Substack has partnered with Pangram to enable readers to scan posts, notes and comments for machine-generated text. No AI detector is flawless, but Pangram is pretty much the gold standard.
AI is nothing more than a tool. Like all tools it has its strengths and weaknesses. It is not a substitute for people or judgement. The management people who think so are idiots. But they were always idiots. Overpaid idiots.
In my line of work, clients or potential clients approach my firm and ask me to make stuff for them. Engineering. Details not important
Often I get small or solo operators trying to get some product off the ground.
I had this guy who just could not stop using AI. In his product design, emails to me, whatsapp messages, you name it
Eventually I just called him and told him that communicating with me using a machine is extremely insulting and easy to spot, and makes him sound like a retard. He didnt say much but sent a message later. Guess what was in it
This is not meant to promote AI, but I have an important, quite intelligent business associate whose written communications, whether due to an inborn linguistic handicap or from playing too much football in an American High School, were often so baffling, I had to ask for clarification. Then all of a sudden their communicative effectiveness rose to never-suspected heights: and he confessed having delegated his writing to AI. For the purposes for which we communicate, that is a fine solution -- but, I hasten to add, not a universal one.
AI in this case is allowing a smart guy to sound like he is supposed to. In my case it amplified the moron. I told the fellow that simply being ignorant and clueless was a better and more honest approach, and would actually let ppl help him. Advice didnt take
Yes, well, I asked him about it, and he readily admitted having delegated his semantics to AI. But as I said, doing so HAS improved our communications ... just as a lousy car can still get you from A to Z..
My son's girlfriend asked me what I thought the difference was between AI and humans. I off handedly said, it was our errors/mistakes. I have returned to the statement not made with deep thought and realized it was profound for me. How can one have the lessons and triumph of Hansel and Gretel is mistakes in judgment are not made? For heaven's sake, one cannot have Christianity with no error nor its counterpart the power of redemption.
I understand where you are coming from and appreciate your sentiment. However, AI is *not* the arbiter of truth. Far from it. To imagine that whatever question you pose receives the “correct” answer may prove your undoing. I have experienced personally error in AI response, and as Eugyppius seems to allude, careful and thoughtful questioning can even make your AI concede to your points.
The above touches upon my greatest fear—that AI will make the incompetent appear competent. Incompetence being an inability to understand and question the response produced via AI inquiry.
I absolutely agree, but people are treating it like it is. I read the essays from the Cosmos Institute. A recent one discussed how representatives of the world's great religions were there to restrain AI models. The writer seemed to believe this was a good thing. I responded that the premise made no sense. But it did highlight the problem. As humans create these LLMs, what are the starting points? Is it Bentham's Utilitarianism? That could go very badly. What is interesting is that there seems to be an agreement that there is an objective truth out there. So, ironically, as Plato cannot be taught in universities for its triggering effects, AI may become one of its students.
I’ve been following Eugy on X and he has been harassing one individual in particular, who is guilty of getting tons of likes and only writes AI slop (I mean the stuff is REALLY bad and meaningless and he gets tons of followers). Eugy keeps posting his AI meter after the posts and it declares the post to be 100% AI, I’ve been laughing my ass off the whole time. The X algorithm and Grok should be able to identify AI and demote its visibility, likewise promote certain accounts who have many insightful things to say without the use of AI. So many platforms are becoming TikTok AI crap, and this is the dumbing down of our future. It’s akin to the educational “experts” who demanded we teach classes based on the lowest IQ in the room, for fear of leaving them behind. Otherwise, X should offer a special version with zero AI and just human interaction, no cheesy videos, so real and proper discussions can take place again.
I personally use AI for medical studies and journals for research on my daughter’s autism and seizures, and it has led to her seizures completely stopping with the use of a different form of B9 (folate). So, until they start blocking/controlling it, I’m sure going to keep using it!
Many years ago, with the advent of email, my mother made the logical observation that; as the quantity of information increased, the quality would necessarily decrease. Obviously there are exceptions to this, but it's ultimately accurate. She and her siblings exchanged a "round-robin" letter, and they would laboriously hand-write or manually type the family news and drop it in the mail. Now we must skim through the random thoughts of millions of people we don't know and try to find some semblance of meaning, often ignorant of the people who should be the closest to us.
A big part of how AI detectors work, is by assessing the 'averageness' of a text. AIs produce word sequences that are widely expected/predictable, whereas humans more likely to throw in unpredictable vocabulary items. Very hard for AIs to avoid this, as it's fundamentally how they work. You can indeed prompt AIs to avoid some of their more annoying rhetorical tricks (tricolon, opposing antithesis, concluding participial phrases, overly uniform sentence length).
how would they avoid the very principle by which they function? anyway we have data (including a few independent studies) on Pangram's accuracy.
the way students, journalists and others avoid AI detection in the real world, is not through prompting to avoid averageness or whatever, but through editing or revising AI output to add more human material. this works, because what neither Pangram nor anybody is very good at right now, is detecting AI-assisted text.
Pangram is allegedly 95% accurate. If we take them at their word it means that over the course of rating 10,000 Substack posts, it will claim that as many as 500 real posts are actually AI-written. This will be used by losers to cast doubt on legitimate writers such as yourself. This is not acceptable.
it’s over 95% accurate in identifying AI text as AI generated. that is, it has a modest but real false negative rate. in 10k AI generated posts, it will let 500 slide as human-generated.
the false positive rate is much smaller - approaching zero in the studies. several false flags in 10,000 posts is still a concern but all interventions have downsides.
adding, this is all a little artificial. what Pangram and other AI detectors do, is develop a probability score. then they set cutoffs for the binary AI-or-not output. the Pangram cutoff is set to avoid false positives at the expense of false negatives.
I did not use AI to translate 400+ pages of my fathers squiggly handwriting in Dutch. A diary of his year as a POW in 1944 until the "end" of the war and a frightening account of his long walk home back to Rotterdam. He was taken up in a razzia. What a harrowing experience.
AI is like the easy short version. It is helpful to get quickie information.
Was he a POW or a forced laborer? I knew someone who had been "van de straat opgepakt" and hauled off to labor in a German ammo factory; late in the war that seems to have been a common procedure.
Well, that's the key, isn't it? Respect. Respect for the writer, because the writer has respect for the reader. (What a story that must have been, BTW!)
It's a circular relationship. But thanks to AI, the hermeneutic circle now has a little dark dot in the centre. Zero respect for me as a reader? -> Zero respect for you as a writer, AI. Into the dot* you go, thanks for nothing and for wasting my time!
* (if you zoom in, it's actually an anal sphincter)
I couldn't agree more about the problem. AI is ruining content platforms of all kinds. A video by a real human about how AI is distorting history was eye-opening; Google searches return answers to historical questions based on AI-generated YouTube videos that are just plain factually wrong. (https://youtu.be/iUfP6amE7UA?si=dzgzWP88_oRJw9nC)
What to do about AI slop is less clear. I disagree with statements like "despite its considerable powers, AI will probably never be a great prose stylist". "Never" is a long time! There's no reason to think that AI won't keep getting better in many ways, including writing content that no AI detector can reliably distinguish from being human-written.
So what's to be done? That's a tough one. Some kind of chain-of-custody verification, where you not only have to prove you're a human, but show your work in creating the content? There's an opportunity for whoever can come up with a good solution.
It's a toughie. No doubt some windbag in Westminster, Wormtongue in Whitehall or Brigand in Brussels will put up Digital ID as the solution. "For the children', of course. They'll use the Swedes as the paradigm - look they've been injected with ID chips already & it's so convenient. . . etc..
Lately I have watched a couple of YouTube videos by Ben Hodges, the retired US general who lives in Frankfurt. His topic is invariably the Ukraine war, and I don't hear much that I disagree with, but in the course of the video he starts conveying a weird kind of body language that doesn't look natural because of its repetitive limitation. It actually induces me to quit watching, because it's creepy; but I suppose it saves him -- the real General -- time.
Interesting thought. In a sense, it’s not. Those dependent upon AI are those too limited to handle the alternative (hard work and study) and they become permanently addicted to such use (as we now find those who use GLP-1 drugs).
I was a copy editor for a while and the editor introduced me to the long hyphens, which I really like. As for letting AI write anything for me, I disabled spell check and prompting as I was getting lazy. Doing it myself again has been a struggle. I still don't automatically check what I've written.
However, my wife occasionally does. I recently wrote a book review for a fantastic book about faith, love, and determination.
I asked my wife to read that review before I posted it. She did and told me that I was giving too much away, that I shouldn’t highlight everything. Then, she asked AI to look at it and it said the same thing! The AI rewrote it, using my own words but eliminating a lot! I actually liked that one better than the original.
I forgot to mention that my dear friend is a specialty lawyer in Special Education (disability) and frequently combats school districts. She was approached by Harvard grads who want to train AI on her brain and writings and musings to create an app for parents of kids with disabilities. I don’t even know what to think about this. It’s genuinely scary.
> More rarely, I use AI to summarise books and scientific papers, particularly when I’m not sure of their relevance to whatever it is I’m working on.
This is akin to the way I have most often used LLMs myself. It's really useful being able to ask questions of "someone" who has *really* read the entire manual for a programming language cover to cover, and also every StackExchange post on the topic as well, and moreover can actually remember all of it. With a series of decent prompts, I can generate an 80% outline of a piece of code I need in a very short period of time. With a few more, I can refine those to about a 95% solution. Then I go after it myself, make sure everything makes sense, write benign test cases, ask it to tell me the *real* function call in the place it just made one up, or completely screwed the syntax for, and get it to 100%.
But that only works because I have 40 years of experience of doing it completely by hand behind me, and experience looking things up in the manuals on paper. So it's also critical that various companies not neglect to hire those kids fresh out of college who will some day replace me, and give *them* some of that experience as well. They can likely short circuit some of it, because the tools they have access to are more powerful than the ones I had in my youth, but they'll need to have it, and the companies will need to have *them*. Eventually, at least. Since I can pretty much guarantee that at least *one* of these days, I'm gonna die. 🤪
Wow! This is a very insightful and thought-provoking analysis of the growing role of AI in online writing. I particularly agree with the distinction between using AI as a tool for research and using it to outsource the actual act of thinking and writing. The risk is not simply that AI-generated prose is bland, but that it gradually replaces authentic human perspective with an optimized simulation of having a perspective. In the end, perhaps the most important thing we can preserve online is the feeling that there is actually a person behind the words.
Always love a beautiful Sam Kriss rant. I will read it - when I can be bothered to fire up the VPN/do whatever it takes nowadays...
Yes, you guessed it: this is classed as "Age-Restricted Content". And I refuse on principle to go through "age verification". And no, I can't message him to tell him, because that's also Too Dangerous for me.
I think that anyone and everyone involved in the invention of this ******* Online SAAAAFETY Act have just managed to jump the queue in my personal First Against the Wall list, overtaking even the AI-slingers...
AI is nothing more than a tool. Like all tools it has its strengths and weaknesses. It is not a substitute for people or judgement. The management people who think so are idiots. But they were always idiots. Overpaid idiots.
> It is not a substitute for people or judgement.
"Well, I mean, it's the *saw* that's the expert at cutting wood, so I just let the circular saw do its thing."
Those management people should be replaced by AI.
Not going to happen. They’re in charge. No chance they fire themselves. They’ll fire everyone else first.
Eugyppius,
In my line of work, clients or potential clients approach my firm and ask me to make stuff for them. Engineering. Details not important
Often I get small or solo operators trying to get some product off the ground.
I had this guy who just could not stop using AI. In his product design, emails to me, whatsapp messages, you name it
Eventually I just called him and told him that communicating with me using a machine is extremely insulting and easy to spot, and makes him sound like a retard. He didnt say much but sent a message later. Guess what was in it
more AI?
but of course...
This is not meant to promote AI, but I have an important, quite intelligent business associate whose written communications, whether due to an inborn linguistic handicap or from playing too much football in an American High School, were often so baffling, I had to ask for clarification. Then all of a sudden their communicative effectiveness rose to never-suspected heights: and he confessed having delegated his writing to AI. For the purposes for which we communicate, that is a fine solution -- but, I hasten to add, not a universal one.
AI in this case is allowing a smart guy to sound like he is supposed to. In my case it amplified the moron. I told the fellow that simply being ignorant and clueless was a better and more honest approach, and would actually let ppl help him. Advice didnt take
Yes, well, I asked him about it, and he readily admitted having delegated his semantics to AI. But as I said, doing so HAS improved our communications ... just as a lousy car can still get you from A to Z..
My son's girlfriend asked me what I thought the difference was between AI and humans. I off handedly said, it was our errors/mistakes. I have returned to the statement not made with deep thought and realized it was profound for me. How can one have the lessons and triumph of Hansel and Gretel is mistakes in judgment are not made? For heaven's sake, one cannot have Christianity with no error nor its counterpart the power of redemption.
I understand where you are coming from and appreciate your sentiment. However, AI is *not* the arbiter of truth. Far from it. To imagine that whatever question you pose receives the “correct” answer may prove your undoing. I have experienced personally error in AI response, and as Eugyppius seems to allude, careful and thoughtful questioning can even make your AI concede to your points.
The above touches upon my greatest fear—that AI will make the incompetent appear competent. Incompetence being an inability to understand and question the response produced via AI inquiry.
I absolutely agree, but people are treating it like it is. I read the essays from the Cosmos Institute. A recent one discussed how representatives of the world's great religions were there to restrain AI models. The writer seemed to believe this was a good thing. I responded that the premise made no sense. But it did highlight the problem. As humans create these LLMs, what are the starting points? Is it Bentham's Utilitarianism? That could go very badly. What is interesting is that there seems to be an agreement that there is an objective truth out there. So, ironically, as Plato cannot be taught in universities for its triggering effects, AI may become one of its students.
I’ve been following Eugy on X and he has been harassing one individual in particular, who is guilty of getting tons of likes and only writes AI slop (I mean the stuff is REALLY bad and meaningless and he gets tons of followers). Eugy keeps posting his AI meter after the posts and it declares the post to be 100% AI, I’ve been laughing my ass off the whole time. The X algorithm and Grok should be able to identify AI and demote its visibility, likewise promote certain accounts who have many insightful things to say without the use of AI. So many platforms are becoming TikTok AI crap, and this is the dumbing down of our future. It’s akin to the educational “experts” who demanded we teach classes based on the lowest IQ in the room, for fear of leaving them behind. Otherwise, X should offer a special version with zero AI and just human interaction, no cheesy videos, so real and proper discussions can take place again.
I personally use AI for medical studies and journals for research on my daughter’s autism and seizures, and it has led to her seizures completely stopping with the use of a different form of B9 (folate). So, until they start blocking/controlling it, I’m sure going to keep using it!
> The X algorithm and Grok should be able to identify AI and demote its visibility
Except their incentives all align in the other direction, at least right up to the point where the platform becomes unusable and everyone abandons it.
Pangram is not reliable enough for this purpose. We need a better tool.
Many years ago, with the advent of email, my mother made the logical observation that; as the quantity of information increased, the quality would necessarily decrease. Obviously there are exceptions to this, but it's ultimately accurate. She and her siblings exchanged a "round-robin" letter, and they would laboriously hand-write or manually type the family news and drop it in the mail. Now we must skim through the random thoughts of millions of people we don't know and try to find some semblance of meaning, often ignorant of the people who should be the closest to us.
What happens if one includes in the AI prompt: “… please write in a style that won’t be flagged by Pangram”? 😵💫
A big part of how AI detectors work, is by assessing the 'averageness' of a text. AIs produce word sequences that are widely expected/predictable, whereas humans more likely to throw in unpredictable vocabulary items. Very hard for AIs to avoid this, as it's fundamentally how they work. You can indeed prompt AIs to avoid some of their more annoying rhetorical tricks (tricolon, opposing antithesis, concluding participial phrases, overly uniform sentence length).
They avoid it all the time. Pangram is not reliable enough for this purpose.
how would they avoid the very principle by which they function? anyway we have data (including a few independent studies) on Pangram's accuracy.
the way students, journalists and others avoid AI detection in the real world, is not through prompting to avoid averageness or whatever, but through editing or revising AI output to add more human material. this works, because what neither Pangram nor anybody is very good at right now, is detecting AI-assisted text.
Pangram is allegedly 95% accurate. If we take them at their word it means that over the course of rating 10,000 Substack posts, it will claim that as many as 500 real posts are actually AI-written. This will be used by losers to cast doubt on legitimate writers such as yourself. This is not acceptable.
no, you’re confusing fpr and fnr.
it’s over 95% accurate in identifying AI text as AI generated. that is, it has a modest but real false negative rate. in 10k AI generated posts, it will let 500 slide as human-generated.
the false positive rate is much smaller - approaching zero in the studies. several false flags in 10,000 posts is still a concern but all interventions have downsides.
adding, this is all a little artificial. what Pangram and other AI detectors do, is develop a probability score. then they set cutoffs for the binary AI-or-not output. the Pangram cutoff is set to avoid false positives at the expense of false negatives.
$$$$, petalitres of water and TWh are being expended as we speak to train a model which will succeed in that task. This is "progress" at work.
I did not use AI to translate 400+ pages of my fathers squiggly handwriting in Dutch. A diary of his year as a POW in 1944 until the "end" of the war and a frightening account of his long walk home back to Rotterdam. He was taken up in a razzia. What a harrowing experience.
AI is like the easy short version. It is helpful to get quickie information.
Was he a POW or a forced laborer? I knew someone who had been "van de straat opgepakt" and hauled off to labor in a German ammo factory; late in the war that seems to have been a common procedure.
Well, that's the key, isn't it? Respect. Respect for the writer, because the writer has respect for the reader. (What a story that must have been, BTW!)
It's a circular relationship. But thanks to AI, the hermeneutic circle now has a little dark dot in the centre. Zero respect for me as a reader? -> Zero respect for you as a writer, AI. Into the dot* you go, thanks for nothing and for wasting my time!
* (if you zoom in, it's actually an anal sphincter)
I couldn't agree more about the problem. AI is ruining content platforms of all kinds. A video by a real human about how AI is distorting history was eye-opening; Google searches return answers to historical questions based on AI-generated YouTube videos that are just plain factually wrong. (https://youtu.be/iUfP6amE7UA?si=dzgzWP88_oRJw9nC)
What to do about AI slop is less clear. I disagree with statements like "despite its considerable powers, AI will probably never be a great prose stylist". "Never" is a long time! There's no reason to think that AI won't keep getting better in many ways, including writing content that no AI detector can reliably distinguish from being human-written.
So what's to be done? That's a tough one. Some kind of chain-of-custody verification, where you not only have to prove you're a human, but show your work in creating the content? There's an opportunity for whoever can come up with a good solution.
It's a toughie. No doubt some windbag in Westminster, Wormtongue in Whitehall or Brigand in Brussels will put up Digital ID as the solution. "For the children', of course. They'll use the Swedes as the paradigm - look they've been injected with ID chips already & it's so convenient. . . etc..
Lately I have watched a couple of YouTube videos by Ben Hodges, the retired US general who lives in Frankfurt. His topic is invariably the Ukraine war, and I don't hear much that I disagree with, but in the course of the video he starts conveying a weird kind of body language that doesn't look natural because of its repetitive limitation. It actually induces me to quit watching, because it's creepy; but I suppose it saves him -- the real General -- time.
How is AI slop writing different from GLP-1 weight loss?
Interesting thought. In a sense, it’s not. Those dependent upon AI are those too limited to handle the alternative (hard work and study) and they become permanently addicted to such use (as we now find those who use GLP-1 drugs).
This post hits with a quiet thunder. Not a whisper, but a roar.
I don't need panagram to know if something is AI generated. It always uses those long hyphens, etc.
Thats just one of many giveaways.
I will say for doing shit I hate like spreadsheets etc. AI is a godsend.
I was a copy editor for a while and the editor introduced me to the long hyphens, which I really like. As for letting AI write anything for me, I disabled spell check and prompting as I was getting lazy. Doing it myself again has been a struggle. I still don't automatically check what I've written.
I don’t use AI and probably never will!
However, my wife occasionally does. I recently wrote a book review for a fantastic book about faith, love, and determination.
I asked my wife to read that review before I posted it. She did and told me that I was giving too much away, that I shouldn’t highlight everything. Then, she asked AI to look at it and it said the same thing! The AI rewrote it, using my own words but eliminating a lot! I actually liked that one better than the original.
I forgot to mention that my dear friend is a specialty lawyer in Special Education (disability) and frequently combats school districts. She was approached by Harvard grads who want to train AI on her brain and writings and musings to create an app for parents of kids with disabilities. I don’t even know what to think about this. It’s genuinely scary.
> More rarely, I use AI to summarise books and scientific papers, particularly when I’m not sure of their relevance to whatever it is I’m working on.
This is akin to the way I have most often used LLMs myself. It's really useful being able to ask questions of "someone" who has *really* read the entire manual for a programming language cover to cover, and also every StackExchange post on the topic as well, and moreover can actually remember all of it. With a series of decent prompts, I can generate an 80% outline of a piece of code I need in a very short period of time. With a few more, I can refine those to about a 95% solution. Then I go after it myself, make sure everything makes sense, write benign test cases, ask it to tell me the *real* function call in the place it just made one up, or completely screwed the syntax for, and get it to 100%.
But that only works because I have 40 years of experience of doing it completely by hand behind me, and experience looking things up in the manuals on paper. So it's also critical that various companies not neglect to hire those kids fresh out of college who will some day replace me, and give *them* some of that experience as well. They can likely short circuit some of it, because the tools they have access to are more powerful than the ones I had in my youth, but they'll need to have it, and the companies will need to have *them*. Eventually, at least. Since I can pretty much guarantee that at least *one* of these days, I'm gonna die. 🤪
> Much AI content thus reads kind of like ad copy
Ha! That's the comment I've been making for *years*. Everything written by an LLM sounds like it was written by a marketer.
So, y'know, of *all* of the people who have something to fear in the "losing your job to AI" department... it couldn't happen to a nicer bunch. 🤣
Wow! This is a very insightful and thought-provoking analysis of the growing role of AI in online writing. I particularly agree with the distinction between using AI as a tool for research and using it to outsource the actual act of thinking and writing. The risk is not simply that AI-generated prose is bland, but that it gradually replaces authentic human perspective with an optimized simulation of having a perspective. In the end, perhaps the most important thing we can preserve online is the feeling that there is actually a person behind the words.
Sorry 🤖🤖🤖
Very well done. But Sam Kriss’ piece still retains the title as Top Late-Human Prose Response to AI Writing:
https://samkriss.substack.com/p/if-you-let-ai-do-your-writing-i-will
Always love a beautiful Sam Kriss rant. I will read it - when I can be bothered to fire up the VPN/do whatever it takes nowadays...
Yes, you guessed it: this is classed as "Age-Restricted Content". And I refuse on principle to go through "age verification". And no, I can't message him to tell him, because that's also Too Dangerous for me.
I think that anyone and everyone involved in the invention of this ******* Online SAAAAFETY Act have just managed to jump the queue in my personal First Against the Wall list, overtaking even the AI-slingers...