I agree that falling back on Pangram - or any other AI detector - is ultimately a dead end that's likely to have as many distorting effects on what and how we produce text as LLMs themselves, but I have a tangential question about one of the points on your chart, the "literature review."
I'm not interested in the ethics so much as the purpose of the activity. Isn't the point of a literature review to personally process the ideas and concepts of the other source material? For me, I would think that means reading them and then writing my own summaries as an artifact of what I've taken from the experience. An AI summary of the literature means I don't do any of that work.
This seems like an example of the AI is like bringing a forklift to the gym critique of LLM use. The point of the literature review isn't just so a literature review exists. It's to build the knowledge of the reviewer, right?
Good questions. I think “literature review” covers several different activities, and the ethics vary across them. I mostly agree when we are talking about deep lit reviews, especially for teaching or learning: if the assignment is meant to make you read, process, and synthesize the field yourself, AI replacement defeats the point. I’ve written before about my own worries about skill atrophy here.
For research writing, I’d distinguish two kinds of citations. Some papers are essential to your argument: they provide evidence, define the disagreement, or shape the frame. Those need to be read carefully. Other citations are more symbolic or positional: “this relates to a broader literature on X,” “similar work exists in Y,” “for background, see...” Academics already treat those differently, with or without AI.
So my view is that AI can help with the second kind of lit-review prose, especially when you already know the area and are placing your work among neighboring literatures. It should not create the impression that you deeply engaged with papers you only skimmed or did not read. My guess is that hallucinations become much more likely when people use AI to simulate engagement with arguments they have not actually understood.
Helpful, thanks. I’m wondering if you have any thoughts/insights into how much relying on AI summary of those related, but not core papers might be contributing to so-called “frankencitations” the inclusion of sources that don’t actually exist. (See this interview I did with someone who edits a journal and whose work is being heavily frankencited: https://www.insidehighered.com/opinion/columns/just-visiting/2026/05/06/what-happens-when-fake-citations-layer-bs-bs
I like the term "frankencitations" but this should be already easily solved with good AI-driven verification mechanisms of ensuring that every single thing cited has a DOI aligning with some other source, for example.
A bigger problem with either humans or AI is that a lot of the citations are pretty superficial. So many people, for example, cited Benedict Anderson to argue that nationalism is made up based on the title of his book even though he is decidedly not arguing that. I noticed that at least 20% of my cites on Google Scholar are also along those lines--very superficial read of my work based mostly on the abstract.
I don’t know enough about academic publishing citation to evaluate the ease of that kind of checking, but my understanding is that the frankencitations are leaving a credible digital trail in their wake and that kind of automated detection might not be so straightforward and will keep getting harder.
That bigger problem is the thing I’m most wondering about and don’t really understand. It suggests a lot of performative B.S. around academic publishing - which okay - but I wish the existence of a technology that can make convincing B.S. at scale prompted a reckoning around what’s valued in these spaces.
I use AI very much the way that you do, with a lot of iteration, back-and-forth, dictating, and heavily editing. When I finish with it, I consider it mine. We need to teach people to use it that way and not just hit publish.
I don't think disclosing it is the answer because people have knee-jerk reactions to AI, even when it's well written. I know you use AI because I've read some of your previous articles where you've talked about it. But I'll be honest with you, my AI radar doesn't go off when I read your stuff the way it does with many other people who don't edit the output and simply provide it as is.
If the final result still sounds like the person who wrote it, isn't that the whole point?
As for the winners of literary competitions, that's a whole other situation, and I don't know how they're going to deal with that. What exactly are we judging at that point: the writing, the editing, the prompting, or some combination of all three?
In the end, I don't think any detector is going to be foolproof. The bigger question is what constitutes AI-generated content. Is the way you write AI-generated? Is the way I use it AI-generated? How do we determine authorship? At some point, almost everybody is going to be using AI to help them write in some capacity.
Thanks, Betty, that’s flattering. I agree with almost all of this. When I go back to some of my earlier AI-assisted writing though, I can see plenty of things I would handle differently now, including little phrases and rhythms that would annoy me today. The funny part is that the strongest AI radar often belongs to people who use these tools a lot themselves :)
In any case, my guess is that all this becomes much less interesting in a couple of years because the incentives all point toward widespread use. What I’m trying to do is speed up the awkward transition, including less moral panic and better social norms.
Sorry, I didn't mean to be hostile at all. I've written plenty about the problem of skill atrophy in my AI series and the struggles in teaching. But I certainly don't think it warrants indescriminate ai policing.
I think I'm still less of a pessimist on disclosure and *more* of a pessimist on detectors, but I recognize that both are moving targets with sometimes pernicious incentives baked in.
I hope this landscape levels out before the decade is over.
Alexander, this is an exceptionally clear and nuanced take — one of the best I’ve read on the topic.
You’re absolutely right: the real danger lies not in the tool itself, but in how quickly we’ve allowed "provenance" to replace substance. When a probabilistic “guess” from a black-box algorithm starts functioning as a status signal and moral verdict, we enter dangerous territory. What begins as a helpful detector quietly mutates into a new form of grammar nazism — or, as I call it, the architecture of accusation.
The moment we start judging writing by how “human” it appears to a machine rather than by its actual quality and insight, we lose something essential about intellectual life.
I explored this exact mechanism — how “probably” becomes a verdict and how fear of AI is being institutionalized — in my recent essay: “Probably. How a single hedging word became the most powerful tool of institutional accusation in modern intellectual life — and what it is actually built to do.”
Your piece and mine complement each other well: you diagnose the cultural shift, while I try to map the deeper systemic and philosophical architecture behind it.
I’m still not that convinced by Pangram - I test it with Claude-generated text and it frequently comes back as 100% human when I’ve done no editing. On occasion it also identified 100% AI or a high score - but not at all consistently as far as I can tell. I find this sudden trusting Pangram on everything shift really odd and just a little suss.
It's interesting: I think I agree with you in principle about pretty much everything here. But I recently got a paid subscription to Pangram and as a creative writer, I can't stop myself from checking text when I suspect it's AI -- or at least I have found myself reaching for the detector when there was no actual need for me to discover the provenance of the text. As a result, I think I've gotten better at spotting AI text. I'm at least better at predicting when Pangram will give it a high AI score, anyway. This kind of suggests another, wider social purpose for AI detectors: to train people to notice and interrogate the patterns AI produces.
I wrote a long essay about the Shy Girl controversy and in there I talked about how "dissolving distinctions disadvantages difficulty"... In other words, when we don't know or don't care how something is produced, the easiest process wins. But knowing and caring how something is produced has historically been considered a valuable skill. Taste, discernment, whatever you want to call it. Imo we can value AI text for what it is but also become more skilled at seeing what it is not. Does that make sense?
I swing hard from the measured and thoughtful positions you’ve laid out here to wild eyed fury that stick swingers like Sam Kriss and Paul Kingsnorth inhabit. There’s nothing wrong with supermarket whole meal bread, it’s often fortified with cool things like niacin and riboflavin – it’s fine for food at an academic conference.
But really, what I want is caramelised crust that my teeth have to crash through to eat the wild fermented sourdough of human expression.
Enjoyed your post and agree with the general thrust of it.
One point I'd like to complicate is 'attention is scarce'. Another reading is that readers have correctly determined that much credentialed output is not particularly good, nor worth their time. Attention scarcity as verdict, not consequence.
You frame detection as an imperfect shortcut competing with others, but the incumbent shortcuts lost credibility because they substituted for reading. Pangram extends that problem. AI-detection anxiety is then, at least partly, due to the collapse of venue as a quality signal, and enforcement of provenance norms is the old gatekeepers trying to re-establish the legitimacy of their proxies.
I agree. I have a similar writing process to yours. It's multiple drafts and back and forths. And I appreciate being able to work through ideas through voice chat and not having to type all the time. It's interesting- I work in an industry where the information just needs to get across, so we often don't use a lot of dressing and style. I suspect that what people are so upset about is that the style or padding is not a natural way of speaking and so using a "translator" is somehow offensive. As long as nothing was hallucinated, then the arguments and thoughts are originally the author's but translated to proper grammar and academic jargon, if necessary. Just my two cents. I am curious- how are you using codex for this?
Well put. There is a continuous spectrum here which is (unsurprisingly) represented in much of 'life, the universe and everything'. It is time to discard the 'AI is evil' first call and allow that AI as a label/category/framework can be something very nuanced and is often beneficial well beyond the individual when used in balanced and appropriate ways. Our problem is that we are still thinking in the 'wild west' phase, even where pockets of expertise exist. It will take a while as the pendulum moves and our shared patience is important. I feel for those who are labelled harshly or 'failed' for what in the future will be normalised and value-adding. We still have a lot to learn in this regard.
I agree that falling back on Pangram - or any other AI detector - is ultimately a dead end that's likely to have as many distorting effects on what and how we produce text as LLMs themselves, but I have a tangential question about one of the points on your chart, the "literature review."
I'm not interested in the ethics so much as the purpose of the activity. Isn't the point of a literature review to personally process the ideas and concepts of the other source material? For me, I would think that means reading them and then writing my own summaries as an artifact of what I've taken from the experience. An AI summary of the literature means I don't do any of that work.
This seems like an example of the AI is like bringing a forklift to the gym critique of LLM use. The point of the literature review isn't just so a literature review exists. It's to build the knowledge of the reviewer, right?
What am I missing?
Good questions. I think “literature review” covers several different activities, and the ethics vary across them. I mostly agree when we are talking about deep lit reviews, especially for teaching or learning: if the assignment is meant to make you read, process, and synthesize the field yourself, AI replacement defeats the point. I’ve written before about my own worries about skill atrophy here.
For research writing, I’d distinguish two kinds of citations. Some papers are essential to your argument: they provide evidence, define the disagreement, or shape the frame. Those need to be read carefully. Other citations are more symbolic or positional: “this relates to a broader literature on X,” “similar work exists in Y,” “for background, see...” Academics already treat those differently, with or without AI.
So my view is that AI can help with the second kind of lit-review prose, especially when you already know the area and are placing your work among neighboring literatures. It should not create the impression that you deeply engaged with papers you only skimmed or did not read. My guess is that hallucinations become much more likely when people use AI to simulate engagement with arguments they have not actually understood.
Helpful, thanks. I’m wondering if you have any thoughts/insights into how much relying on AI summary of those related, but not core papers might be contributing to so-called “frankencitations” the inclusion of sources that don’t actually exist. (See this interview I did with someone who edits a journal and whose work is being heavily frankencited: https://www.insidehighered.com/opinion/columns/just-visiting/2026/05/06/what-happens-when-fake-citations-layer-bs-bs
I like the term "frankencitations" but this should be already easily solved with good AI-driven verification mechanisms of ensuring that every single thing cited has a DOI aligning with some other source, for example.
A bigger problem with either humans or AI is that a lot of the citations are pretty superficial. So many people, for example, cited Benedict Anderson to argue that nationalism is made up based on the title of his book even though he is decidedly not arguing that. I noticed that at least 20% of my cites on Google Scholar are also along those lines--very superficial read of my work based mostly on the abstract.
I don’t know enough about academic publishing citation to evaluate the ease of that kind of checking, but my understanding is that the frankencitations are leaving a credible digital trail in their wake and that kind of automated detection might not be so straightforward and will keep getting harder.
That bigger problem is the thing I’m most wondering about and don’t really understand. It suggests a lot of performative B.S. around academic publishing - which okay - but I wish the existence of a technology that can make convincing B.S. at scale prompted a reckoning around what’s valued in these spaces.
I use AI very much the way that you do, with a lot of iteration, back-and-forth, dictating, and heavily editing. When I finish with it, I consider it mine. We need to teach people to use it that way and not just hit publish.
I don't think disclosing it is the answer because people have knee-jerk reactions to AI, even when it's well written. I know you use AI because I've read some of your previous articles where you've talked about it. But I'll be honest with you, my AI radar doesn't go off when I read your stuff the way it does with many other people who don't edit the output and simply provide it as is.
If the final result still sounds like the person who wrote it, isn't that the whole point?
As for the winners of literary competitions, that's a whole other situation, and I don't know how they're going to deal with that. What exactly are we judging at that point: the writing, the editing, the prompting, or some combination of all three?
In the end, I don't think any detector is going to be foolproof. The bigger question is what constitutes AI-generated content. Is the way you write AI-generated? Is the way I use it AI-generated? How do we determine authorship? At some point, almost everybody is going to be using AI to help them write in some capacity.
Thanks, Betty, that’s flattering. I agree with almost all of this. When I go back to some of my earlier AI-assisted writing though, I can see plenty of things I would handle differently now, including little phrases and rhythms that would annoy me today. The funny part is that the strongest AI radar often belongs to people who use these tools a lot themselves :)
In any case, my guess is that all this becomes much less interesting in a couple of years because the incentives all point toward widespread use. What I’m trying to do is speed up the awkward transition, including less moral panic and better social norms.
It seems like "moral panic" is a pretty reasonable response to people not writing for themselves, which seems like a fairly likely outcome.
We outsource most things to technology at this point, and that's ok.
You seem to be one of those people who lamented the loss of handwriting skills when the first typewriters came out.
Thanks for the hostile response. Regardless, you don't think there is any trade off at all here?
Sorry, I didn't mean to be hostile at all. I've written plenty about the problem of skill atrophy in my AI series and the struggles in teaching. But I certainly don't think it warrants indescriminate ai policing.
Nuanced and well reasoned as usual, sir.
I think I'm still less of a pessimist on disclosure and *more* of a pessimist on detectors, but I recognize that both are moving targets with sometimes pernicious incentives baked in.
I hope this landscape levels out before the decade is over.
Thanks, Quinn!
On the topic other commenters are on, it's a balance regarding grasping something you are meant to understand versus getting the one-sentence summary.
This is nothing new, it's a 20 year old issue that started with the internet's mass-adoption and Sparknotes.
I completely agree with the piece and that we obviously have things that AI are going to help us optimize, probably many creative things.
And that is good. Better is good.
Alexander, this is an exceptionally clear and nuanced take — one of the best I’ve read on the topic.
You’re absolutely right: the real danger lies not in the tool itself, but in how quickly we’ve allowed "provenance" to replace substance. When a probabilistic “guess” from a black-box algorithm starts functioning as a status signal and moral verdict, we enter dangerous territory. What begins as a helpful detector quietly mutates into a new form of grammar nazism — or, as I call it, the architecture of accusation.
The moment we start judging writing by how “human” it appears to a machine rather than by its actual quality and insight, we lose something essential about intellectual life.
I explored this exact mechanism — how “probably” becomes a verdict and how fear of AI is being institutionalized — in my recent essay: “Probably. How a single hedging word became the most powerful tool of institutional accusation in modern intellectual life — and what it is actually built to do.”
https://olegmaltsev.substack.com/p/probably-the-most-dangerous-word-in-intellectual-life
Your piece and mine complement each other well: you diagnose the cultural shift, while I try to map the deeper systemic and philosophical architecture behind it.
Thank you for writing it.
Thanks, Oleg!
I’m still not that convinced by Pangram - I test it with Claude-generated text and it frequently comes back as 100% human when I’ve done no editing. On occasion it also identified 100% AI or a high score - but not at all consistently as far as I can tell. I find this sudden trusting Pangram on everything shift really odd and just a little suss.
It's interesting: I think I agree with you in principle about pretty much everything here. But I recently got a paid subscription to Pangram and as a creative writer, I can't stop myself from checking text when I suspect it's AI -- or at least I have found myself reaching for the detector when there was no actual need for me to discover the provenance of the text. As a result, I think I've gotten better at spotting AI text. I'm at least better at predicting when Pangram will give it a high AI score, anyway. This kind of suggests another, wider social purpose for AI detectors: to train people to notice and interrogate the patterns AI produces.
I wrote a long essay about the Shy Girl controversy and in there I talked about how "dissolving distinctions disadvantages difficulty"... In other words, when we don't know or don't care how something is produced, the easiest process wins. But knowing and caring how something is produced has historically been considered a valuable skill. Taste, discernment, whatever you want to call it. Imo we can value AI text for what it is but also become more skilled at seeing what it is not. Does that make sense?
I swing hard from the measured and thoughtful positions you’ve laid out here to wild eyed fury that stick swingers like Sam Kriss and Paul Kingsnorth inhabit. There’s nothing wrong with supermarket whole meal bread, it’s often fortified with cool things like niacin and riboflavin – it’s fine for food at an academic conference.
But really, what I want is caramelised crust that my teeth have to crash through to eat the wild fermented sourdough of human expression.
Too romantic?
Enjoyed your post and agree with the general thrust of it.
One point I'd like to complicate is 'attention is scarce'. Another reading is that readers have correctly determined that much credentialed output is not particularly good, nor worth their time. Attention scarcity as verdict, not consequence.
You frame detection as an imperfect shortcut competing with others, but the incumbent shortcuts lost credibility because they substituted for reading. Pangram extends that problem. AI-detection anxiety is then, at least partly, due to the collapse of venue as a quality signal, and enforcement of provenance norms is the old gatekeepers trying to re-establish the legitimacy of their proxies.
The question never asked is 'is this good?'
So called accuracy is a red herring. AI detectors have no place in education. https://www.tandfonline.com/doi/full/10.1080/1360080X.2026.2622146
I agree. I have a similar writing process to yours. It's multiple drafts and back and forths. And I appreciate being able to work through ideas through voice chat and not having to type all the time. It's interesting- I work in an industry where the information just needs to get across, so we often don't use a lot of dressing and style. I suspect that what people are so upset about is that the style or padding is not a natural way of speaking and so using a "translator" is somehow offensive. As long as nothing was hallucinated, then the arguments and thoughts are originally the author's but translated to proper grammar and academic jargon, if necessary. Just my two cents. I am curious- how are you using codex for this?
Well put. There is a continuous spectrum here which is (unsurprisingly) represented in much of 'life, the universe and everything'. It is time to discard the 'AI is evil' first call and allow that AI as a label/category/framework can be something very nuanced and is often beneficial well beyond the individual when used in balanced and appropriate ways. Our problem is that we are still thinking in the 'wild west' phase, even where pockets of expertise exist. It will take a while as the pendulum moves and our shared patience is important. I feel for those who are labelled harshly or 'failed' for what in the future will be normalised and value-adding. We still have a lot to learn in this regard.