> running Pangram against every article that enters The Consensus feed is financially impossible.
Surely reading every article in the feed is even more costly than running Pangram on it? If there are so many articles that Pangram is too expensive, that would also suggest that there are too many articles for them to be meaningfully curated, which seems like a bigger issue.
I think we should start to shift away from pure concern that a text is written by AI and toward strictly rewarding better writing. I realize that only some of this involves literal pattern matching against annoying AI ticks that grate like fingernails on a chalkboard.
But moving forward the goal should remain that writing should be comprehensible and not overly reliant on templates and other annoyances.
[edit] some good disagreements here which I appreciate. Thanks! I am continuing to think about this.
To be honest, if I am given any reason to suspect someone put less effort into writing something than they expect me to put into reading it, I'm not going to, if I have a choice. And it seems most prevalent in areas of life where you don't have a choice, so to me it feels like an abuse of of the social contract.
I could put zero effort into writing a quick post on some piece of programming knowledge. But if someone needed that information, I suspect that they would feel happy to read the sentences and not care about the process.
I don't think this "effort" argument for technical knowledge pieces makes sense to me. If there is something I need, I do some research and I look for signals of authoritative information. I don't really care about how much effort it takes to write a blog post.
> I could put zero effort into writing a quick post on some piece of programming knowledge
For starters, if you're writing something yourself, it's not "zero-effort". It might be low-effort, but you still put some effort into it. And low-effort doesn't mean worthless, as you illustrate ("but if someone needed that information...").
Consider that the refrain seems to be: "I won't put more energy reading something than you put into writing it". A "quick" post takes little energy to write, and little energy to read. That's fine!
Why do you trust something that is LLM-generated is authoritative?
Even if, on the most good-faith scenario, someone did all the research them told an LLM to add more pizzazz for some reason, there is still a chance that the LLM hallucinate a lie here and there that the author might have skipped when proof-reading if they proof-read it at all.
I can't trust a machine specialized in generating convincing bullshit.
but where do you draw the line, say somebody is strongly dyslexic and/or has issues and spends two hours back & forth with the llm to articulate their thoughts then posts, that to me is a lot more valuable than somebody typing an off the cuff comment with tiktok in the background.
AI to an extent removes a lot of signals that in the past people used to figure out if the writer spent effort on the writing (length, lexicon, grammatical correctness, you name it).
Personally as long as whatever is there is not full of load bearing seams I prefer to give the benefit of the doubt even if the "voice" sounds AI (and TBH it feels the "AI voice" issues are mostly "opus pre 5.5 voice" issues, talking to gemini does not feel very "AI", same as fable, opus 5.5 occasionally has its load bearing moments but nowhere near as much.
No. There's a problem with fully-AI writing that is independent of quality: you can produce a practically infinite supply of it at nearly zero cost. This means that if we don't push back, it's simply bound to crowd out human expression. You already see this on your social media feed, if you're paying attention.
I see inherent value in hearing from and talking to other humans. There is, unfortunately, a lot of AI content hustlers who want to rob me of this so that they can get more links to their website, more eyeballs on their product, etc.
The answer here isn't to demand that submarine marketers and SEO dudes prompt better so that the text is more pleasant to read. It's to cordon it off wholesale. And yeah, that sucks for some corner cases, but there's nothing else we can really do.
it's the same thing with rsearch papers right now as well. it seems to be the first time in history where one can produce "intellectual" content at a rate exponentially faster than one can read it or process it.
When people communicate to me I want them to maximize my comprehension bitrate. AI output is usually dense in incomprehensibility and sparse in bitrate. But the prompt they fed into the AI? Much higher bitrate. And if it's hard to understand, well I can always ask my own AI to help me interpret.
I understand it is a little selfish to want people to maximize my comprehension (what about the thousands of other readers?). I also understand there can purposes other than communication in writing (fiction -> pleasure, poetry -> art). I agree with you that we shouldn't care if it's AI or human writing for these other purposes. But when the goal is one human transferring bits from their brain to mine, playing telephone with an AI cannot be more efficient (it necessarily decreases channel capacity).
There are two exceptions.
1. When AI is good enough to "read people's minds" it can extract extra bits than what they explicitly meant to include in their prompt... but at that point why am I even bothering to listen to them, and not have the more powerful, smarter AI tell me what they think?
2. AI can run experiments or prove things, and I shouldn't have my AI duplicate that work. In this case I would want them to explain what tasks they set the AI out to do, and then provide the results. I would not complain if the results are an AI writeup, though usually I would want to know what in the results the human author felt was important.
No, I am capable of assessing whether a text is written well. I do want to know if someone cared enough to write it themselves though. That's an important piece of metadata.
It depends on the goal of the text. I don't care if technical documentation is written by human or machine. I care if it's correct and helpful. I don't care if a letter providing emotional support for a lost loved one is a perfectly accurate, beautifully structured piece of text. I care that it's heartfelt, coming from a person that meant what they wrote. Such metadata is important for the latter, but irrelevant for the former.
This is something that gets glossed over a lot in the debate. "Writing to convey functional information" has a different set of success criteria than "Writing to convey emotion and shared human experience". One might still legitimately argue that knowing that a text is LLM-generated undermines confidence in the former (there's no guarantee that a coherent mind exists that can clarify ambiguities); but much _less_ so than for the latter.
About three times as many people speak English as a second/foreign language compared to native speakers, not to mention people with no English fluency at all. That's a lot of people who might use an AI writing assistant for legitimate reasons that go beyond poor character. Same as we might do if we needed to communicate in Farsi or whatever.
> Can this be gamed? Teaching the AI to write less like stock Claude and stock ChatGPT? Yes, absolutely. And honestly that would be great. Because half of the irritation is that all these AI-written article sound the same!
A lot of the annoying AI writing is actually good writing patterns done far too often. Rule of three is something every writer should use but not three times in one paragraph, etc.
Yes, definitely. The search for a way to root out some kind of ephemeral origin or quality besides what is reflected in the text is fundamentally pointless and doesn't improve anything.
> I think we should start to shift away from pure concern that a text is written by AI and toward strictly rewarding better writing.
I am a human supremacist. I believe humans have intrinsic value and what they produce has intrinsic value (no matter how poorly written it might be). I do not think that the output of a machine has any value, nor am I interested in reading it.
Will we? I will value output from people I personally care about, because I want to know more about them. But what about the other 8 billion people? Do you value their output because they are human or for some other reason?
In an abstract sense, I agree that I value human output because I value humans, even if I never see those humans or their output. But I only can process 10 bits/second, while there is probably a million times that much being posted to the internet. I simply don't have the capacity to care about it except in the abstract sense.
I'm fine with AI writing as long its not bad. The issue is that its far easier to produce large amounts of poorly made AI writing over well made AI writing.
> I'm fine with AI writing as long its not bad. The issue is that its far easier to produce large amounts of poorly made AI writing over well made AI writing.
AI writing is some average over all writing the model was trained on (with some RLHF on top). It is mediocre by definition.
The AIs generation isn't freeform completetion from the beginning of sequence token. It's not an average over all string seen. It's a conditioned generation based on the context (i.e. prompt).
If the context can be above average quality the result should be too as long as the AI is sufficiently smart. Especially if people iterate on the produced work to push it up in quality.
However if you don't have a good topic and don't iterate to improve the result then you get incredibly poor writing.
Same reason I value home made cooking, home made furniture, and my kids drawings. Not everything is about the result itself. The process is important to me. Knowing that there is a human on the other side matters. In fact, learning that there is/isn't changes my perception entirely.
You didn't really answered my question. Why is it important to you to know that there is a human on the other side of things? You just restated what your preference is without giving the 'why' of the preference, which is what I'm curious about.
TL;DR: If you didn't put effort into writing it, you clearly don't value it enough for me to put any effort into reading it. I don't give a damn about readability.
Longer: We are now locked in a titanic, global war on attention. While once it was impossible to output content faster than it can be consumed, this relationship has inverted to the extreme. So the social norms that we evolved prior to this inversion simply cannot still apply.
While once it took tremendous resources or effort, it is now trivial and cheap to saturate my attention. And as a result, my attention is under non-stop assault.
To guard my precious attention and preserve it for the subjects I deem most worthy, I am forced to choose between what does and does not deserve my attention.
The level of effort an author invested in the creation of their work is a decent proxy to their actual assessment of its worth. I will continue to use that metric as a filter, and have found it to be helpful for maintaining the health of my mind and my relationships.
It's a game of cat and mouse, as every AI tell machine detector will have an equivalent of a reverse engineering algo back to human writing, the same with cryptographic watermarking.
Essentially another layer and commodity that users will pay more for, as humans value speed and volume over authenticity.
Same reason almost no one writes with pens or typewriters any longer, as they are niche artifacts.
> Can this be gamed? Teaching the AI to write less like stock Claude and stock ChatGPT? Yes, absolutely. And honestly that would be great. Because half of the irritation is that all these AI-written article sound the same! If they're all being written with each individial's, each company's, own rules and preferences and character that'd probably be nice?
No. There is no link. But crucially, they have given you modern source-code for the entire thing. You should be able to use an LLM agent to produce a similar tool. And the diversity of such tools may well restore diversity of textual conversation.
I read it as it's an in-house tool used at theconsensus.dev for labeling articles in their feed, because using Pangram on every article would cost too much.
> But Cynthia Dunlop's survey on tech reader preferences has shown that people overwhelming dislike AI-written articles, preferring even "bad" human writing to "good" AI writing.
I would absolutely ignore surveys on this, and pay attention to studies on this (or run your own if you have a mix of slop and human-written articles on your site.) People don't know their preferences.
It's fine to do a survey of how people think the economy is doing in order to figure out how they might vote; it's dumb to do a survey of how people think the economy is doing in order to figure out how the economy is doing.
People absolutely know what they experienced, they’re just guessing about what went wrong and how to fix it. “I hate AI” doesn’t necessarily mean “I can only please this person by not using AI”, but it sure does mean there’s something about your AI this person hates. “The economy’s doing poorly” doesn’t tell you about Nvidia’s stock price, but it does tell you about this person’s experiences and information exposure.
yep, there's a massive say-do gap in play - very similar to the way people say they prefer no makeup but behavioral studies reveal a preference for light/tastefully-applied makeup
> running Pangram against every article that enters The Consensus feed is financially impossible.
Surely reading every article in the feed is even more costly than running Pangram on it? If there are so many articles that Pangram is too expensive, that would also suggest that there are too many articles for them to be meaningfully curated, which seems like a bigger issue.
I think we should start to shift away from pure concern that a text is written by AI and toward strictly rewarding better writing. I realize that only some of this involves literal pattern matching against annoying AI ticks that grate like fingernails on a chalkboard.
But moving forward the goal should remain that writing should be comprehensible and not overly reliant on templates and other annoyances.
[edit] some good disagreements here which I appreciate. Thanks! I am continuing to think about this.
To be honest, if I am given any reason to suspect someone put less effort into writing something than they expect me to put into reading it, I'm not going to, if I have a choice. And it seems most prevalent in areas of life where you don't have a choice, so to me it feels like an abuse of of the social contract.
I could put zero effort into writing a quick post on some piece of programming knowledge. But if someone needed that information, I suspect that they would feel happy to read the sentences and not care about the process.
I don't think this "effort" argument for technical knowledge pieces makes sense to me. If there is something I need, I do some research and I look for signals of authoritative information. I don't really care about how much effort it takes to write a blog post.
> I could put zero effort into writing a quick post on some piece of programming knowledge
For starters, if you're writing something yourself, it's not "zero-effort". It might be low-effort, but you still put some effort into it. And low-effort doesn't mean worthless, as you illustrate ("but if someone needed that information...").
Consider that the refrain seems to be: "I won't put more energy reading something than you put into writing it". A "quick" post takes little energy to write, and little energy to read. That's fine!
Why do you trust something that is LLM-generated is authoritative?
Even if, on the most good-faith scenario, someone did all the research them told an LLM to add more pizzazz for some reason, there is still a chance that the LLM hallucinate a lie here and there that the author might have skipped when proof-reading if they proof-read it at all.
I can't trust a machine specialized in generating convincing bullshit.
but where do you draw the line, say somebody is strongly dyslexic and/or has issues and spends two hours back & forth with the llm to articulate their thoughts then posts, that to me is a lot more valuable than somebody typing an off the cuff comment with tiktok in the background.
AI to an extent removes a lot of signals that in the past people used to figure out if the writer spent effort on the writing (length, lexicon, grammatical correctness, you name it).
Personally as long as whatever is there is not full of load bearing seams I prefer to give the benefit of the doubt even if the "voice" sounds AI (and TBH it feels the "AI voice" issues are mostly "opus pre 5.5 voice" issues, talking to gemini does not feel very "AI", same as fable, opus 5.5 occasionally has its load bearing moments but nowhere near as much.
No. There's a problem with fully-AI writing that is independent of quality: you can produce a practically infinite supply of it at nearly zero cost. This means that if we don't push back, it's simply bound to crowd out human expression. You already see this on your social media feed, if you're paying attention.
I see inherent value in hearing from and talking to other humans. There is, unfortunately, a lot of AI content hustlers who want to rob me of this so that they can get more links to their website, more eyeballs on their product, etc.
The answer here isn't to demand that submarine marketers and SEO dudes prompt better so that the text is more pleasant to read. It's to cordon it off wholesale. And yeah, that sucks for some corner cases, but there's nothing else we can really do.
it's the same thing with rsearch papers right now as well. it seems to be the first time in history where one can produce "intellectual" content at a rate exponentially faster than one can read it or process it.
So who is it for if no humans are consuming those papers?
When people communicate to me I want them to maximize my comprehension bitrate. AI output is usually dense in incomprehensibility and sparse in bitrate. But the prompt they fed into the AI? Much higher bitrate. And if it's hard to understand, well I can always ask my own AI to help me interpret.
I understand it is a little selfish to want people to maximize my comprehension (what about the thousands of other readers?). I also understand there can purposes other than communication in writing (fiction -> pleasure, poetry -> art). I agree with you that we shouldn't care if it's AI or human writing for these other purposes. But when the goal is one human transferring bits from their brain to mine, playing telephone with an AI cannot be more efficient (it necessarily decreases channel capacity).
There are two exceptions.
1. When AI is good enough to "read people's minds" it can extract extra bits than what they explicitly meant to include in their prompt... but at that point why am I even bothering to listen to them, and not have the more powerful, smarter AI tell me what they think?
2. AI can run experiments or prove things, and I shouldn't have my AI duplicate that work. In this case I would want them to explain what tasks they set the AI out to do, and then provide the results. I would not complain if the results are an AI writeup, though usually I would want to know what in the results the human author felt was important.
No, I am capable of assessing whether a text is written well. I do want to know if someone cared enough to write it themselves though. That's an important piece of metadata.
It depends on the goal of the text. I don't care if technical documentation is written by human or machine. I care if it's correct and helpful. I don't care if a letter providing emotional support for a lost loved one is a perfectly accurate, beautifully structured piece of text. I care that it's heartfelt, coming from a person that meant what they wrote. Such metadata is important for the latter, but irrelevant for the former.
This is something that gets glossed over a lot in the debate. "Writing to convey functional information" has a different set of success criteria than "Writing to convey emotion and shared human experience". One might still legitimately argue that knowing that a text is LLM-generated undermines confidence in the former (there's no guarantee that a coherent mind exists that can clarify ambiguities); but much _less_ so than for the latter.
About three times as many people speak English as a second/foreign language compared to native speakers, not to mention people with no English fluency at all. That's a lot of people who might use an AI writing assistant for legitimate reasons that go beyond poor character. Same as we might do if we needed to communicate in Farsi or whatever.
I don't think we should paper over "the largest theft of labor in human history.”[1]
1. https://www.computerworld.com/article/4234688/microsoft-know...
The article talks about this too:
> Can this be gamed? Teaching the AI to write less like stock Claude and stock ChatGPT? Yes, absolutely. And honestly that would be great. Because half of the irritation is that all these AI-written article sound the same!
A lot of the annoying AI writing is actually good writing patterns done far too often. Rule of three is something every writer should use but not three times in one paragraph, etc.
That's a fine ideal, such as writing that doesn't use phrases like "moving forward." Similarly grating even if it's not an "AI tell".
Yes, definitely. The search for a way to root out some kind of ephemeral origin or quality besides what is reflected in the text is fundamentally pointless and doesn't improve anything.
> I think we should start to shift away from pure concern that a text is written by AI and toward strictly rewarding better writing.
I am a human supremacist. I believe humans have intrinsic value and what they produce has intrinsic value (no matter how poorly written it might be). I do not think that the output of a machine has any value, nor am I interested in reading it.
No, absolutely not. We will continue to value output that purely shows human thinking and effort.
Will we? I will value output from people I personally care about, because I want to know more about them. But what about the other 8 billion people? Do you value their output because they are human or for some other reason?
Say there was a Chinese artist in a room...
I'm not going to bother engaging with online posts if they all turn into generated replies.
> Do you value their output because they are human
Their output does have intrinsic value yes...
In an abstract sense, I agree that I value human output because I value humans, even if I never see those humans or their output. But I only can process 10 bits/second, while there is probably a million times that much being posted to the internet. I simply don't have the capacity to care about it except in the abstract sense.
I'm fine with AI writing as long its not bad. The issue is that its far easier to produce large amounts of poorly made AI writing over well made AI writing.
> I'm fine with AI writing as long its not bad. The issue is that its far easier to produce large amounts of poorly made AI writing over well made AI writing.
AI writing is some average over all writing the model was trained on (with some RLHF on top). It is mediocre by definition.
The AIs generation isn't freeform completetion from the beginning of sequence token. It's not an average over all string seen. It's a conditioned generation based on the context (i.e. prompt).
If the context can be above average quality the result should be too as long as the AI is sufficiently smart. Especially if people iterate on the produced work to push it up in quality.
However if you don't have a good topic and don't iterate to improve the result then you get incredibly poor writing.
Why?
Same reason I value home made cooking, home made furniture, and my kids drawings. Not everything is about the result itself. The process is important to me. Knowing that there is a human on the other side matters. In fact, learning that there is/isn't changes my perception entirely.
You didn't really answered my question. Why is it important to you to know that there is a human on the other side of things? You just restated what your preference is without giving the 'why' of the preference, which is what I'm curious about.
I strongly disagree.
TL;DR: If you didn't put effort into writing it, you clearly don't value it enough for me to put any effort into reading it. I don't give a damn about readability.
Longer: We are now locked in a titanic, global war on attention. While once it was impossible to output content faster than it can be consumed, this relationship has inverted to the extreme. So the social norms that we evolved prior to this inversion simply cannot still apply.
While once it took tremendous resources or effort, it is now trivial and cheap to saturate my attention. And as a result, my attention is under non-stop assault.
To guard my precious attention and preserve it for the subjects I deem most worthy, I am forced to choose between what does and does not deserve my attention.
The level of effort an author invested in the creation of their work is a decent proxy to their actual assessment of its worth. I will continue to use that metric as a filter, and have found it to be helpful for maintaining the health of my mind and my relationships.
No thank you.
It's a game of cat and mouse, as every AI tell machine detector will have an equivalent of a reverse engineering algo back to human writing, the same with cryptographic watermarking.
Essentially another layer and commodity that users will pay more for, as humans value speed and volume over authenticity.
Same reason almost no one writes with pens or typewriters any longer, as they are niche artifacts.
FTFA:
> Can this be gamed? Teaching the AI to write less like stock Claude and stock ChatGPT? Yes, absolutely. And honestly that would be great. Because half of the irritation is that all these AI-written article sound the same! If they're all being written with each individial's, each company's, own rules and preferences and character that'd probably be nice?
> They are not eligible for the sparkle icon on the feed page
This is the opposite of how most places use the sparkle icon!
It was relaxing to see some minor typos and stylistic awkwardness in this article, I have to be honest.
> This is the opposite of how most places use the sparkle icon!
I heard this from someone else today too. I'll have to think of a different icon perhaps then.
> It was relaxing to see some minor typos and stylistic awkwardness in this article, I have to be honest.
whoops
> people overwhelming dislike
Just a little thing. :)
AI labs should really focus now on making AI that can write like humans.
Ironically, AI Tells in job descriptions means the company actually pays more. I just pulled these numbers the other day: https://www.linkedin.com/posts/zmohiuddin_the-more-a-job-pos...
Pay more, or advertise fake AI slop jobs with higher pay?
Isn't this the same as Simon Willison's LLM cliche highlighter?
https://tools.simonwillison.net/llm-cliche-highlighter
Is there a link I'm missing to the tool this article is discussing?
No. There is no link. But crucially, they have given you modern source-code for the entire thing. You should be able to use an LLM agent to produce a similar tool. And the diversity of such tools may well restore diversity of textual conversation.
I don't see the source code, where did you find that?
[Edit] Or are you implying that readers should direct an LLM at the entire article and expect an equivalent tool to be produced
Haha, yes, the latter. The textual description is the critical source code. Here, it tells you that you:
* compare pre-LLM and post-LLM writing using 2023 as your index year
* create some 100+ fixed rules rather than a sophisticated model
* use a fixed threshold to kill
I read it as it's an in-house tool used at theconsensus.dev for labeling articles in their feed, because using Pangram on every article would cost too much.
That's right.
going with something low-tech and gamable is better imo - pangram would flag the good stuff too and cause collateral damage
Supposedly, SlopShape https://arxiv.org/abs/2609.15369 can ID not just human vs bot but which bot via structural patterns in the writing.
Next we need a "Verbose useless content" detector and we'll be set up to GAN the whole internet into https://xkcd.com/810/ "Mission Accomplished".
For those wanting something to play with, https://adamdunkels.github.io/is-it-slop/ from May 2026, comes to mind.
> But Cynthia Dunlop's survey on tech reader preferences has shown that people overwhelming dislike AI-written articles, preferring even "bad" human writing to "good" AI writing.
I would absolutely ignore surveys on this, and pay attention to studies on this (or run your own if you have a mix of slop and human-written articles on your site.) People don't know their preferences.
It's fine to do a survey of how people think the economy is doing in order to figure out how they might vote; it's dumb to do a survey of how people think the economy is doing in order to figure out how the economy is doing.
People absolutely know what they experienced, they’re just guessing about what went wrong and how to fix it. “I hate AI” doesn’t necessarily mean “I can only please this person by not using AI”, but it sure does mean there’s something about your AI this person hates. “The economy’s doing poorly” doesn’t tell you about Nvidia’s stock price, but it does tell you about this person’s experiences and information exposure.
yep, there's a massive say-do gap in play - very similar to the way people say they prefer no makeup but behavioral studies reveal a preference for light/tastefully-applied makeup
people love to openly hate on ai/assisted writing meanwhile https://www.salahadawi.com/hacker-news-ai-detector reveals a lot of it is being voted up to the frontpage