the cost of humans avoiding them. Companies will eventually train their models to be undetectable, but society would be better if they didn't.
Except that the entire point of the article is that they're not AI idioms. They're not "watermarks for text." They're legitimate language constructions that LLMs tend to overuse, but that real humans also use. Real humans do, in fact, say "align with" all the time, just as often as "corresponds."
And you can pry my em dashes from my cold, dead hands.
The article is not God, just because it claims something doesn't mean we have to accept it.
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.
What's worse is neurodivergent writing, including my own, often resemble AI output. Now it feels like I'm having to alter my own voice in online discussions just to specifically avoid being accused of pasting an AI response.
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.
Well reading between the lines I don’t think they’re saying all of those uses are AI. They’re legitimate constructs, like the em-dash, en-dash, and hyphen, all of which I used to use regularly. But now they’re AI tells so I use them sparingly.
I agree with the feeling. But if you agree with the analysis of the article, this cat & mouse game ultimately amounts to stop disclosing our reasoning threads through commonly accepted linguistic structures. That's quite a price to pay as a society...
"So, if we publicly shame people whose text looks like it might have been written by a machine – because it mimics the language used for human reasoning – and people stop writing in ways that they internalize as "AI writing" out of fear of false detection, it sends a signal that your language for reasoning must be policed, or you too could be held up to public scrutiny."
This is honestly both terrifying and well articulated.
It’s unlikely this is true. LLMs are way more mad-libs / templates than we like to admit, that’s (ironically) not a judgement about their capability, it’s primarily just an observation. But it’s also what plain old SFT, which I believe is the primary culprit, ends up imparting.
This is how early forms of "reasoning" in LLMs worked: just literally inserting words like "Wait...", "Hmm...", "Let me reconsider...", "But is it really..." into the token stream.
Is this not how current forms of reasoning work? It seems like the open models still output things like that, and the closed ones all just summarize their thinking instead to avoid distillation, but probably do the same thing internally.
I think the basic idea is the same (not being a frontier lab researcher I couldn’t say for sure), but there are different techniques, such as “reasoning tokens” that aren’t literally words, and more interesting structures than just sticking them into the stream.
nice article, but i think as a non native english speaker, i always use the model in english for reasoning and then translate the output to my language. most of these considerations do not apply. because the translation step is taking out alot of these language artifacts
Do you manually translate or translate with an LLM? While reading, I was wondering how common these kinds of written tics are in languages outside English.
> In the end, shaming people for writing that gets flagged as AI can lead people to sidestep structures the model has learned from us
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.
Its not model collapse nor does it have anything to do with training data frequency. It's simply RLHF where the humans hired to tune the conversational style of these LLMs preferred certain idioms over others and so the reward function for these LLMs gravitated toward using them.
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.
> Because if Pangram's AI system found me guilty, that's the end of my career. That's literally extortion.
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.
"Hyphen functioning as an em dash" is an expected human thing as it's what's easy to type. It's specifically an actual em dash which got bulldozed, much to the dismay of those who bothered to put the unicode character in.
If you read The Mac is Not a Typewriter in 1992—thus burning Option-Shift-hyphen into your typing patterns for life, along with a dogmatic love for serif body fonts—you're the real victim here.
Or those of us that use a full featured editor when writing md!
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 signal is not the same thing as a guarantee. Both of your points so far, i.e. your provided text & that bots often bother to replace em dashes to avoid detection, actually support that it is a signal though.
The stronger yet signal is both combined! This glyph, that emoji, a given sentence structure, that formatting, a certain phrase. The more you notice -> the stronger the signal, the more you miss/discard -> the weaker the signal.
Alternatively, no one sounds like an llm, an llm sounds like someone, typically those close to the median of the training corpus. If AI were genuinly capable of novelty, it would be a big deal, tech bros having enough work ethic to design new detectable prose for an llm is a mssive reach and has no real evidence supporting it, else why do tech bros only tackle the easier issues? Things we have massive well labelled corpi for? Why is it never dishwashing and folding laundry?
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.
Problem is, everything gets poisoned by AI these days, and it gets worse when there's some sort of reward attached. Karma points in the case of Reddit and HN, in Wikipedia you got a ton of commercial actors and propaganda/distortion campaigns.
And that's why everyone on the receiving end of the AI slop deluge is so paranoid.
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.
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."
And you can pry my em dashes from my cold, dead hands.
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.
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.
This is honestly both terrifying and well articulated.
High praise to the blog author.
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?
It is bad writing.
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.
Sometimes it’s not just about the Ys but also the Qs.
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.
- "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.
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.
And that's why everyone on the receiving end of the AI slop deluge is so paranoid.