On a personal level, I don't get all this hype of: "Was AI involved or not?"
I am a simple guy. All I want to know boils down to simple questions: Is the story good and the writing serves the story? Are the thoughts/ideas supported by the writing?
It might be a bigger problem in academia, where you have to evaluate if the student has done the thinking, understood the material.
But way back in the 1980s, I had a friend who earned quite some money by writing for his fellow students. So, that problem is not new.
Well, a while back I wrote a piece about the similarities between AI Slop and Kitsch. AI might become its own art form. But the provenance of an art piece is an intrinsic component of its emotional story. Would the Mona Lisa be as important today if its provenance would be not be from the mind of a human artist? I highly doubt it. It would still be the same art piece, just a lot less relevant.
Yes, and there is one more area where the classifier can fail. Suppose it succedes in tracking a genuine pattern from machine-to-text. Yet, a pattern by itself is of no use except if it can fit the conceptual organization of those watching it. If it doesn’t, we can infer anything at all from it. This relates to the opaqueness of how Pangram verdicts are reached… a test to know if two people mean the same thing by the same sentence is if their use is sensitive to the same evidence in favour and against it (its a old logical positivist test). Now we cannot possibly know if those detectors mean anything by their verdicts because they are sensitive to evidence that are completly opaque and irrelevant to us, like em-dashes.
We're only a few days in and the new pet ai has made a plethora of mistakes across all parts of the spectrum, from calling human work ai, to calling ai work human.
In other words it's useless at what it does, but potentially damaging and destructive.
Michael, this is an exceptionally sharp and honest piece.
Your experiment perfectly exposes one of the most dangerous aspects of Pangram: the "All-Clear" (100% human) verdict is, in many ways, more misleading than a cautious “probably AI.” When a tool confidently labels heavily AI-assisted text as purely human, it creates a false sense of security — which institutions then eagerly use as “proof” of authenticity.
This confirms what I’ve been arguing: we are not dealing with a reliable detector, but with a probabilistic instrument that is being treated as authoritative evidence. The real problem emerges when “probably” or “all clear” moves from the screen into real-world decisions about reputation, careers, and academic integrity.
I explored this systemic issue in depth 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."
You rightly focus on the technical and pedagogical failures. I go one step further: this is rapidly becoming an entire architecture of accusation, where fear is monetized and human judgment is outsourced to a black box.
Thank you for this important work. Highly recommended reading.
Michael, Jason from Pangram here. I am curious which of your blogs you ran Pangram on that came back Human. I checked your last nine and your top nine, and all of them come back either with significant AI or mixed, including this one.
I've already checked multiple articles across multiple authors and you DO call out ai as human and human as ai, and that carries potential real life ramifications.
Please, by all means, send the shareable links to whatever you think we are getting wrong, especially any false positives, which is a far bigger issue than missing some AI writing. Please also verify that no AI assistance was used in the process of editing those texts, since, as Michael will attest, when authors use AI to edit and write, even if it is their original idea, Pangram can detect it. Anecdotes of authenticity aren't really that helpful, since we cannot access the ground truth. The best practice is simply always to disclose AI writing.
Don't you think YOU should get it right in the first place, which 30 minutes on here will show you that you haven't, rather than falsely accusing people and then run around after you?
Reading just this site today your guess rate is far below the one you claim, but even if we took your patently absurd claims, at the rate the population writes you are literally falsely defaming hundreds of people.
That's terrible.
I'm in the final third of my life, so I'm not effected, or bothered, but for some young aspiring writer you may have just damaged their career for your own profit and greed.
There is no "post a link" that fixes that.
& why?
You know your software is flawed so what is its actual purpose, apart from to make you money by exploiting and cashing in on people's fears with a "solution" you already know doesn't work?
Is it possible you are running an old model? The offer still stands to give you access to an unlimited trial account to the current model I am running so that you can verify our accuracy.
I used whatever model was current on the day I published that piece. It is very easy to get around any version of Pangram. But the issue is that getting around Pangram makes the writing worse. There really is no point in using AI detection.
I cannot verify what you did or did not see, but it is in fact the case that all of your current posts are detected by Pangram as having AI as you admit, so I cannot verify the error you report. It's true that anyone attempting to bypass detection, with enough skill, might have some success, but when we detect something as having AI, we have a very high success rate with a very low false positive rate. That is something that many readers find valuable and worth it, including myself. As someone who has spent the last 25 years teaching English and Philosophy, I certainly find great epistemological and pedagogical value in knowing whether what I am reading has AI in it or not. Happy to explore the reasons why with you at some point.
If you do not believe me, that’s fine. I do not really care. But there are quite a few people who wrote about how Pangram is misidentifying their human writing as AI generated, which is even worse. If you think you can reliably identify AI written text, you are, quite frankly, completely misguided.
On a personal level, I don't get all this hype of: "Was AI involved or not?"
I am a simple guy. All I want to know boils down to simple questions: Is the story good and the writing serves the story? Are the thoughts/ideas supported by the writing?
It might be a bigger problem in academia, where you have to evaluate if the student has done the thinking, understood the material.
But way back in the 1980s, I had a friend who earned quite some money by writing for his fellow students. So, that problem is not new.
I think provenance is an art thing. I would argue that if you want real art, it needs to be primarily human made.
I disagree on that.
AI does nothing without human input.
The only question that remains is: Is it good? Does it touch me? Does it tell a story?
Well, a while back I wrote a piece about the similarities between AI Slop and Kitsch. AI might become its own art form. But the provenance of an art piece is an intrinsic component of its emotional story. Would the Mona Lisa be as important today if its provenance would be not be from the mind of a human artist? I highly doubt it. It would still be the same art piece, just a lot less relevant.
Yes, and there is one more area where the classifier can fail. Suppose it succedes in tracking a genuine pattern from machine-to-text. Yet, a pattern by itself is of no use except if it can fit the conceptual organization of those watching it. If it doesn’t, we can infer anything at all from it. This relates to the opaqueness of how Pangram verdicts are reached… a test to know if two people mean the same thing by the same sentence is if their use is sensitive to the same evidence in favour and against it (its a old logical positivist test). Now we cannot possibly know if those detectors mean anything by their verdicts because they are sensitive to evidence that are completly opaque and irrelevant to us, like em-dashes.
We're only a few days in and the new pet ai has made a plethora of mistakes across all parts of the spectrum, from calling human work ai, to calling ai work human.
In other words it's useless at what it does, but potentially damaging and destructive.
https://eliaswinter.substack.com/p/the-purity-machine?r=5f7ps0&utm_medium=ios
Michael, this is an exceptionally sharp and honest piece.
Your experiment perfectly exposes one of the most dangerous aspects of Pangram: the "All-Clear" (100% human) verdict is, in many ways, more misleading than a cautious “probably AI.” When a tool confidently labels heavily AI-assisted text as purely human, it creates a false sense of security — which institutions then eagerly use as “proof” of authenticity.
This confirms what I’ve been arguing: we are not dealing with a reliable detector, but with a probabilistic instrument that is being treated as authoritative evidence. The real problem emerges when “probably” or “all clear” moves from the screen into real-world decisions about reputation, careers, and academic integrity.
I explored this systemic issue in depth 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
You rightly focus on the technical and pedagogical failures. I go one step further: this is rapidly becoming an entire architecture of accusation, where fear is monetized and human judgment is outsourced to a black box.
Thank you for this important work. Highly recommended reading.
Michael, Jason from Pangram here. I am curious which of your blogs you ran Pangram on that came back Human. I checked your last nine and your top nine, and all of them come back either with significant AI or mixed, including this one.
I've already checked multiple articles across multiple authors and you DO call out ai as human and human as ai, and that carries potential real life ramifications.
You are potentially a lot ore toxic than ai slop.
Please, by all means, send the shareable links to whatever you think we are getting wrong, especially any false positives, which is a far bigger issue than missing some AI writing. Please also verify that no AI assistance was used in the process of editing those texts, since, as Michael will attest, when authors use AI to edit and write, even if it is their original idea, Pangram can detect it. Anecdotes of authenticity aren't really that helpful, since we cannot access the ground truth. The best practice is simply always to disclose AI writing.
Don't you think that is a bit back to front?
Don't you think YOU should get it right in the first place, which 30 minutes on here will show you that you haven't, rather than falsely accusing people and then run around after you?
Reading just this site today your guess rate is far below the one you claim, but even if we took your patently absurd claims, at the rate the population writes you are literally falsely defaming hundreds of people.
That's terrible.
I'm in the final third of my life, so I'm not effected, or bothered, but for some young aspiring writer you may have just damaged their career for your own profit and greed.
There is no "post a link" that fixes that.
& why?
You know your software is flawed so what is its actual purpose, apart from to make you money by exploiting and cashing in on people's fears with a "solution" you already know doesn't work?
Is it possible you are running an old model? The offer still stands to give you access to an unlimited trial account to the current model I am running so that you can verify our accuracy.
I used whatever model was current on the day I published that piece. It is very easy to get around any version of Pangram. But the issue is that getting around Pangram makes the writing worse. There really is no point in using AI detection.
I cannot verify what you did or did not see, but it is in fact the case that all of your current posts are detected by Pangram as having AI as you admit, so I cannot verify the error you report. It's true that anyone attempting to bypass detection, with enough skill, might have some success, but when we detect something as having AI, we have a very high success rate with a very low false positive rate. That is something that many readers find valuable and worth it, including myself. As someone who has spent the last 25 years teaching English and Philosophy, I certainly find great epistemological and pedagogical value in knowing whether what I am reading has AI in it or not. Happy to explore the reasons why with you at some point.
Jason
If you do not believe me, that’s fine. I do not really care. But there are quite a few people who wrote about how Pangram is misidentifying their human writing as AI generated, which is even worse. If you think you can reliably identify AI written text, you are, quite frankly, completely misguided.