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.
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.
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.
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.
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).
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 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 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.
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.
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 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).
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.
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?
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.
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.
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.
"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.
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?
> 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 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.
> 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).
> 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.
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).
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.
Glad those words proved prophetic!
[0] https://rustfoundation.org/media/how-the-rust-standard-libra...
AAUGH IT BURNS
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.
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?
It's common enough that it's training me to recognize and recoil from AI tics through sheer classical conditioning.
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 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.
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'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.
(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).
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?
> 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.
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.
The chief grief appears to be phoning in the whole process.
I wish that were true, but I fear it may not be.
https://arstechnica.com/ai/2026/07/canadian-legislator-reads...
Pretty much sums up the issue re: workplace lazy AI dumping on folks as well.
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.
[0] https://oxide-and-friends.transistor.fm/episodes/ai-detectio...
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 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.
this goes 10x for all the slide decks and google docs and wikislop everyone's trying to pass off as an accomplishment lately
If you want to read something good, read a good book.
Is obvious AI-assisted writing better or worse than an obvious PR quid pro quo and/or cross-promotion?
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.
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.