Nowadays we use LLMs mostly for doing agentic-based work. LLMs new Pareto frontier only make the headlines if they push the boundaries on benchmarks that are deterministic tasks. So models are encouraged to focus on these deterministic tasks that are, in nature, structured texts. I think that this makes models more “plastic” or “polished”, as opposed to natural and pleasant to read. User-based benchmarks, like LLM Arena, are for me the best we can do in order to rank models in this way, but come with its own drawback (subjective evaluation, prone to spam or techniques to promote a giving model).
My revolt is against the cognitive stress of reading generated text. A trope typically indicates I’m in for an uphill read.
I recently read this William Zinsser quote that inspired a nickname for this: Clotted Claude [1].
> Nobody has made the point better than George Orwell in his translation into modern bureaucratic fuzz of this famous verse from Ecclesiastes:
> > I returned and saw under the sun, that the race is not to the swift, nor the battle to the strong, neither yet bread to the wise, nor yet riches to men of understanding, nor yet favor to men of skill; but time and chance happeneth to them all.
> Orwell's version goes:
> > Objective consideration of contemporary phenomena compels the conclusion that success or failure in competitive activities exhibits no tendency to be commensurate with innate capacity, but that a considerable element of the unpredictable must invariably be taken into account.
> First notice how the two passages look. The first one at the top invites us to read it. The words are short and have air around them; they convey the rhythms of human speech. The second one is clotted with long words. It tells us instantly that a ponderous mind is at work. We don't want to go anywhere with a mind that expresses itself in such suffocating language. We don't even start to read.
> The Ecclasiast looked under the sun, but there was something he didn't understand. Something that wasn't right. Something that was not as it was supposed to be. And here is what the Ecclesiast didn't understand. Here is what nobody understood. Not then. Not in the years that followed. Not now. It was not the swift who won the race. Not the strong who won the battle. Not the wise who earned the bread. Not the men of understanding who gained the riches. Not the men of skill who gained the favor. And here is what I found: to any story of success, there is an element of unpredictability and chance.
(There are really just two possible outcomes: either the article is right or in, say, two years, we will be all writing and talking like this, as in humans learning from mediamatically reinforced human feedback.)
I am a bit of a luddite in this domain and have so far managed to resist the lure of using the generator to expand my thoughts, and I still catch myself writing "it's not just an X it's a Y" and other generator type tells. If it infecting my patterns it is totally entering the wider subconscious as "How to write" (Sighs)
Despite its bizarre look, the sentence is evocative and eloquent. It does make me get a clear mental image from the very first word. It leaves little room for roaming and guessing, as it firmly nails elements one by one, and, by the time I reach the end of the sentence, I get the full meaning almost immediately.
This sentence is not randomly written; this is crafted with intention. TBH, it would take me hours, if not days, to write a sentence this much condensed and easy to understand. I seriously like it.
Perhaps, this is more about context -- which style to use in which situation. I'm only guessing here, but, since Orwell is offering an interpretation, he probably chose to be more clinical. He probably had a point to make and didn't want to risk vagueness up-front.
Well, I can tell what sort of angle you most enjoy. Anyways -- I think there is GOOD writing and BAD writing, but only subjectively. So if you enjoy it, power to you. It's certainly not random, but it is the sort of verbosity that turns off 99 percent of the people that would read it given a comparison. I find the former rather eloquent.
When I read the Orwell's version, I immediately had the same feeling I had when reading mathematical proofs. I hate so much to hunt the preceding text for anaphora resolution... It's just such a bad and pretentious way of writing. It's hostile to the reader with the side of flaunting author's superiority.
It's like having to sit through a party with acclaimed academics: every single one is so full of themselves, they will constantly one-up each other by belittling everyone in their workplace s.a. to make you feel how great of an intellect they possess and how much more they would accomplish, had they not been surrounded by all these bumbling idiots.
My take is that the generated stuff is terrible for communications.
It feels great to use, direct your machine minion to fill out your thoughts for you, but holy hell does it suck to be on the receiving end. Least of all is the disrespect, they don't care enough to even talk to you but worse is having to try and reason through that big incoherent blob.
Probably to only reasonable thing to do is to try and get your own mechanical agents to produce summaries. Inventing the lossy expansion algorithm(like compression but things get bigger on the wire), And we wept.
Now I am all depressed because it is probably inevitable, apparently thinking is hard and in general people are all to happy to outsource it to the machines.
The problem is not the LLM prose appearing everywhere, it's the legions of AI-boosters appearing in every thread attacking anyone who complains.
Apparently, even though they want to spew AI prose everywhere, they want it read by humans, not by other bots, so when a few holdout places are insisting that prose be human authored they fight very hard against the rule.
I don't think I've ever seen someone on HN say that. Many people would say that AI makes them more productive at coding, not that the output is nice to read.
This thread, in particular, stands out - reader makes the claim that Pangram found that the US constitution was 100% AI generated, when others tried they found 0% (or close to it) https://news.ycombinator.com/item?id=48378191
Those are not that. The first two are people complaining about other people incontinently identifying text as AI, because it's annoying to listen to unreliable hunches and aspersions. The second two are complaining about AI detectors not being very reliable. The claim made in the last one is a casual anecdote about "an AI detector", presumably told because it's amusing. It isn't a vehement statement about how you must accept slop into your life.
I would say that AI prose is often still a bit iffy, but I would disagree with anyone who would want to argue that this is any indication that AI prose will always be bad in the future.
i am a big fan of llms and the possibilities they enable. but i also find this type of behavior extremely rude! ai;dr for life. :is-your-human-around: is my preferred emoji for reacting to such behavior
>iv. Never use the passive where you can use the active.
Orwell himself routinely ignores it, even in the first sentence of the essay:
>Most people who bother with the matter at all would admit that the English language is in a bad way, but it is generally assumed that we cannot by conscious action do anything about it.
The second clause could be rewritten in active voice by changing it to "but people generally assume". But this would make the writing worse, and Orwell, as a good writer, probably didn't even consider the option of making it worse, and therefore didn't notice the passive voice.
Passive voice is an essential tool for all good writers of English. I always give the example of the opening of Pride and Prejudice [0]:
>It is a truth universally acknowledged, that a single man in possession of a good fortune, must be in want of a wife.
The joke doesn't work in active voice. If you attribute this acknowledgement to some specific group of people then it's simply false, not a comedic exaggeration.
Technically, I think, that first sentence, by Austen, uses a passive participle but does not use a passive voice for any finite verb. I don't think that advice to avoid "the passive" is intended to apply to that situation.
For example, nobody would seriously suggest avoiding the passive participle in a sentence like "Put the broken plate in the bin". ("Put the plate that someone broke into the bin"?)
> It's a bit of a pet peeve when people include quotes on a blog post without linking or otherwise references their source.
You’re peeved with good reason. It’s the blog equivalent of posting a screenshot of an article to social media. People, please post your sources! In the age of misinformation, that’s more important than ever.
A bigger question I'm interested in is why do LLMs speak like that in the first place? Is that really what you get if you took the average of the English language? It would be difficult for me to believe that.
Is there something about tuning for desirable qualities that forces LLMs to have this voice?
LLMs get a lot of finetuning, but I suspect there are two things that can cause this kind of writing:
Firstly, some parts of the RLHF involve human graders on the LLM's performance. I suspect their general bias towards a punchy, persuasive writing style could come from what biases the graders towards preferring that response, especially in shorter segments and when the grader is not focused on writing style
Secondly, later parts of the finetuning involve reinforcement learning on achieving certain tasks which are automatically graded: stuff like coding tasks. I think this can create a kind of feedback loop where the style drifts further, and you get the kind of LLM tics which are even more extreme (it might be that they incidentally help somehow with the actual tasks, or it might be a drift that comes from the grader also now being an LLM or some of this finetuning happening on output from other models). The more recent claude models seem to suffer from this a lot, moreso than earlier ones.
It appears LLMs are much better at writing during a debate than when explicitly asked to write. The moment LLMs are tasked with composing a blog article or intro for a book they introduce all the nuances that identify the output as AI slop.
The first is of course from the King James Bible which for centuries was essentially a standard that all English speaking peoples aspired to. If you find that version difficult I would expect much literary writing before the 1940s also seems difficult. This is just to say I recommend reading the King James even if you are an atheist, as I am.
I also have to say that the first strikes me as being written by someone that might be smarter than I am, the second as being written by someone significantly less intelligent than I, yet somehow placed by society in a position of authority over me.
In the context of newly written work in the modern era, I would argue it's best to use grammatical constructions that are used in modern 20th/21st century English, at least most of the time. Those who wrote the KJV were trying to be expressive but the whole point was to do so in language that ordinary people would be familiar with.
> the first strikes me as being written by someone that might be smarter than I am
This is why Joseph Smith tried to imitate the language of the King James Bible in the Book of Mormon, albeit not very successfully.
Yeah, I found Orwell's easier to understand and quite fast to read as well, but I think it's just because of the style of writing of the first one. It's from an earlier style of prose that I'm just not used to.
I also don't have a problem with large words as long as I'm well familiar with the words. The length of a word has nothing to do with the complexity of its meaning. We just have a limit to the number of pronounceable combinations of 5 letters.
The second one was very clear and to the point. Parsing it was rewarded with instant understanding and I enjoyed the word choice. The first one was just annoying; I could tell it was just listing a bunch of pointless analogies to try to make its point sound more grandiose so I immediately started skimming, and didn't come away feeling like it meant much other than "we all die in the end". The second one made an actual point and was the one that made me want to read it. The first one was the chore for me.
I wonder if this is due to experience with reading technical documentation?
I think the second requires deeper concentration, but is still quite readable compared to the kind of low-content engagement / SEO stuff one read on the internet even before LLMs
The grammar is more straightforward. It has some extraneous words, and makes conspicuously bad choices of vocabulary, but it's still a more direct statement.
Other translations keep Zinsser's preferred lack of fuzz but avoid using "is ... to" for possession.
For example, the Lexham English Bible:
> I looked again and saw under the sun that the race does not belong to the swift, the battle does not belong to the mighty, food does not belong to the wise, wealth does not belong to the intelligent, and success does not belong to the skillful, for time and chance befalls all of them.
This could be shortened to "success does not belong to the skillful, for time and chance befalls all" with no meaning lost. It's self-indulgent fluff. Meanwhile, Orwell's actually adds more to the statement - much better signal to noise.
I would love to use Pangram but they simply don’t allow signing up with my custom email domain. The error was “This email address can't be used for signup. Please use a different email.” I’m not about to create a Gmail is to use your service. To me the attack on the decentralized nature on Internet infrastructure is no less serious than the attack on the human provenance of writing itself.
Disclaimer: I do not like to read LLM-generated text any more than anyone else.
IMHO a big problem with Pangram in particular is that they market it as a reliable tool that can be used to catch students cheating. This can obviously have disastrous effects on young lives, because it is not as reliable as they suggest.
Per their own benchmarks, they do not achieve 100% accuracy even on text that is published on the Internet, and which is likely encoded into the models themselves.
There is validity to their goals, but that is overshadowed by the irresponsible way in which it is marketed.
(All of this, swirling in a context where students are being told that they absolutely must become proficient at using LLMs to do exactly this kind of work by the highest levels of state and federal governments, faculty leadership, as well as the leaders of the workforce into which they hope to graduate. The message to youth is extremely muddled at best.)
Agreed, I like the Oxide podcast as well, even though it’s hardware I’m unlikely to ever see never mind use, it’s nice that someone somewhere is still trying to be what they are trying to be.
Having dealt with Enterprise Hardware(TM) in a previous job, it’s refreshing simply to see someone look at that pile of crap and go “it doesn’t have to be that way” and then actually set out to prove it.
They have a ton of rfds publicly available that are worth reading. Recently, read this one: https://rfd.shared.oxide.computer/rfd/0161 because I'm researching clickhouse for my work (there is also a podcast ep on it). Even if you never use their hardware, just reading their work around the software they use/make is incredibly valuable as an engineer.
I have (an API, not an extension), but pangram is way to expensive for me to run, current (v4) pricing is $0.05 per 100 words. Should be absolutely doable if we split the cost between users tho.
I tried searching for good opensource / reasonably priced alternatives, pangram themselves even have some of their older architecture and training data on github/hugging face, but i never got that working reliably enough.
That is honestly the highest possible praise -- thank you. And when this piece was starting to boil inside of me last night (triggered, I'm sorry to report, by an obviously LLM-authored guest blog entry from the Rust Foundation[0]), I messaged one of my colleagues: "Time to do what I do best: bluntly say what lots of people are thinking."
I would go even further, I want a browser extension that scans all words on every page and colours them more and more transparent as the likelihood of llm prose is increased.
Years ago I used a rudimentary (text matching) Greasemonkey script that hid Reddit posts and comments from accounts matching a few behavioral/history signals.
I wonder if that idea could be modernized now for this.
Are you footing the bill yourself, or would you be supporting BYOK? Not sure whether logging in with user's account would also be a good workaround or not.
Unfortunately, it's not a browser extension and doesn't seem to have an API. I'd make a browser extension for this myself if it didn't involve paying for expensive Pangram usage.
For me it’s AI videos or music / narration that is beyond off putting. What’s worse now it seems people are writing their YouTube scripts with Claude et al. so at times even if it is a human creator you can clearly and immediately tell the words are not their own. To those creators I have but one message: IT SUCKS. I’d rather have you ramble incoherently in your mic then reading an LLM script and I will remove you from my feed immediately. I concur with the author on all accounts. We all can tell the BS people are selling us, unoriginal ideas, shallow concepts, open ended questions that hint at exactly nothing. Don’t be an LLM echo
This. And the structure as well, like the repetition of the same points over and over.
I'm not entirely against using AI to help content creators improve their narrative, like finding common storytelling mistakes. But that's very different than using yourself as merely an avatar for LLM content.
Great piece and interesting data. The rate of LLM-based writing rejection among developers is even higher than I thought it would be.
To me, the glaring question is: What are we doing? The act of writing exists to 1) externalize and organize one's own thoughts for the purpose of considering and revising those thoughts; and 2) share one's own thoughts with other minds.
When we hand writing to a machine, we hand thinking to a machine, denying both our humanity and our role in the conversation.
No. I have good anecdata: readers cannot reliably distinguish my own prose from LLM-written one apart from cases where LLMs use odd metaphors or one of their specific patterns. I've been specifically experimenting with that.
What kind of prompting are you using to get those results? Anything I have claude or codex write carries a ton of distinctive characteristics. Obsession with "bit-for-bit identical", "it's not the X it's the Y Z" and so on.
It's driving me nuts, I constantly have to prompt it to "explain in plain, simple English"
Schwitzgebel, Strasser, and Crosby fine-tuned GPT-3 on Dennett's corpus and asked whether readers could pick Dennett's real answers to ten philosophical questions from four machine-generated alternatives, with no cherry-picking beyond mechanical length filters. Even Dennett experts averaged only 5.1 out of 10 (well below the 80% the authors predicted), blog readers got 4.8, and lay participants barely beat chance — though experts did rate Dennett's answers as more Dennett-like overall. Schwitzgebel stresses this isn't a Turing test (one-shot text is far easier to fake than extended interaction), but argues it foreshadows a future where machine outputs are humanlike enough that their moral status becomes genuinely uncertain, motivating his "Design Policy of the Excluded Middle": build machines that clearly lack moral status or clearly have it, not ambiguous ones in between.
My own take is : don't focus on the symbols on paper. focus on the facts about the world it is talking about. Isn't objectivity all about the facts? In future AI will have all the memory about what I have already read and it will just furnish the delta new information in the blog/writing so that I don't spend time on refreshing what I already know.
I see this at work. People are "writing" specs and design proposals with bots. This is noticeable and is a huge turn off. I don't have issues with using bots to aid research, but I'm not reading the doc you slopped together.
I work with a guy that I swear is addicted to LLMs. He uses them for literally all communication, often dropping mountains of text for design specs that could have been written with half the words. Even on a 1:1 Zoom call, he'll type things into Claude and then read me the response! It's infuriating, and I've told him on a number of occasions, in as many polite ways as I can, that I would prefer to speak and work with him instead of Claude, but he just can't break the addiction.
How can you tell if people can accurately identify AI generated text?
If a person reads AI generated text and does not notice, they by definition will not know about it.
There have been numerous cases of people accessing human created content as being AI.
There are instances where it seems relatively uncontroversial that it is AI generated, but without knowing both the amount of AI content people are exposed toand the amount that they register I don't think you can draw a conclusion of the overall state.
I don't think this is about edge cases where someone has successfully disguised the writing to some degree: the current crop of LLMs have some pretty blatant (and frankly annoying) habits by default, ones that are hard to miss once you have read a decent amount of their output. If I had to describe them broadly, I would say they are a collection of habits which are common in certain kinds of persuasive and emotive writing, but are usually applied way out of proportion to the topic at hand, which tends to make the result quite grandiose, overly dramatic, and tiring to read: a LLM will often write a TODO app README like it's a cross between a thriller novel, a political speech, and a bombshell news article. There's lots of specific tics (and just by sheer volume and uniformity almost any habit an LLM picks up is going to rapidly shoot into cliche regardless of its own merit) but this is the general effect which I think is objectionable independent of the source of the text.
I do think the sensitivity to it can vary a lot: it depends a lot on how much and how closely you read the text, and how much exposure you have to LLM writing. Certainly it seems like a lot of people just don't really notice, or at least don't care much.
> the current crop of LLMs have some pretty blatant
This is just "em dash redux." Except now we've moved on to accusing anyone who does "It's not X. It's Y." of being AI. In six months, it'll be "use of the word 'petrichor'" or something.
I do think there is a tendency to over-index on one or two particularly straightforward tells, and for any given feature of LLM writing you can find places where people do also use that feature (they had to learn it from somewhere, and in a lot of cases it is good writing practice — for the context in which it is used). But I'm not talking about just that, but also the general tone issue: it's bad writing regardless because it's in most cases just not appropriate for the context it's been written in.
(TBH I think the biggest likelihood for false positives comes from heavy LLM users picking up their tics: it's a natural tendency and I've already seen a few cases where it seems like that has happened).
Idk, it's more like "your writing is cliché and I don't feel like reading it because I've already read something that sounded similar countless times and it wasn't worth the read". The source of the clichés being an LLM. And maybe now humans are writing the same way as LLM output, I still am not going to read all that, sorry. If I see a sea of clichés, I'm going the other way.
I'm also not reading pumpkin spice murder mysteries for a similar reason. I'm also not reading stories where everybody clapped. Actually, I'm already familiar with petrichor, so unless someone has surrounded the word "petrichor" with non-cliché prose, I'm also not going to read all that.
(a) believe this is human prose
(b) enjoy reading this prose
(c) would enjoy reading 100 READMEs like this.
As for invoking petitio principii and questioning other commenters' logical coherence [0], can you politely shove the argumentum ad Latinum up your ass?
Some people are still losing their shit over em-dashes, with no other tells, and humans can't use the not x; y construction anymore either, regardless of any other merit to the writing.
LLM writing is verbose and meandering, but people are making a much bigger deal over this stuff than necessary for virtue signalling purposes. You don't want to read someone else's LLM writing? Get a summary of the page from yours. No time wasted, no pretentious posturing, and you don't make the error of assuming because the piece was written by an LLM that there was no thought put into the subject or there's no value in what is being communicated.
This is very handwavy and dismissive. It is pretty safe to assume that must of us catch it most of the time because the simple fact is so many people just copy and paste whatever the LLM outputs without even trying to edit it or mask that they used one. We’ve all seen so many examples of the exact same cadence and verbiage that we’ve learned how to identify it pretty reliably. The ones who are “slipping past us” are actually putting in the work make not just pasting raw LLM outputs, which is the real issue here. If somebody has edited it meaningfully after the fact then it’s not the same crime.
> It is pretty safe to assume that must of us catch it most of the time because the simple fact is so many people just copy and paste whatever the LLM outputs without even trying to edit it or mask that they used one.
This statement does not logically cohere. "We can spot it because so many people make it easy to spot." You don't see how this is just petitio principii in action?
What is your point? It’s not that complicated. Obvious slop is obvious. Maybe there are some humans out there who sound like Claude but I’m not going to force myself through 900 slop blog posts on the off chance that one of them might actually be written by a human.
Maybe some people stop reading LLM slop purely because it violates their moral principles or whatever but most people bail out because slop is mentally painful to read. If you are a human and you write like today’s AI find a different writing style, not because reads like AI, but because it reads like shit.
AI writing just means "writing I don't like" now. Just like Nazi means whatever and whoever I politically disagree with. Words have lost their meaning.
You're right that Nazi doesn't mean Nazi anymore. It means neonazi / white supremacist / white nationalist, which is a much broader group of people that, for some baffling reason, are under the impression that people don't care about their fascism and racism anymore.
Clearly not, or this wouldn't be something people discuss at all.
There are lots of people who belong to the above groups, sure, but at least here in Germany Nazi is now applied to basically anyone who doesn't vote green, it's ridiculous.
Just like "violence" can now mean speech you don't agree with, "genocide" means military action you don't agree with, and nobody seems to know what "woman" means anymore, Maybe the solution is to stop using words.
I am bad at recognizing LLM writing off the bat, though I am getting better. It's pretty common that the writing is good enough to get me reading on a topic I am interested in; then, once I am invested in the piece, it turns out to be shallow, wildly incomplete, or simply wrong.
It's common enough that it's training me to recognize and recoil from AI tics through sheer classical conditioning.
> we readers shouldn’t be expected to labor to understand a sentence that the writer themselves didn’t work to create.
What about answers that an LLM gave to a question that we ourselves asked? Should we “labor” to understand that answer?
I think the argument, as presented in this and other similar pieces of critique, is too simplistic.
I do understand the criticism, but I think it should be framed in a different manner. The problem, when we read a long form piece by an author, is that we imagine that there’s another “mind” at the other side. We imagine that we are following the reasoning within the mind of a fellow human being, the writer. There’s an implied sort of “intimacy” to it. And the breach is when we are fooled into thinking that we are engaged in human communication, only to discover that there is a machine on the other side.
When we ask questions to an AI, this problem does not exist, because we are fully aware that the entity on the other side is not a human being.
Yet there is no doubt that the reply from an AI can contain information that is very much worthy of our time, and of our “labor” and effort to understand it.
So I think this ultimately will be about disclosure. As long as we are being made aware of the percentage of AI use in a text, explicitly or implicitly, I think we will actually grow to accept it.
For me, writing is an activity of expressing my feelings and conveying my thoughts. I rarely left that to LLM simply because one does not contract out activities one cherishes.
I (am kinda forced to) use LLM to generate maybe 40% of the code at work, that is after my review and modifications. But I pretty much wrote all of the comments by myself. I can get into the flow by writing comments.
I liked your piece, and agree with almost all of it, but I'm surprised by your faith in the accuracy of Pangram at detecting AI writing. Is your faith based on testing it with lots of writing of known origins, or are you just saying that it reaches the same conclusion that you do as a talented human?
In particular, I wondered if you have tried running all of your own writings through it to verify that it thinks you are human. I was struck by Freddie deBoer's recent piece where he did this and said it often failed: https://freddiedeboer.substack.com/p/i-wouldnt-say-pangram-i...
What percentage of false positive rejections would you find acceptable? Would you accept this even if it forced you to change the way you write?
>I wondered if you have tried running all of your own writings through it to verify that it thinks you are human. I was struck by Freddie deBoer's recent piece where he did this
Maybe I'm taking the "all" too literally here, but I read the article, and I'm not seeing anywhere the author ran a substantial portion of his corpus through Pangram to determine the false positive rate. That would be really interesting to see.
He does
* give an example of a piece of his writing that was, when ran in segments, flagged as generated (which he disputes)
* multiply the size of his corpus by pangram's published false positive rate and estimate that a few of his pieces would be flagged
* get the pangram model to label a piece "100% AI" when it only has 3 generated sentences
* demonstrate the ability to intentionally trigger a false positive
My experience is using Pangram quite often with lots of writing of all flavors (including a bunch of known origin).
As for my own writing, I didn't do this experiment, but one of my co-workers did -- and over 176 posts spanning 22 years, all 176 (well, 177 now with my latest) are 100% human. This is not hugely surprising in that (in addition to me having actually written them!) my voice is very... distinctive. What would be more entertaining would be to try to get an LLM to write like me and fool Pangram that way. I still think that this would be difficult based on the experiences that I've heard, but it wouldn't surprise me if you could pull it off (and I would assuredly find the result entertaining!).
In the dimensions that we use Pangram in the most actionable sense (namely, to audit our own public writing), I am unconcerned about false positives, and leave it to Oxide authors to rework/recast as needed. (Though it sounds like Freddie didn't even need to do that -- he just needed to provide a longer sample.)
Now that you’ve seen it can be brittle (e.g. if a small sample is provided, per this single case), would it be sensible to add a disclaimer to the post? It’s a great ad for the tool (& I’d love for a perfect tool to exist!), so it’ll sell subscriptions & we wanna make sure that some teacher out there doesn’t falsely accuse a kid, or engineer doesn’t think worse of their colleague unfairly, etc.
False negatives are mentioned, but the false positive is what could hurt people.
This is exactly how we should approach problems. Not through just throwing more resources at it (fighting GPU compute with GPU compute) but by being clever.
Thanks a lot for sharing!
Feeding it samples of a long-going conversation with Gemini 3.1 Pro is interesting. The first message seems to get flagged instantly, but later ones sometimes pass as human. Or at least more human-ish.
If I read the blogpost correctly, you've only "trained" on prompt<->response and not interactive sessions?
Interesting! Amusingly, if I feed that detector this blog post, it identifies it as confidently robot (97 out of 100 test passages). And running through my last five blog entries, they are all over the map, with three deemed at least "likely robot." Looking further back in time (and taking a somewhat random example), a blog entry from 2008, "Concurrency's Shysters"[0], is also deemed as similarly confidently robot (also 97 out of 100); do you expect this high a false positive rate?
"you probably shouldn’t let it write it for you if you actually expect the rest of us to read it." definitely resonates with me.
Where I kind of disagree is that I don't think most readers will revolt. I think the mountain of LLM slop has actually changed people's behavior in more ways than one. Some are already relying on LLMs to summarize articles: then it doesn't matter to them who wrote it, they're just consuming machine-condensed content with no way to tell if a human or an LLM wrote the original piece. Or if their summarizer hallucinated.
In a way, The ubiquity of AI-generated material will force the world to acknowledge the superiority of the human mind. Already on Youtube there are channels proudly claiming their music was not generated by AI. Will the software industry have similar disclaimers? (some already have).
It is not clear to me what the author is SPECIFICALLY against.
Only saying "LLM writing" is honestly lazy writing. Specifically what?
I get the glaring cases, I get the idea that if the prose is generated then maybe also the idea, I get the feeling when reading a complete LLM authored piece.
But that doesn't help the piece, because - beside those glaring cases - most writing today is a mix between authors ideas and LLM prose.
A lot of similar pieces have not considered a post is both the first and final work of a thing: opinions, and experience, and less of much in cited facts; with AI, that there even was a revision pass at all.
I guess there is a kind of participatory element to the discourse where, if you want an audience, there is an editing process. Whereas in other cases, we wrote these as progress notes on an unknown journey, breadcrumbs or upturned stones to mark a path to the horizon.
Maybe it's the difference between writing as a mode of discovery, retreading the mental arc of a solution, and writing something honed to leave a mark.
If you need AI for the draft, then it does not need to be written at all. You dont have a thing to say, you just have requirement to produce a lot of words.
And in that case, no one needs to read it ai or not.
"I have fantasized about sentencing the author to read them aloud, certain that they themselves will be unable to endure the slop that they are foisting upon the rest of us.)"
>to use an LLM to write is to void the social contract between writer and reader: we readers shouldn’t be expected to labor to understand a sentence that the writer themselves didn’t work to create.
Pretty much sums up the issue re: workplace lazy AI dumping on folks as well.
I'm confused and disturbed by the need to invoke Pangram (a model) as the arbiter of slop here. Slop, like smut, is self evident. You know it when you see it. Yes, some effort may be required before realizing that something is slop, which, yeah, is annoying, but that's nothing in comparison to outsourcing your shiite detection to a model! What do you get, except the loss of self worth, by needing a model to have the confidence to call something slop?
> Why do people have this reaction? Beyond having to endure aggravating stylistic tics, when reading a piece that has had substantial LLM assistance, we — the readers — don’t know what is real and what isn’t.
This is well said. But, here too, I would pause and reflect on what it means to (think you) know what is real and what isn't in a pre-LLM setting. For example, authority bias predates LLMs, and can have disastrous consequences.
The part about false positives was telling. The author doesn't want to demonize human trash, that's not fashionable, they're only concerned about virtue signalling.
When I write technical documentation, here's how I use LLMs:
* If I need to learn something before I write about it, I rely on LLMs heavily to answer questions that I have about other source materials, e.g. to clear up ambiguities.
* I've recently started prompting it to find grammatical and spelling errors.
* And I've prompted it to find technical errors, places where I'm just wrong.
For all the prompting, I additionally tell it to not rewrite anything or offer any prose suggestions. It can keep all that to itself, thank you.
And I verify what it gives back for correctness.
(I'd encourage non-native speakers to use LLMs in much the same way. Don't sacrifice your human voice by letting the AI rewrite your words. Personally, I'd very much rather hear it from you, blemishes and all, than hear it from an AI.)
But if I could step back for a minute:
Why write anything?
If your writing goal is to flood the zone and make as much money as humanly possible from ads, then hell yeah, paperclip the everliving shit out of that.
But if your writing goal is to learn material or share material, then put that LLM on the back burner and don't use it to directly generate your text. It's bad for you, and the results are subpar.
When I'm learning something, I can go through reams of tokens and then, once I understand it, I digest that to single a paragraph about the topic. The paragraph is as concise and as helpful as I can make it. Now, I could just share the prompts that I went through with those pages of back-and-forth with the LLM... but wouldn't you rather just read the concise paragraph that gets the point across?
It's not hard to be better than an AI at writing for humans, so the minimum low bar to aim for is "better than an AI". And we can all get there with a small amount of practice. The real goal is to greatly exceed the LLMs' capabilities for sharing information.
Finally, I think everyone should write a lot. Blogs, morning pages, fiction, technical books, letters, whatever. Especially when it comes to technical content, nothing makes you do your research like putting your ass out in the ether to get flamed by 5 billion people. And teachers the world over know the best way to learn something is to teach it. Pick a topic, research, and write it up more clearly and concisely than anyone else ever has. You'll learn so much, and your readers will, as well. Writing fires up your brain. Don't give that up to an LLM.
“Personally, I'd very much rather hear it from you, blemishes and all, than hear it from an AI.”
Yes! Well said. If I already know someone, reading their own words, technical or businesss or personal, is meaningful to me. Warts and all. And if I don’t yet know the author then I definitely want to read their own words so I can get to know them.
Either way, taking the time to think and then write is a gift and I respect that.
You're marginalizing yourself. I don't have hard data for writing, but I do for another area: YouTube Thumbnails. AI generated thumbnails outperform human thumbnails, often with a +2-5% delta in CTR. Yet "so many" people loudly complain about how they HATE AI thumbnails and block channels that have them. Clearly the incentive is there, and (in the case of YT) these are mobs of angry people who don't really matter but scream and bitch as if they did.
I hate to say this but AI assisted short pitch deks from founder to angel (usually their first time) have improved with AI. But also, they follow the same formula so are sorta obvious. Still, the decks are generally more business focused than typical founders early deck being very product/solution oriented.
AI detection should NEVER be used in an educational setting where the only acceptable false positive rate is 0%. That being a rate that which will never be achieved.
Yeah, even Pangram is a bit problematic here. It has been notoriously fragile. Minor edits can flip scores from 100%-human to 100%-AI, because Pangram is crazy sensitive to local and surfacial features of text. Simple consulting with LLMs for word choices can result in 100%-AI score. Insane.
Look, in the future we may get to a point where LLMs are indistinguishable from humans in writing style.
Even then, I would say that using an LLM is robbing you of the process of writing, a process that is crucial to developing and understanding your own ideas.
Think about the last time you wrote something for consumption and the sentence to sentence thought processes you’re going through. I bet a lot of that was “is that right?” Or “does that make sense?” Or “am I communicating this at the level of my reader?”.
All of that is fundamental to your readers understanding, but more importantly, its fundamental to YOUR understanding.
This meme of trying to make it sound like LLM text is so obvious is a joke. It’s literally not, you can tell it to write in literally any style and given just a bit of an example of a person’s writing style, frontier models copy it completely and effectively. This argument can probably be leveled at vanilla raw output from an LLM, but even the slightest attempt at obfuscation bears solid fruit.
If someone uses an LLM to write and is able to tailor their writing such that it isn't obviously written by an LLM, then I'm fine with it! But two of my otherwise-favourite news sources -- the Hacker News front page and FT Alphaville -- are inundated by articles where the LLM usage is blindingly obvious.
Well, give it a shot -- you'll likely find that that technique doesn't work nearly as well (at least with Pangram 4) as you think it might. When we had Max on the podcast[0], Adam explicitly asked him about exactly this (after all, you can give an LLM access to Pangram and let it iterate!), and Max reported that someone had attempted to do this -- and ended up burning through $700 in tokens and had a "sad Claude." Another interesting bit: according to Max, newer models are diverging more from human writing not less. I think that that was more anecdotal than quantified, but an interesting comment nonetheless.
99% of college essays and pretty much everything “product” in corporate America is now LLM generated with some marginal oversight. It passes muster for the most part.
This isn't very effective on any models released in recent years. With older ones, you used to be able to influence writing style significantly by just putting examples in the context, but newer models have gone through so much assistant RLHF, they really want to revert back to their default "assistant voice" during their turn.
You can still influence their writing style in a broad manner that might look correct at a glance, but the repetitive little patterns that give it away will always be there - if it was that easy to get rid of them, don't you think the AI labs themselves would've done it before releasing the models?
I think nowadays defeating the detection probably looks like finetuning a smaller LLM and getting it to paraphrase the text from the other one (or just using a more obscure finetune: it'll probably have its own cliches and habits but it will be at least different). As an added bonus this also likely removes the fingerprinting from the output as well. But I think most people are not going to bother with this.
> To those who read broadly, the hand of the LLM is so clear it’s as if the writer’s intellectual fly is open
I dunno, man, according to Hardcover, I've read 76 fiction books this year, and I can't tell. All the "AI tells" fail the vibe check. I'm a writer and I get flagged by many of them.
I vaguely recall that researchers were able to train people to tell, but only for a minority language that AIs likely aren't particularly good at mimicking, and after training.
This whole thing reminds me of how "you can recognize a vegan because they'll tell you." There, you have a ton of false negatives (i.e., since you aren't polling people to find out if they're vegan, you're only flagging the obvious vegans and missing all the regular people who happen to be vegan).
Except here, it's a bunch of false positives and negatives I bet. You don't really have a way of knowing, so you're accusing some people (without complete accuracy) and missing some people (without complete accuracy). But you have no way of knowing, so you're just like "hell yeah, my vibes tell me I'm right."
I think anyone claiming 100% accuracy is wrong, but the recent Claude models, for example, have a writing style that is sufficiently distinct that claiming people can't recognize it is like claiming you can't recognize the styles of particular famous authors. Yes, particular elements of their writing are going to be used by others, and it's possible to disguise their style or emulate it deliberately, but it's pretty hard to accidentally write like them.
> Indeed, Pangram has become important to so many of us that I was thrilled when Pangram Labs co-founder and CEO Max Spero joined us recently on Oxide and Friends.
Is obvious AI-assisted writing better or worse than an obvious PR quid pro quo and/or cross-promotion?
This is (obviously?) false, but considering how well-capitalized we are at the moment, you do have me wondering what a quid pro quo would be for; perhaps in this fictional universe Pangram has lucked into some of the PCIe clock buffers that we've been scrambling to secure enough of?
By "quid pro quo" I wasn't suggesting that Pangram's PR people's podcast placement was pay-for-play, just that they traded access for your positioning of their tech and their exec in your content marketing efforts.
That's pretty normal, but the point is that a blog post which is 35% Pangram promotion may not actually be less annoying than the use of AI to help write blog posts.
Yeah, fair -- and definitely not: I am earnestly just a fan of what they built (and I also think it's really important as a way of getting a check against rampant LLM use).
I agree. Most human-generated content I've run across on the internet over the past couple decades has been fairly low quality and, to be honest, most of the self-admitted LLM-generated content is higher quality. I'm continually surprised that so many consider any content written by humans to automatically be more worth their time to read.
I'm much more interested in the content itself than the author that wrote it.
Ending with a spin on "if you didn't bother to write it why should people bother to read it." Way to rail on cliche repetitive slop, with more cliched slop. Will the irony never cease.
> Today, legitimate businesses are very careful about how they use bulk e-mail
I don’t think this is remotely true. Sure, they’re legally obliged to let you unsubscribe, and sure, it’s not dick pills, but every US company will immediately send you a newsletter when you purchase something, review requests and, if they/you use Shop for checkout, expect an abandoned cart reminder.
PR pieces and software companies don’t write tutorials to be helpful, they are advertising to you. If the LLM can do it for cheap, they really don’t care.
Nowadays we use LLMs mostly for doing agentic-based work. LLMs new Pareto frontier only make the headlines if they push the boundaries on benchmarks that are deterministic tasks. So models are encouraged to focus on these deterministic tasks that are, in nature, structured texts. I think that this makes models more “plastic” or “polished”, as opposed to natural and pleasant to read. User-based benchmarks, like LLM Arena, are for me the best we can do in order to rank models in this way, but come with its own drawback (subjective evaluation, prone to spam or techniques to promote a giving model).
My revolt is against the cognitive stress of reading generated text. A trope typically indicates I’m in for an uphill read.
I recently read this William Zinsser quote that inspired a nickname for this: Clotted Claude [1].
> Nobody has made the point better than George Orwell in his translation into modern bureaucratic fuzz of this famous verse from Ecclesiastes:
> > I returned and saw under the sun, that the race is not to the swift, nor the battle to the strong, neither yet bread to the wise, nor yet riches to men of understanding, nor yet favor to men of skill; but time and chance happeneth to them all.
> Orwell's version goes:
> > Objective consideration of contemporary phenomena compels the conclusion that success or failure in competitive activities exhibits no tendency to be commensurate with innate capacity, but that a considerable element of the unpredictable must invariably be taken into account.
> First notice how the two passages look. The first one at the top invites us to read it. The words are short and have air around them; they convey the rhythms of human speech. The second one is clotted with long words. It tells us instantly that a ponderous mind is at work. We don't want to go anywhere with a mind that expresses itself in such suffocating language. We don't even start to read.
[1] https://blog.kierangill.xyz/clotted-claude
In contemporary YouTube-script wording:
> The Ecclasiast looked under the sun, but there was something he didn't understand. Something that wasn't right. Something that was not as it was supposed to be. And here is what the Ecclesiast didn't understand. Here is what nobody understood. Not then. Not in the years that followed. Not now. It was not the swift who won the race. Not the strong who won the battle. Not the wise who earned the bread. Not the men of understanding who gained the riches. Not the men of skill who gained the favor. And here is what I found: to any story of success, there is an element of unpredictability and chance.
(There are really just two possible outcomes: either the article is right or in, say, two years, we will be all writing and talking like this, as in humans learning from mediamatically reinforced human feedback.)
> in, say, two years, we will be all writing and talking like this
Shoot me now.
Very depressing, and I believe it.
I am a bit of a luddite in this domain and have so far managed to resist the lure of using the generator to expand my thoughts, and I still catch myself writing "it's not just an X it's a Y" and other generator type tells. If it infecting my patterns it is totally entering the wider subconscious as "How to write" (Sighs)
I think Orwell did a great job there actually.
Despite its bizarre look, the sentence is evocative and eloquent. It does make me get a clear mental image from the very first word. It leaves little room for roaming and guessing, as it firmly nails elements one by one, and, by the time I reach the end of the sentence, I get the full meaning almost immediately.
This sentence is not randomly written; this is crafted with intention. TBH, it would take me hours, if not days, to write a sentence this much condensed and easy to understand. I seriously like it.
Perhaps, this is more about context -- which style to use in which situation. I'm only guessing here, but, since Orwell is offering an interpretation, he probably chose to be more clinical. He probably had a point to make and didn't want to risk vagueness up-front.
Well, I can tell what sort of angle you most enjoy. Anyways -- I think there is GOOD writing and BAD writing, but only subjectively. So if you enjoy it, power to you. It's certainly not random, but it is the sort of verbosity that turns off 99 percent of the people that would read it given a comparison. I find the former rather eloquent.
When I read the Orwell's version, I immediately had the same feeling I had when reading mathematical proofs. I hate so much to hunt the preceding text for anaphora resolution... It's just such a bad and pretentious way of writing. It's hostile to the reader with the side of flaunting author's superiority.
It's like having to sit through a party with acclaimed academics: every single one is so full of themselves, they will constantly one-up each other by belittling everyone in their workplace s.a. to make you feel how great of an intellect they possess and how much more they would accomplish, had they not been surrounded by all these bumbling idiots.
My take is that the generated stuff is terrible for communications.
It feels great to use, direct your machine minion to fill out your thoughts for you, but holy hell does it suck to be on the receiving end. Least of all is the disrespect, they don't care enough to even talk to you but worse is having to try and reason through that big incoherent blob.
Probably to only reasonable thing to do is to try and get your own mechanical agents to produce summaries. Inventing the lossy expansion algorithm(like compression but things get bigger on the wire), And we wept.
Now I am all depressed because it is probably inevitable, apparently thinking is hard and in general people are all to happy to outsource it to the machines.
I came to this post and saw that Orwell had forgotten to mention money as the prerequisite of success, which was very unlike him.
The problem is not the LLM prose appearing everywhere, it's the legions of AI-boosters appearing in every thread attacking anyone who complains.
Apparently, even though they want to spew AI prose everywhere, they want it read by humans, not by other bots, so when a few holdout places are insisting that prose be human authored they fight very hard against the rule.
I don't think I've ever seen someone on HN say that. Many people would say that AI makes them more productive at coding, not that the output is nice to read.
> I don't think I've ever seen someone on HN say that.
A 5m search got me the following:
https://news.ycombinator.com/item?id=49410941
https://news.ycombinator.com/item?id=49411042
https://news.ycombinator.com/item?id=49059571
This thread, in particular, stands out - reader makes the claim that Pangram found that the US constitution was 100% AI generated, when others tried they found 0% (or close to it) https://news.ycombinator.com/item?id=48378191
Those are not that. The first two are people complaining about other people incontinently identifying text as AI, because it's annoying to listen to unreliable hunches and aspersions. The second two are complaining about AI detectors not being very reliable. The claim made in the last one is a casual anecdote about "an AI detector", presumably told because it's amusing. It isn't a vehement statement about how you must accept slop into your life.
Sorry, to me any complaint about people complaining about prose is a vehement statement about accepting AI.
I feel that if one doesn't want to send that message, they shouldn't be attempting to convince others that rejection of AI prose must stop.
I would say that AI prose is often still a bit iffy, but I would disagree with anyone who would want to argue that this is any indication that AI prose will always be bad in the future.
i am a big fan of llms and the possibilities they enable. but i also find this type of behavior extremely rude! ai;dr for life. :is-your-human-around: is my preferred emoji for reacting to such behavior
The problem is that the AI commenters ironically want human readers for their generated content.
I feel that if you want humans to read your stuff, those humans insisting that you write your own stuff is not an unreasonable position to take.
Orwell's entire essay (Politics and the English Language) is well worth reading if you haven't before:
https://www.orwellfoundation.com/the-orwell-foundation/orwel...
(I'm guessing Zinsser's comments are from "On Writing Well", which you can also find online even though it is still under copyright.)
It's a bit of a pet peeve when people include quotes on a blog post without linking or otherwise references their source.
But Orwell's rule 4 should be ignored:
>iv. Never use the passive where you can use the active.
Orwell himself routinely ignores it, even in the first sentence of the essay:
>Most people who bother with the matter at all would admit that the English language is in a bad way, but it is generally assumed that we cannot by conscious action do anything about it.
The second clause could be rewritten in active voice by changing it to "but people generally assume". But this would make the writing worse, and Orwell, as a good writer, probably didn't even consider the option of making it worse, and therefore didn't notice the passive voice.
Passive voice is an essential tool for all good writers of English. I always give the example of the opening of Pride and Prejudice [0]:
>It is a truth universally acknowledged, that a single man in possession of a good fortune, must be in want of a wife.
The joke doesn't work in active voice. If you attribute this acknowledgement to some specific group of people then it's simply false, not a comedic exaggeration.
[0]: https://www.gutenberg.org/ebooks/42671
Technically, I think, that first sentence, by Austen, uses a passive participle but does not use a passive voice for any finite verb. I don't think that advice to avoid "the passive" is intended to apply to that situation.
For example, nobody would seriously suggest avoiding the passive participle in a sentence like "Put the broken plate in the bin". ("Put the plate that someone broke into the bin"?)
> It's a bit of a pet peeve when people include quotes on a blog post without linking or otherwise references their source.
You’re peeved with good reason. It’s the blog equivalent of posting a screenshot of an article to social media. People, please post your sources! In the age of misinformation, that’s more important than ever.
Why dont they train LLMs not to speak like that? Is it some tragedy of the commons here?
You ever wonder why recent claude models speak in riddles? I dunno, maybe all those "rare" books? They may have been rare for a reason.
A bigger question I'm interested in is why do LLMs speak like that in the first place? Is that really what you get if you took the average of the English language? It would be difficult for me to believe that.
Is there something about tuning for desirable qualities that forces LLMs to have this voice?
LLMs get a lot of finetuning, but I suspect there are two things that can cause this kind of writing:
Firstly, some parts of the RLHF involve human graders on the LLM's performance. I suspect their general bias towards a punchy, persuasive writing style could come from what biases the graders towards preferring that response, especially in shorter segments and when the grader is not focused on writing style
Secondly, later parts of the finetuning involve reinforcement learning on achieving certain tasks which are automatically graded: stuff like coding tasks. I think this can create a kind of feedback loop where the style drifts further, and you get the kind of LLM tics which are even more extreme (it might be that they incidentally help somehow with the actual tasks, or it might be a drift that comes from the grader also now being an LLM or some of this finetuning happening on output from other models). The more recent claude models seem to suffer from this a lot, moreso than earlier ones.
They are certainly trying to
english is my second, third language, and i understood the second immediately, but not the first. go figure.
It appears LLMs are much better at writing during a debate than when explicitly asked to write. The moment LLMs are tasked with composing a blog article or intro for a book they introduce all the nuances that identify the output as AI slop.
Am I the only one that finds the second one much easier to parse?
The first is of course from the King James Bible which for centuries was essentially a standard that all English speaking peoples aspired to. If you find that version difficult I would expect much literary writing before the 1940s also seems difficult. This is just to say I recommend reading the King James even if you are an atheist, as I am.
I also have to say that the first strikes me as being written by someone that might be smarter than I am, the second as being written by someone significantly less intelligent than I, yet somehow placed by society in a position of authority over me.
In the context of newly written work in the modern era, I would argue it's best to use grammatical constructions that are used in modern 20th/21st century English, at least most of the time. Those who wrote the KJV were trying to be expressive but the whole point was to do so in language that ordinary people would be familiar with.
> the first strikes me as being written by someone that might be smarter than I am
This is why Joseph Smith tried to imitate the language of the King James Bible in the Book of Mormon, albeit not very successfully.
yes, I am very schooled in Smith's attempts.
Yeah, I found Orwell's easier to understand and quite fast to read as well, but I think it's just because of the style of writing of the first one. It's from an earlier style of prose that I'm just not used to.
I also don't have a problem with large words as long as I'm well familiar with the words. The length of a word has nothing to do with the complexity of its meaning. We just have a limit to the number of pronounceable combinations of 5 letters.
I had to reread the first one but I wanted to read it! The second one was understood on the first scan but was a chore to read.
Does that make sense?
The second one was very clear and to the point. Parsing it was rewarded with instant understanding and I enjoyed the word choice. The first one was just annoying; I could tell it was just listing a bunch of pointless analogies to try to make its point sound more grandiose so I immediately started skimming, and didn't come away feeling like it meant much other than "we all die in the end". The second one made an actual point and was the one that made me want to read it. The first one was the chore for me.
The first communicated a feeling. A brief flutter in the soul of a picture of recognition, painted for your mind's eye with care by the author.
The second transcribed considerable information bandwidth through intentionally structured word choice for maximal density.
That was exactly my experience, and I found it to be a little bit depressing.
I wonder if this is due to experience with reading technical documentation?
I think the second requires deeper concentration, but is still quite readable compared to the kind of low-content engagement / SEO stuff one read on the internet even before LLMs
The grammar is more straightforward. It has some extraneous words, and makes conspicuously bad choices of vocabulary, but it's still a more direct statement.
This is why almost 100% of hard and social scientific discourse reads almost exactly like this these days.
I found it an improvement too
Second helped explain the first, appreciated having both (but I’m no genius or nothin’).
Other translations keep Zinsser's preferred lack of fuzz but avoid using "is ... to" for possession.
For example, the Lexham English Bible:
> I looked again and saw under the sun that the race does not belong to the swift, the battle does not belong to the mighty, food does not belong to the wise, wealth does not belong to the intelligent, and success does not belong to the skillful, for time and chance befalls all of them.
Easiest! Loses some color, gains much clarity.
This could be shortened to "success does not belong to the skillful, for time and chance befalls all" with no meaning lost. It's self-indulgent fluff. Meanwhile, Orwell's actually adds more to the statement - much better signal to noise.
Writing is fluff. What do you want, a list of bullet points? There would be no books. Maybe you have a career in writing 2 page books?
Seconded
I would love to use Pangram but they simply don’t allow signing up with my custom email domain. The error was “This email address can't be used for signup. Please use a different email.” I’m not about to create a Gmail is to use your service. To me the attack on the decentralized nature on Internet infrastructure is no less serious than the attack on the human provenance of writing itself.
Thanks for writing about that. I appreciate you taking the time to call out a bad actor like that.
Weird. I have recently created an account on my meme domain ($something.party) just fine.
So what you're saying is that Pangram's heuristics misidentified a legitimate input, on your very first encounter?
I'm surprised it's not on some of the third party model routers.
Disclaimer: I do not like to read LLM-generated text any more than anyone else.
IMHO a big problem with Pangram in particular is that they market it as a reliable tool that can be used to catch students cheating. This can obviously have disastrous effects on young lives, because it is not as reliable as they suggest.
Per their own benchmarks, they do not achieve 100% accuracy even on text that is published on the Internet, and which is likely encoded into the models themselves.
There is validity to their goals, but that is overshadowed by the irresponsible way in which it is marketed.
(All of this, swirling in a context where students are being told that they absolutely must become proficient at using LLMs to do exactly this kind of work by the highest levels of state and federal governments, faculty leadership, as well as the leaders of the workforce into which they hope to graduate. The message to youth is extremely muddled at best.)
Someone should make a browser extension to label HN posts with Pangram results of the top 100 posts, so I don't waste my time reading crap.
Always a pleasure reading Bryan's writing; it's like Bryan is sitting there with you and saying the words (hard to convey the feeling).
Agreed, I like the Oxide podcast as well, even though it’s hardware I’m unlikely to ever see never mind use, it’s nice that someone somewhere is still trying to be what they are trying to be.
Having dealt with Enterprise Hardware(TM) in a previous job, it’s refreshing simply to see someone look at that pile of crap and go “it doesn’t have to be that way” and then actually set out to prove it.
They have a ton of rfds publicly available that are worth reading. Recently, read this one: https://rfd.shared.oxide.computer/rfd/0161 because I'm researching clickhouse for my work (there is also a podcast ep on it). Even if you never use their hardware, just reading their work around the software they use/make is incredibly valuable as an engineer.
I have (an API, not an extension), but pangram is way to expensive for me to run, current (v4) pricing is $0.05 per 100 words. Should be absolutely doable if we split the cost between users tho.
I tried searching for good opensource / reasonably priced alternatives, pangram themselves even have some of their older architecture and training data on github/hugging face, but i never got that working reliably enough.
That is honestly the highest possible praise -- thank you. And when this piece was starting to boil inside of me last night (triggered, I'm sorry to report, by an obviously LLM-authored guest blog entry from the Rust Foundation[0]), I messaged one of my colleagues: "Time to do what I do best: bluntly say what lots of people are thinking."
Glad those words proved prophetic!
[0] https://rustfoundation.org/media/how-the-rust-standard-libra...
“Not every one of those 11,970 carries the same weight, and the paper is careful about that rather than rounding it up.”
AAUGH IT BURNS
Right?! When I hit "That work was genuinely valuable" I literally hollered in exasperation.
For me it was “the results speak for themselves” and then simply a (large) number of automated tests run that never had human eyes.
Yes, quantity famously has a quality all its own, but perhaps not where correctness checks for something this central is concerned.
I would go even further, I want a browser extension that scans all words on every page and colours them more and more transparent as the likelihood of llm prose is increased.
Years ago I used a rudimentary (text matching) Greasemonkey script that hid Reddit posts and comments from accounts matching a few behavioral/history signals.
I wonder if that idea could be modernized now for this.
I’m working on something like this! My issue is that using Pangram for it ends up being quite expensive.
Are you footing the bill yourself, or would you be supporting BYOK? Not sure whether logging in with user's account would also be a good workaround or not.
There's a website that does this: https://www.salahadawi.com/hacker-news-ai-detector
Unfortunately, it's not a browser extension and doesn't seem to have an API. I'd make a browser extension for this myself if it didn't involve paying for expensive Pangram usage.
AI slop really needs to be auto-flagged. It’s been like 20% of front page links I click on lately.
AI comments are auto-flagged, but not posts.
For me it’s AI videos or music / narration that is beyond off putting. What’s worse now it seems people are writing their YouTube scripts with Claude et al. so at times even if it is a human creator you can clearly and immediately tell the words are not their own. To those creators I have but one message: IT SUCKS. I’d rather have you ramble incoherently in your mic then reading an LLM script and I will remove you from my feed immediately. I concur with the author on all accounts. We all can tell the BS people are selling us, unoriginal ideas, shallow concepts, open ended questions that hint at exactly nothing. Don’t be an LLM echo
This. And the structure as well, like the repetition of the same points over and over.
I'm not entirely against using AI to help content creators improve their narrative, like finding common storytelling mistakes. But that's very different than using yourself as merely an avatar for LLM content.
Great piece and interesting data. The rate of LLM-based writing rejection among developers is even higher than I thought it would be.
To me, the glaring question is: What are we doing? The act of writing exists to 1) externalize and organize one's own thoughts for the purpose of considering and revising those thoughts; and 2) share one's own thoughts with other minds.
When we hand writing to a machine, we hand thinking to a machine, denying both our humanity and our role in the conversation.
> do you think readers can’t tell?
No. I have good anecdata: readers cannot reliably distinguish my own prose from LLM-written one apart from cases where LLMs use odd metaphors or one of their specific patterns. I've been specifically experimenting with that.
What kind of prompting are you using to get those results? Anything I have claude or codex write carries a ton of distinctive characteristics. Obsession with "bit-for-bit identical", "it's not the X it's the Y Z" and so on.
It's driving me nuts, I constantly have to prompt it to "explain in plain, simple English"
https://schwitzsplinters.blogspot.com/2022/07/results-comput...
Schwitzgebel, Strasser, and Crosby fine-tuned GPT-3 on Dennett's corpus and asked whether readers could pick Dennett's real answers to ten philosophical questions from four machine-generated alternatives, with no cherry-picking beyond mechanical length filters. Even Dennett experts averaged only 5.1 out of 10 (well below the 80% the authors predicted), blog readers got 4.8, and lay participants barely beat chance — though experts did rate Dennett's answers as more Dennett-like overall. Schwitzgebel stresses this isn't a Turing test (one-shot text is far easier to fake than extended interaction), but argues it foreshadows a future where machine outputs are humanlike enough that their moral status becomes genuinely uncertain, motivating his "Design Policy of the Excluded Middle": build machines that clearly lack moral status or clearly have it, not ambiguous ones in between.
My own take is : don't focus on the symbols on paper. focus on the facts about the world it is talking about. Isn't objectivity all about the facts? In future AI will have all the memory about what I have already read and it will just furnish the delta new information in the blog/writing so that I don't spend time on refreshing what I already know.
I see this at work. People are "writing" specs and design proposals with bots. This is noticeable and is a huge turn off. I don't have issues with using bots to aid research, but I'm not reading the doc you slopped together.
I work with a guy that I swear is addicted to LLMs. He uses them for literally all communication, often dropping mountains of text for design specs that could have been written with half the words. Even on a 1:1 Zoom call, he'll type things into Claude and then read me the response! It's infuriating, and I've told him on a number of occasions, in as many polite ways as I can, that I would prefer to speak and work with him instead of Claude, but he just can't break the addiction.
How can you tell if people can accurately identify AI generated text?
If a person reads AI generated text and does not notice, they by definition will not know about it.
There have been numerous cases of people accessing human created content as being AI.
There are instances where it seems relatively uncontroversial that it is AI generated, but without knowing both the amount of AI content people are exposed toand the amount that they register I don't think you can draw a conclusion of the overall state.
I don't think this is about edge cases where someone has successfully disguised the writing to some degree: the current crop of LLMs have some pretty blatant (and frankly annoying) habits by default, ones that are hard to miss once you have read a decent amount of their output. If I had to describe them broadly, I would say they are a collection of habits which are common in certain kinds of persuasive and emotive writing, but are usually applied way out of proportion to the topic at hand, which tends to make the result quite grandiose, overly dramatic, and tiring to read: a LLM will often write a TODO app README like it's a cross between a thriller novel, a political speech, and a bombshell news article. There's lots of specific tics (and just by sheer volume and uniformity almost any habit an LLM picks up is going to rapidly shoot into cliche regardless of its own merit) but this is the general effect which I think is objectionable independent of the source of the text.
I do think the sensitivity to it can vary a lot: it depends a lot on how much and how closely you read the text, and how much exposure you have to LLM writing. Certainly it seems like a lot of people just don't really notice, or at least don't care much.
> the current crop of LLMs have some pretty blatant
This is just "em dash redux." Except now we've moved on to accusing anyone who does "It's not X. It's Y." of being AI. In six months, it'll be "use of the word 'petrichor'" or something.
I do think there is a tendency to over-index on one or two particularly straightforward tells, and for any given feature of LLM writing you can find places where people do also use that feature (they had to learn it from somewhere, and in a lot of cases it is good writing practice — for the context in which it is used). But I'm not talking about just that, but also the general tone issue: it's bad writing regardless because it's in most cases just not appropriate for the context it's been written in.
(TBH I think the biggest likelihood for false positives comes from heavy LLM users picking up their tics: it's a natural tendency and I've already seen a few cases where it seems like that has happened).
Idk, it's more like "your writing is cliché and I don't feel like reading it because I've already read something that sounded similar countless times and it wasn't worth the read". The source of the clichés being an LLM. And maybe now humans are writing the same way as LLM output, I still am not going to read all that, sorry. If I see a sea of clichés, I'm going the other way.
I'm also not reading pumpkin spice murder mysteries for a similar reason. I'm also not reading stories where everybody clapped. Actually, I'm already familiar with petrichor, so unless someone has surrounded the word "petrichor" with non-cliché prose, I'm also not going to read all that.
I went to /show and chose a random project with a GitHub repo. Here's the README:
https://github.com/ucsandman/declick/blob/main/README.md
Please let me know if you
As for invoking petitio principii and questioning other commenters' logical coherence [0], can you politely shove the argumentum ad Latinum up your ass?[0] https://news.ycombinator.com/item?id=49582713
Hot off the presses of /new!
https://github.com/prathish-ks/isthmus/blob/main/README.md
https://github.com/prathish-ks/isthmus/blob/main/docs/thesis...
https://github.com/prathish-ks/isthmus/blob/main/docs/host-d...
https://github.com/prathish-ks/isthmus/blob/main/docs/threat...
https://github.com/prathish-ks/isthmus/blob/main/docs/threat...
https://github.com/prathish-ks/isthmus/blob/main/docs/baseli...
But Wait, There's More!
https://github.com/prathish-ks/isthmus/tree/main/docs
What—do—you—think————is this human?
Pieces that people aren't revolted by will be fine. Readers aren't revolting because of a flood of high quality writing though.
Some people are still losing their shit over em-dashes, with no other tells, and humans can't use the not x; y construction anymore either, regardless of any other merit to the writing.
LLM writing is verbose and meandering, but people are making a much bigger deal over this stuff than necessary for virtue signalling purposes. You don't want to read someone else's LLM writing? Get a summary of the page from yours. No time wasted, no pretentious posturing, and you don't make the error of assuming because the piece was written by an LLM that there was no thought put into the subject or there's no value in what is being communicated.
This is very handwavy and dismissive. It is pretty safe to assume that must of us catch it most of the time because the simple fact is so many people just copy and paste whatever the LLM outputs without even trying to edit it or mask that they used one. We’ve all seen so many examples of the exact same cadence and verbiage that we’ve learned how to identify it pretty reliably. The ones who are “slipping past us” are actually putting in the work make not just pasting raw LLM outputs, which is the real issue here. If somebody has edited it meaningfully after the fact then it’s not the same crime.
> It is pretty safe to assume that must of us catch it most of the time because the simple fact is so many people just copy and paste whatever the LLM outputs without even trying to edit it or mask that they used one.
This statement does not logically cohere. "We can spot it because so many people make it easy to spot." You don't see how this is just petitio principii in action?
What is your point? It’s not that complicated. Obvious slop is obvious. Maybe there are some humans out there who sound like Claude but I’m not going to force myself through 900 slop blog posts on the off chance that one of them might actually be written by a human.
Maybe some people stop reading LLM slop purely because it violates their moral principles or whatever but most people bail out because slop is mentally painful to read. If you are a human and you write like today’s AI find a different writing style, not because reads like AI, but because it reads like shit.
AI writing just means "writing I don't like" now. Just like Nazi means whatever and whoever I politically disagree with. Words have lost their meaning.
You're right that Nazi doesn't mean Nazi anymore. It means neonazi / white supremacist / white nationalist, which is a much broader group of people that, for some baffling reason, are under the impression that people don't care about their fascism and racism anymore.
Clearly not, or this wouldn't be something people discuss at all.
There are lots of people who belong to the above groups, sure, but at least here in Germany Nazi is now applied to basically anyone who doesn't vote green, it's ridiculous.
Just like "violence" can now mean speech you don't agree with, "genocide" means military action you don't agree with, and nobody seems to know what "woman" means anymore, Maybe the solution is to stop using words.
I am bad at recognizing LLM writing off the bat, though I am getting better. It's pretty common that the writing is good enough to get me reading on a topic I am interested in; then, once I am invested in the piece, it turns out to be shallow, wildly incomplete, or simply wrong.
It's common enough that it's training me to recognize and recoil from AI tics through sheer classical conditioning.
> we readers shouldn’t be expected to labor to understand a sentence that the writer themselves didn’t work to create.
What about answers that an LLM gave to a question that we ourselves asked? Should we “labor” to understand that answer?
I think the argument, as presented in this and other similar pieces of critique, is too simplistic.
I do understand the criticism, but I think it should be framed in a different manner. The problem, when we read a long form piece by an author, is that we imagine that there’s another “mind” at the other side. We imagine that we are following the reasoning within the mind of a fellow human being, the writer. There’s an implied sort of “intimacy” to it. And the breach is when we are fooled into thinking that we are engaged in human communication, only to discover that there is a machine on the other side.
When we ask questions to an AI, this problem does not exist, because we are fully aware that the entity on the other side is not a human being.
Yet there is no doubt that the reply from an AI can contain information that is very much worthy of our time, and of our “labor” and effort to understand it.
So I think this ultimately will be about disclosure. As long as we are being made aware of the percentage of AI use in a text, explicitly or implicitly, I think we will actually grow to accept it.
For me, writing is an activity of expressing my feelings and conveying my thoughts. I rarely left that to LLM simply because one does not contract out activities one cherishes.
I (am kinda forced to) use LLM to generate maybe 40% of the code at work, that is after my review and modifications. But I pretty much wrote all of the comments by myself. I can get into the flow by writing comments.
Hi Bryan,
I liked your piece, and agree with almost all of it, but I'm surprised by your faith in the accuracy of Pangram at detecting AI writing. Is your faith based on testing it with lots of writing of known origins, or are you just saying that it reaches the same conclusion that you do as a talented human?
In particular, I wondered if you have tried running all of your own writings through it to verify that it thinks you are human. I was struck by Freddie deBoer's recent piece where he did this and said it often failed: https://freddiedeboer.substack.com/p/i-wouldnt-say-pangram-i...
What percentage of false positive rejections would you find acceptable? Would you accept this even if it forced you to change the way you write?
>I wondered if you have tried running all of your own writings through it to verify that it thinks you are human. I was struck by Freddie deBoer's recent piece where he did this
Maybe I'm taking the "all" too literally here, but I read the article, and I'm not seeing anywhere the author ran a substantial portion of his corpus through Pangram to determine the false positive rate. That would be really interesting to see.
He does
* give an example of a piece of his writing that was, when ran in segments, flagged as generated (which he disputes)
* multiply the size of his corpus by pangram's published false positive rate and estimate that a few of his pieces would be flagged
* get the pangram model to label a piece "100% AI" when it only has 3 generated sentences
* demonstrate the ability to intentionally trigger a false positive
My experience is using Pangram quite often with lots of writing of all flavors (including a bunch of known origin).
As for my own writing, I didn't do this experiment, but one of my co-workers did -- and over 176 posts spanning 22 years, all 176 (well, 177 now with my latest) are 100% human. This is not hugely surprising in that (in addition to me having actually written them!) my voice is very... distinctive. What would be more entertaining would be to try to get an LLM to write like me and fool Pangram that way. I still think that this would be difficult based on the experiences that I've heard, but it wouldn't surprise me if you could pull it off (and I would assuredly find the result entertaining!).
In the dimensions that we use Pangram in the most actionable sense (namely, to audit our own public writing), I am unconcerned about false positives, and leave it to Oxide authors to rework/recast as needed. (Though it sounds like Freddie didn't even need to do that -- he just needed to provide a longer sample.)
Now that you’ve seen it can be brittle (e.g. if a small sample is provided, per this single case), would it be sensible to add a disclaimer to the post? It’s a great ad for the tool (& I’d love for a perfect tool to exist!), so it’ll sell subscriptions & we wanna make sure that some teacher out there doesn’t falsely accuse a kid, or engineer doesn’t think worse of their colleague unfairly, etc.
False negatives are mentioned, but the false positive is what could hurt people.
To human writing. Thank you!
I did some experiments with very simple AI detection. You can get a very long way with simple ngram probabilities.
https://meatbag.atomic14.com/
https://www.atomic14.com/2026/08/18/detecting-claude-with-le...
It’s very hard to make reliable though. Different models have different characteristics and you can prompt your way out of being detected.
This is exactly how we should approach problems. Not through just throwing more resources at it (fighting GPU compute with GPU compute) but by being clever.
Thanks a lot for sharing!
Feeding it samples of a long-going conversation with Gemini 3.1 Pro is interesting. The first message seems to get flagged instantly, but later ones sometimes pass as human. Or at least more human-ish.
If I read the blogpost correctly, you've only "trained" on prompt<->response and not interactive sessions?
Interesting! Amusingly, if I feed that detector this blog post, it identifies it as confidently robot (97 out of 100 test passages). And running through my last five blog entries, they are all over the map, with three deemed at least "likely robot." Looking further back in time (and taking a somewhat random example), a blog entry from 2008, "Concurrency's Shysters"[0], is also deemed as similarly confidently robot (also 97 out of 100); do you expect this high a false positive rate?
[0] https://bcantrill.dtrace.org/2008/11/03/concurrencys-shyster...
It really depends - it’s trained on fairly limited data (things I could generate from Claude opus 5 and ChatGPT (pre-Astra).
It’s now quite hard to get non AI training data…
"you probably shouldn’t let it write it for you if you actually expect the rest of us to read it." definitely resonates with me.
Where I kind of disagree is that I don't think most readers will revolt. I think the mountain of LLM slop has actually changed people's behavior in more ways than one. Some are already relying on LLMs to summarize articles: then it doesn't matter to them who wrote it, they're just consuming machine-condensed content with no way to tell if a human or an LLM wrote the original piece. Or if their summarizer hallucinated.
In a way, The ubiquity of AI-generated material will force the world to acknowledge the superiority of the human mind. Already on Youtube there are channels proudly claiming their music was not generated by AI. Will the software industry have similar disclaimers? (some already have).
It is not clear to me what the author is SPECIFICALLY against.
Only saying "LLM writing" is honestly lazy writing. Specifically what?
I get the glaring cases, I get the idea that if the prose is generated then maybe also the idea, I get the feeling when reading a complete LLM authored piece.
But that doesn't help the piece, because - beside those glaring cases - most writing today is a mix between authors ideas and LLM prose.
A lot of similar pieces have not considered a post is both the first and final work of a thing: opinions, and experience, and less of much in cited facts; with AI, that there even was a revision pass at all.
I guess there is a kind of participatory element to the discourse where, if you want an audience, there is an editing process. Whereas in other cases, we wrote these as progress notes on an unknown journey, breadcrumbs or upturned stones to mark a path to the horizon.
Maybe it's the difference between writing as a mode of discovery, retreading the mental arc of a solution, and writing something honed to leave a mark.
You could use AI to come up with a draft, and to critique ahead of posting.
The chief grief appears to be phoning in the whole process.
If you need AI for the draft, then it does not need to be written at all. You dont have a thing to say, you just have requirement to produce a lot of words.
And in that case, no one needs to read it ai or not.
"I have fantasized about sentencing the author to read them aloud, certain that they themselves will be unable to endure the slop that they are foisting upon the rest of us.)"
I wish that were true, but I fear it may not be.
https://arstechnica.com/ai/2026/07/canadian-legislator-reads...
Blocked EVERY single LLM generated blog and content farm with uBlacklist.
Another great day where Google only gives me 2 results on the first page.
Care to share the list? Or is uBlacklist the list?
>to use an LLM to write is to void the social contract between writer and reader: we readers shouldn’t be expected to labor to understand a sentence that the writer themselves didn’t work to create.
Pretty much sums up the issue re: workplace lazy AI dumping on folks as well.
In a way, I think they're doing us a favor with LLM slop: It's rare to have a signal that's 100% accurate at telling me it's safe to stop reading.
I'm confused and disturbed by the need to invoke Pangram (a model) as the arbiter of slop here. Slop, like smut, is self evident. You know it when you see it. Yes, some effort may be required before realizing that something is slop, which, yeah, is annoying, but that's nothing in comparison to outsourcing your shiite detection to a model! What do you get, except the loss of self worth, by needing a model to have the confidence to call something slop?
> Why do people have this reaction? Beyond having to endure aggravating stylistic tics, when reading a piece that has had substantial LLM assistance, we — the readers — don’t know what is real and what isn’t.
This is well said. But, here too, I would pause and reflect on what it means to (think you) know what is real and what isn't in a pre-LLM setting. For example, authority bias predates LLMs, and can have disastrous consequences.
The part about false positives was telling. The author doesn't want to demonize human trash, that's not fashionable, they're only concerned about virtue signalling.
When I write technical documentation, here's how I use LLMs:
* If I need to learn something before I write about it, I rely on LLMs heavily to answer questions that I have about other source materials, e.g. to clear up ambiguities.
* I've recently started prompting it to find grammatical and spelling errors.
* And I've prompted it to find technical errors, places where I'm just wrong.
For all the prompting, I additionally tell it to not rewrite anything or offer any prose suggestions. It can keep all that to itself, thank you.
And I verify what it gives back for correctness.
(I'd encourage non-native speakers to use LLMs in much the same way. Don't sacrifice your human voice by letting the AI rewrite your words. Personally, I'd very much rather hear it from you, blemishes and all, than hear it from an AI.)
But if I could step back for a minute:
Why write anything?
If your writing goal is to flood the zone and make as much money as humanly possible from ads, then hell yeah, paperclip the everliving shit out of that.
But if your writing goal is to learn material or share material, then put that LLM on the back burner and don't use it to directly generate your text. It's bad for you, and the results are subpar.
When I'm learning something, I can go through reams of tokens and then, once I understand it, I digest that to single a paragraph about the topic. The paragraph is as concise and as helpful as I can make it. Now, I could just share the prompts that I went through with those pages of back-and-forth with the LLM... but wouldn't you rather just read the concise paragraph that gets the point across?
It's not hard to be better than an AI at writing for humans, so the minimum low bar to aim for is "better than an AI". And we can all get there with a small amount of practice. The real goal is to greatly exceed the LLMs' capabilities for sharing information.
Finally, I think everyone should write a lot. Blogs, morning pages, fiction, technical books, letters, whatever. Especially when it comes to technical content, nothing makes you do your research like putting your ass out in the ether to get flamed by 5 billion people. And teachers the world over know the best way to learn something is to teach it. Pick a topic, research, and write it up more clearly and concisely than anyone else ever has. You'll learn so much, and your readers will, as well. Writing fires up your brain. Don't give that up to an LLM.
“Personally, I'd very much rather hear it from you, blemishes and all, than hear it from an AI.”
Yes! Well said. If I already know someone, reading their own words, technical or businesss or personal, is meaningful to me. Warts and all. And if I don’t yet know the author then I definitely want to read their own words so I can get to know them.
Either way, taking the time to think and then write is a gift and I respect that.
if its not worth your time to write, its not worth my time to read
this goes 10x for all the slide decks and google docs and wikislop everyone's trying to pass off as an accomplishment lately
You're marginalizing yourself. I don't have hard data for writing, but I do for another area: YouTube Thumbnails. AI generated thumbnails outperform human thumbnails, often with a +2-5% delta in CTR. Yet "so many" people loudly complain about how they HATE AI thumbnails and block channels that have them. Clearly the incentive is there, and (in the case of YT) these are mobs of angry people who don't really matter but scream and bitch as if they did.
I hate to say this but AI assisted short pitch deks from founder to angel (usually their first time) have improved with AI. But also, they follow the same formula so are sorta obvious. Still, the decks are generally more business focused than typical founders early deck being very product/solution oriented.
AI detection should NEVER be used in an educational setting where the only acceptable false positive rate is 0%. That being a rate that which will never be achieved.
Yeah, even Pangram is a bit problematic here. It has been notoriously fragile. Minor edits can flip scores from 100%-human to 100%-AI, because Pangram is crazy sensitive to local and surfacial features of text. Simple consulting with LLMs for word choices can result in 100%-AI score. Insane.
Look, in the future we may get to a point where LLMs are indistinguishable from humans in writing style.
Even then, I would say that using an LLM is robbing you of the process of writing, a process that is crucial to developing and understanding your own ideas.
Think about the last time you wrote something for consumption and the sentence to sentence thought processes you’re going through. I bet a lot of that was “is that right?” Or “does that make sense?” Or “am I communicating this at the level of my reader?”.
All of that is fundamental to your readers understanding, but more importantly, its fundamental to YOUR understanding.
This meme of trying to make it sound like LLM text is so obvious is a joke. It’s literally not, you can tell it to write in literally any style and given just a bit of an example of a person’s writing style, frontier models copy it completely and effectively. This argument can probably be leveled at vanilla raw output from an LLM, but even the slightest attempt at obfuscation bears solid fruit.
If someone uses an LLM to write and is able to tailor their writing such that it isn't obviously written by an LLM, then I'm fine with it! But two of my otherwise-favourite news sources -- the Hacker News front page and FT Alphaville -- are inundated by articles where the LLM usage is blindingly obvious.
Well, give it a shot -- you'll likely find that that technique doesn't work nearly as well (at least with Pangram 4) as you think it might. When we had Max on the podcast[0], Adam explicitly asked him about exactly this (after all, you can give an LLM access to Pangram and let it iterate!), and Max reported that someone had attempted to do this -- and ended up burning through $700 in tokens and had a "sad Claude." Another interesting bit: according to Max, newer models are diverging more from human writing not less. I think that that was more anecdotal than quantified, but an interesting comment nonetheless.
[0] https://oxide-and-friends.transistor.fm/episodes/ai-detectio...
99% of college essays and pretty much everything “product” in corporate America is now LLM generated with some marginal oversight. It passes muster for the most part.
Talking about attempts to bypass ML detection:
>This argument can probably be leveled at vanilla raw output from an LLM, but even the slightest attempt at obfuscation bears solid fruit
Whoops, disproven by bcantrill's comment:
https://news.ycombinator.com/item?id=49582629
Let's talk about the detection ability of corporate normies instead:
>pretty much everything “product” in corporate America is now LLM generated with some marginal oversight. It passes muster for the most part.
Goalposts: moved.
You can adapt the output if you care to do so. Almost all do not care to do so.
They just don't care to put the slightest attempt because they have a blindness to the problem. They are not doing it to intentionally mislead people.
This isn't very effective on any models released in recent years. With older ones, you used to be able to influence writing style significantly by just putting examples in the context, but newer models have gone through so much assistant RLHF, they really want to revert back to their default "assistant voice" during their turn.
You can still influence their writing style in a broad manner that might look correct at a glance, but the repetitive little patterns that give it away will always be there - if it was that easy to get rid of them, don't you think the AI labs themselves would've done it before releasing the models?
I think nowadays defeating the detection probably looks like finetuning a smaller LLM and getting it to paraphrase the text from the other one (or just using a more obscure finetune: it'll probably have its own cliches and habits but it will be at least different). As an added bonus this also likely removes the fingerprinting from the output as well. But I think most people are not going to bother with this.
so, yet another post encouraging the use of AI tools lol... great. that's the literal meaning of the post.
> To those who read broadly, the hand of the LLM is so clear it’s as if the writer’s intellectual fly is open
I dunno, man, according to Hardcover, I've read 76 fiction books this year, and I can't tell. All the "AI tells" fail the vibe check. I'm a writer and I get flagged by many of them.
And according to PhD linguists with expertise in the field, most AI tells are just the equivalent of old wives' tales. https://www.youtube.com/watch?v=ORgKY9AlybA
I vaguely recall that researchers were able to train people to tell, but only for a minority language that AIs likely aren't particularly good at mimicking, and after training.
This whole thing reminds me of how "you can recognize a vegan because they'll tell you." There, you have a ton of false negatives (i.e., since you aren't polling people to find out if they're vegan, you're only flagging the obvious vegans and missing all the regular people who happen to be vegan).
Except here, it's a bunch of false positives and negatives I bet. You don't really have a way of knowing, so you're accusing some people (without complete accuracy) and missing some people (without complete accuracy). But you have no way of knowing, so you're just like "hell yeah, my vibes tell me I'm right."
Research and experts disagree.
I think anyone claiming 100% accuracy is wrong, but the recent Claude models, for example, have a writing style that is sufficiently distinct that claiming people can't recognize it is like claiming you can't recognize the styles of particular famous authors. Yes, particular elements of their writing are going to be used by others, and it's possible to disguise their style or emulate it deliberately, but it's pretty hard to accidentally write like them.
is this... an ad
> Indeed, Pangram has become important to so many of us that I was thrilled when Pangram Labs co-founder and CEO Max Spero joined us recently on Oxide and Friends.
Is obvious AI-assisted writing better or worse than an obvious PR quid pro quo and/or cross-promotion?
This is (obviously?) false, but considering how well-capitalized we are at the moment, you do have me wondering what a quid pro quo would be for; perhaps in this fictional universe Pangram has lucked into some of the PCIe clock buffers that we've been scrambling to secure enough of?
By "quid pro quo" I wasn't suggesting that Pangram's PR people's podcast placement was pay-for-play, just that they traded access for your positioning of their tech and their exec in your content marketing efforts.
That's pretty normal, but the point is that a blog post which is 35% Pangram promotion may not actually be less annoying than the use of AI to help write blog posts.
It is less annoying, at least to me. I can read a semi-promotional post. I just bounce right off LLM writing
Yeah, fair -- and definitely not: I am earnestly just a fan of what they built (and I also think it's really important as a way of getting a check against rampant LLM use).
Your love shines through! I'll work on turning down my "assume the worst" knob, thank you.
The idea that readers "can tell" remains laughable. Readers routinely claim they "can tell" on things that turn out to be entirely hand authored.
If you want to read something good, read a good book.
I agree. Most human-generated content I've run across on the internet over the past couple decades has been fairly low quality and, to be honest, most of the self-admitted LLM-generated content is higher quality. I'm continually surprised that so many consider any content written by humans to automatically be more worth their time to read.
I'm much more interested in the content itself than the author that wrote it.
I don't want to read human slop either.
Ending with a spin on "if you didn't bother to write it why should people bother to read it." Way to rail on cliche repetitive slop, with more cliched slop. Will the irony never cease.
This reader revolted after the second unnecessary (and exclaimed!) parenthetical. You can't please some people.
> Today, legitimate businesses are very careful about how they use bulk e-mail
I don’t think this is remotely true. Sure, they’re legally obliged to let you unsubscribe, and sure, it’s not dick pills, but every US company will immediately send you a newsletter when you purchase something, review requests and, if they/you use Shop for checkout, expect an abandoned cart reminder.
PR pieces and software companies don’t write tutorials to be helpful, they are advertising to you. If the LLM can do it for cheap, they really don’t care.
This entire article is the bad toupee fallacy.
Readers think they don't like LLM-authored text because they only recognize bad LLM-authored text as LLM-authored.
Blind trials have actually shown that readers generally prefer LLM authored books to human-authored ones on the same subject.
The "trials" you're referring to is probably this study on 1000-word short stories
https://www.cambridge.org/core/journals/judgment-and-decisio...