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.
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.
An infinite supply of generative AI is bound to crowd out human expression. And consequently, human communication and culture. This applies to writing, visual arts, mathematics, software.. A horrifying prospect.
> You already see this on your social media feed, if you're paying attention.
Regardless of the original provenance of the content on my feed, its presence on my feed has long been a result of ML algorithms designed to optimize engagement/outrage. For me, LLM content gave me the nudge I needed to finally quit the apps.
That isn't a problem with AI writing per se, at least not in the sense that it's a problem in a particular instance of a person encountering the writing. It's a problem with the existence and availability of the tools. I agree it is a problem, though, and it should be addressed by drastically curtailing the development and availability of the tools.
A little bit of subterfuge isn't bad here. Say you've just read about another goddamn "quietly" justified pushback, with enough other evidence that it's pretty clearly AI. What are your options?
Push back against the AI: IME, you get a lot of backpedaling, denials, unproductive arguments with people arguing that their slop is actually for your benefit, etc. They continue misbehaving.
Argue that the writing is objectively bad: You're still negatively critiquing somebody, so this has opposition from time to time, but people aren't emboldened by their use of AI, so they tend to respond much better and actually change a bit -- at a minimum, reflect on why they're being critiqued and try some better prompts, usually even better outcomes.
The "subterfuge" is just applying that latter strategy to all AI output. Whatever you're using to determine it's AI, recast that as an inherently negative property of the writing.
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.
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.
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.
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.
I did have an experience with AI where I was asserting something min some text I was having it proof. And it actually said “I double checked your statement and here and I think you made a mistake. I rephrased it with the correct background and interpretation” (paraphrased). And it was completely right. I had misinterpreted some data and it went and got the data and double checked me. The final product it produced was much better than what i wrote originally (in large part because it was no longer false). So AI maybe closer than you think to those conditions.
> But the prompt they fed into the AI? Much higher bitrate.
You might be projecting from your own use of AI. IME there’s a large overlap between your exception 1 and the kind of people who copy/paste AI responses.
I.e., when you find yourself dealing with a meat proxy, the original prompt is unlikely to provide anything useful. So as you say, “Why are you even bothering to listen to them.”
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.
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 appreciate people putting effort into writing. For one because I intrinsically value human effort, expression and connection as others have already argued, but also because one can only become a better writer by practicing. I am very much willing to be subjected to someone's practice as long as they genuinely put the effort into it.
We cannot have actual discourse, disagreement and learn from each other, and sometimes even from the process of writing something out itself without writing; or at least not nearly as good — verbal discourse can be valuable as well, but has different limitations and we live in a world where it is getting less and less.
If we let AI write for us we let it think for us.
"Which is why the Matrix was redesigned to this, the peak of your civilization. I say your civilization, because as soon as we started thinking for you, it really became our civilization"
> 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.
the problem in my opinion comes down to a sort of poor "theory of mind" on the part of agents at our current moment.
The AI is really bad at guessing what its reader already will know. Not just before reading, but also from reading the text. This means that it so often explains things that don't need explaining, and emphasizes things that don't need emphasis.
It's not Y but X. except no sane person would ever think it was Y, and it was clear to everyone it is X already (or X is not a widely understandable thing in the first place, or the AI hasn't established a satisfying reason for why this distinction isn't important, or or or).
This is, I think, due to a combination of training on human feedback rewarding surface level good text (people marking preference on their 300th answer of the day are not trying to understand whatever point the agent is making they're just seeing whether it seems like it knows what its talking about) combined with later training by automated grading systems (other llms) that always have perfect context - if I also went and read all the code the agent referenced, read all of its 12 paragraph answers in gave at every turn of this back and forth conversation, knew all the files on the hard drive and had read all the diffs it made, then yeah it would be fine to assume some of the things it does, but I don't do that when im talking with my boss, because he asked me to do it so he doesn't have to understand everything, and he expects me to surface important stuff and minimise other stuff in a sensible way
The next generation of models, it will be far less obvious what is AI-generated and what is human-written.
The vendors have an obvious motivation to make their writing less off-putting, and I don't think there is any hard architectural limit preventing them from doing so, it is just a matter of prioritising it and investing the necessary engineering resources.
> 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.
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.
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?
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.
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.
Part of better writing is that a human wrote it because writing reflects the person who wrote it. I don't want to read clear writing if it doesn't represent communication from a human being. What's the point? I want to do more than learn or understand a piece of information. I want to simply communicate with others and understand them. And part of that means having a good idea of their ability or lack thereof to communicate.
I hope this never happens and claude/openai retains its idiosyncratic writing style. That way it's easy to spot "grifters". It's also a signal that the fine details (like entire sentences) may be wrong since a clanker with no context wrote it.
> 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.
For sure, there are ~23,000 articles in the feed (which only goes back 5 years, but grows naturally as people publish and as we index more companies in software infrastructure).
The point is to index interesting jobs and articles to 1) help people follow the technologies they're interested in and 2) to help us figure out what to write about.
Articles get tagged with mentioned infrastructure technologies (e.g. Postgres [0] or Lean [1]) and get tagged semantic categories too (e.g. compilers [2] or network protocols [3]). (And all the same tagging happens for jobs as well.) The weekly newsletter [4] features articles and jobs from a few key semantic categories. But you can also build your own RSS feeds or email watchlists for whatever you're interested in.
There's a lot of filtering done to keep out marketing content, listicles, etc. And checking AI tells is the next stage of that.
Limiting to only good quality articles lets us build broader intelligence like the ranking of programming languages [5] and databases [6] and helps us figure out which companies are most focused on the technologies and so on.
Never seen that tool before, but it has an obvious bug: paste in markdown-formatted text with `-` bullet points, and it told me a sign of AI writing was: `Repeated sentence openers · 3 sentences opening “-”`
Now, the text I pasted in so happened to really be AI-generated, but I don't think using markdown bulleted lists is a reliable sign of AI-generation.
Interesting! I tried it with unordered Markdown lists (both nested and single-level) and I couldn't reproduce it. Would you mind sharing a small repro? In some ways I guess the brittleness is to be expected since IIUC, it uses regexes under the hood.
A coworker assumed I shared AI slop because I wrote about 200-300 words of text with nice formatting: different heading text levels, monospace formatting for variable names, used different highlight colours, etc.
Can’t even write pleasant technical documentation by hand after memorizing hotkeys without being falsely accused…
> 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?
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.
This is a wildly online take. Humans highly value authenticity and many professional writers still write their first draft with a pen or pencil.
David Foster Wallace famously described writing with a pen and paper instead of a word processor because it forced him to slow down enough to think deeply, which I find very relevant to this whole debate.
There is no substitute for slow thinking. Your brain can’t really “move up the stack” until you understand the layer you’re on. All abstractions paper over important details. All shortcuts are cut short. You cannot produce quality writing without doing the slow, dlaborious work of choosing your own words.
> 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.
Think its because how every article that was written by AI sounds the same, if they insist on using AI, writers should train their AI to make the article sound more like their writing, instead of just blindly follow what was produced.
> 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
I think when you are writing anything today, you should pass it through the LLM with command to desloppify it, even if you wrote it 100% yourself. AI soured perfectly tolerable language constructs and now a lot of human texts don't sound human.
I'm Japanese and can't write English, so I think in Japanese and have AI translate it. If that makes my comments read like AI, I guess there's nothing I can do about it.
AI is magical for working with multilingual people in multilingual contexts.
I think and write in English, but I have to operate a lot at complex levels in Spanish, so I do the thinking and writing in English and have AI translate it.
As a writer, it pains me to see my own writing, when I read it back in Spanish, sound like AI.
But there's not much I can do about it, considering hiring a personal human translator for every little thing and who can write in a unique way in Spanish to mimic the native equivalent to my style is overwhelmingly unaffordable and too slow for the every day translation type work.
Good translations require context to carry over, and that is not always a 1:1 translation from the source text. Your human text almost always needs to be structurally modified if it is to be properly translated into the target language.
I guess there's something you can do about it: learn the language (I know it's hard and everything, but "nothing I can do" is something I hear an awful lot with AI these days).
I had a terribly hard time in school with English, I've spent so much time crying and yelling over my tests and homeworks - and now the great equalizer has arrived, anyone can achieve anything, the sky's the limit (or the tokens).
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.
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.
An infinite supply of generative AI is bound to crowd out human expression. And consequently, human communication and culture. This applies to writing, visual arts, mathematics, software.. A horrifying prospect.
9 replies →
> You already see this on your social media feed, if you're paying attention.
Regardless of the original provenance of the content on my feed, its presence on my feed has long been a result of ML algorithms designed to optimize engagement/outrage. For me, LLM content gave me the nudge I needed to finally quit the apps.
(Now what am I doing in HN...)
I think it's intentional. When AI produces more than we can read, we have to rely on AI to read it.
It's like a CO2 scrubber that runs on coal.
1 reply →
That isn't a problem with AI writing per se, at least not in the sense that it's a problem in a particular instance of a person encountering the writing. It's a problem with the existence and availability of the tools. I agree it is a problem, though, and it should be addressed by drastically curtailing the development and availability of the tools.
A little bit of subterfuge isn't bad here. Say you've just read about another goddamn "quietly" justified pushback, with enough other evidence that it's pretty clearly AI. What are your options?
Push back against the AI: IME, you get a lot of backpedaling, denials, unproductive arguments with people arguing that their slop is actually for your benefit, etc. They continue misbehaving.
Argue that the writing is objectively bad: You're still negatively critiquing somebody, so this has opposition from time to time, but people aren't emboldened by their use of AI, so they tend to respond much better and actually change a bit -- at a minimum, reflect on why they're being critiqued and try some better prompts, usually even better outcomes.
The "subterfuge" is just applying that latter strategy to all AI output. Whatever you're using to determine it's AI, recast that as an inherently negative property of the writing.
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.
7 replies →
Technopoly more relevant than ever.
The thing you are pushing against...is not real. You will develop a paranoia for no useful reason.
8 replies →
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.
3 replies →
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.
1 reply →
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.
I did have an experience with AI where I was asserting something min some text I was having it proof. And it actually said “I double checked your statement and here and I think you made a mistake. I rephrased it with the correct background and interpretation” (paraphrased). And it was completely right. I had misinterpreted some data and it went and got the data and double checked me. The final product it produced was much better than what i wrote originally (in large part because it was no longer false). So AI maybe closer than you think to those conditions.
> But the prompt they fed into the AI? Much higher bitrate.
You might be projecting from your own use of AI. IME there’s a large overlap between your exception 1 and the kind of people who copy/paste AI responses.
I.e., when you find yourself dealing with a meat proxy, the original prompt is unlikely to provide anything useful. So as you say, “Why are you even bothering to listen to them.”
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.
2 replies →
[dead]
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.
2 replies →
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...
I appreciate people putting effort into writing. For one because I intrinsically value human effort, expression and connection as others have already argued, but also because one can only become a better writer by practicing. I am very much willing to be subjected to someone's practice as long as they genuinely put the effort into it.
And most importantly writing is thinking: https://paulgraham.com/words.html And not writing is not thinking: https://paulgraham.com/writes.html
We cannot have actual discourse, disagreement and learn from each other, and sometimes even from the process of writing something out itself without writing; or at least not nearly as good — verbal discourse can be valuable as well, but has different limitations and we live in a world where it is getting less and less.
If we let AI write for us we let it think for us.
"Which is why the Matrix was redesigned to this, the peak of your civilization. I say your civilization, because as soon as we started thinking for you, it really became our civilization"
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.
Good writing is not taking a list of the top ten literary devices and just pouring them into your text without any intent.
the problem in my opinion comes down to a sort of poor "theory of mind" on the part of agents at our current moment.
The AI is really bad at guessing what its reader already will know. Not just before reading, but also from reading the text. This means that it so often explains things that don't need explaining, and emphasizes things that don't need emphasis.
It's not Y but X. except no sane person would ever think it was Y, and it was clear to everyone it is X already (or X is not a widely understandable thing in the first place, or the AI hasn't established a satisfying reason for why this distinction isn't important, or or or).
This is, I think, due to a combination of training on human feedback rewarding surface level good text (people marking preference on their 300th answer of the day are not trying to understand whatever point the agent is making they're just seeing whether it seems like it knows what its talking about) combined with later training by automated grading systems (other llms) that always have perfect context - if I also went and read all the code the agent referenced, read all of its 12 paragraph answers in gave at every turn of this back and forth conversation, knew all the files on the hard drive and had read all the diffs it made, then yeah it would be fine to assume some of the things it does, but I don't do that when im talking with my boss, because he asked me to do it so he doesn't have to understand everything, and he expects me to surface important stuff and minimise other stuff in a sensible way
The next generation of models, it will be far less obvious what is AI-generated and what is human-written.
The vendors have an obvious motivation to make their writing less off-putting, and I don't think there is any hard architectural limit preventing them from doing so, it is just a matter of prioritising it and investing the necessary engineering resources.
> it will be far less obvious what is AI-generated
Maybe not: https://mindmatters.ai/brief/ai-gen-writing-model-collapse-i...
After playing with googles new model for a day it is a breath of fresh air comparatively - but it has ticks a well.
Unlike code, or math where one can apply a quantifiable measure to what is being added in natural language is much harder to gauge.
1 reply →
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".
> 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.
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.
4 replies →
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...
3 replies →
Why?
5 replies →
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.
Some of my own writing looks like generated by AI. The older it is, the more similar.
My newest texts read more human, or at least I hope so.
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.
> toward strictly rewarding better writing
Part of better writing is that a human wrote it because writing reflects the person who wrote it. I don't want to read clear writing if it doesn't represent communication from a human being. What's the point? I want to do more than learn or understand a piece of information. I want to simply communicate with others and understand them. And part of that means having a good idea of their ability or lack thereof to communicate.
No thank you.
I hope this never happens and claude/openai retains its idiosyncratic writing style. That way it's easy to spot "grifters". It's also a signal that the fine details (like entire sentences) may be wrong since a clanker with no context wrote it.
[flagged]
[dead]
> 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.
For sure, there are ~23,000 articles in the feed (which only goes back 5 years, but grows naturally as people publish and as we index more companies in software infrastructure).
The point is to index interesting jobs and articles to 1) help people follow the technologies they're interested in and 2) to help us figure out what to write about.
Articles get tagged with mentioned infrastructure technologies (e.g. Postgres [0] or Lean [1]) and get tagged semantic categories too (e.g. compilers [2] or network protocols [3]). (And all the same tagging happens for jobs as well.) The weekly newsletter [4] features articles and jobs from a few key semantic categories. But you can also build your own RSS feeds or email watchlists for whatever you're interested in.
There's a lot of filtering done to keep out marketing content, listicles, etc. And checking AI tells is the next stage of that.
Limiting to only good quality articles lets us build broader intelligence like the ranking of programming languages [5] and databases [6] and helps us figure out which companies are most focused on the technologies and so on.
[0] https://theconsensus.dev/feed.html?q=projects%3DPostgreSQL
[1] https://theconsensus.dev/feed.html?q=projects%3DLean
[2] https://theconsensus.dev/feed.html?q=topics%3D%22Compilers+%...
[3] https://theconsensus.dev/feed.html?q=topics%3D%22Network+pro...
[4] https://theconsensus.dev/weekly.html
[5] https://theconsensus.dev/languages.html
[6] https://theconsensus.dev/databases.html
> 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. :)
Isn't this the same as Simon Willison's LLM cliche highlighter?
https://tools.simonwillison.net/llm-cliche-highlighter
Never seen that tool before, but it has an obvious bug: paste in markdown-formatted text with `-` bullet points, and it told me a sign of AI writing was: `Repeated sentence openers · 3 sentences opening “-”`
Now, the text I pasted in so happened to really be AI-generated, but I don't think using markdown bulleted lists is a reliable sign of AI-generation.
Interesting! I tried it with unordered Markdown lists (both nested and single-level) and I couldn't reproduce it. Would you mind sharing a small repro? In some ways I guess the brittleness is to be expected since IIUC, it uses regexes under the hood.
1 reply →
A reliable sign of AI writing is the middot in the tool's output. But that is to be expected, given the tool's author.
A coworker assumed I shared AI slop because I wrote about 200-300 words of text with nice formatting: different heading text levels, monospace formatting for variable names, used different highlight colours, etc.
Can’t even write pleasant technical documentation by hand after memorizing hotkeys without being falsely accused…
> 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?
Immediately reminded my of this classic xkcd: https://xkcd.com/810
hahaha that is a good one.
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.
This is a wildly online take. Humans highly value authenticity and many professional writers still write their first draft with a pen or pencil.
David Foster Wallace famously described writing with a pen and paper instead of a word processor because it forced him to slow down enough to think deeply, which I find very relevant to this whole debate.
There is no substitute for slow thinking. Your brain can’t really “move up the stack” until you understand the layer you’re on. All abstractions paper over important details. All shortcuts are cut short. You cannot produce quality writing without doing the slow, dlaborious work of choosing your own words.
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?
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
1 reply →
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.
1 reply →
Think its because how every article that was written by AI sounds the same, if they insist on using AI, writers should train their AI to make the article sound more like their writing, instead of just blindly follow what was produced.
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".
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?
> 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
I think when you are writing anything today, you should pass it through the LLM with command to desloppify it, even if you wrote it 100% yourself. AI soured perfectly tolerable language constructs and now a lot of human texts don't sound human.
I wonder which will come first, the year of AI writing well without grating ticks, or the year of the Linux desktop.
Weird comparison honestly, (comprehensible) AI writing has only been around for, what, 3 years?
AI labs should really focus now on making AI that can write like humans.
Super thoughtful discussion here. 100% human generated it seems.
For those wanting something to play with, https://adamdunkels.github.io/is-it-slop/ from May 2026, comes to mind.
[flagged]
I'm Japanese and can't write English, so I think in Japanese and have AI translate it. If that makes my comments read like AI, I guess there's nothing I can do about it.
100% this.
AI is magical for working with multilingual people in multilingual contexts.
I think and write in English, but I have to operate a lot at complex levels in Spanish, so I do the thinking and writing in English and have AI translate it.
As a writer, it pains me to see my own writing, when I read it back in Spanish, sound like AI.
But there's not much I can do about it, considering hiring a personal human translator for every little thing and who can write in a unique way in Spanish to mimic the native equivalent to my style is overwhelmingly unaffordable and too slow for the every day translation type work.
Translation is translation. It retains the structure of your human text.
Good translations require context to carry over, and that is not always a 1:1 translation from the source text. Your human text almost always needs to be structurally modified if it is to be properly translated into the target language.
I guess there's something you can do about it: learn the language (I know it's hard and everything, but "nothing I can do" is something I hear an awful lot with AI these days).
I had a terribly hard time in school with English, I've spent so much time crying and yelling over my tests and homeworks - and now the great equalizer has arrived, anyone can achieve anything, the sky's the limit (or the tokens).
[flagged]