On how much text? There's a fair few reports that it's quite inaccurate when given smaller samples of text, and if you go cherry-picking the odds will likely go up even further.
My neighbours regularly remind me that AI is quite different in French - despite being Francophone they tend to use it in English for what they think are better results.
It's entirely plausible detection is totally different in different languages.
Note 'Balance ton Claude' is an oblique reference to 'Balance ton porc', what's commonly considered the French equivalent of the US #MeToo cultural event
People like to admire the author, to see a better version of themselves in the author.... If they find out that image in their head is a lie, they get angy
Not comparable if what was advertised performs significantly differently from 65hp (assuming beef and horse meat taste similar). And if AI-assisted writing is really at the point where it can honestly win prizes, then there's no problem here, unless the prizes in question are specifically for "wholly human-created works". No need to tell furniture shoppers if the screws in the tables were driven by traditional screwdriver or a machine-powered one.
Cat or human meat, or even soy would perform the same as beef, so why bother right? Why label things? Why bother about knowing what you consume?
No need to tell furniture shoppers if the screws in the tables were driven by traditional screwdriver or a machine-powered one.
Why do you think hand made furnitures cost 10x more than ikea then? "BuT ItS JuSt a FuRnItUrE", I swear you slop enjoyers are either completely devoid of critical thinking or don't know anything other than bad faith arguments.
I see you building that straw man but I'll address it anyway. Did you notice my mention of taste? Cat meat is unlikely to taste like beef because they are carnivores, while cows (and horses) are herbivores. Similarly, humans are omnivores (and, tangentially, we - along with most other species - have the evolved instinct not to cannibalize each other as that's a road to extinction). I would expect cows, horses, bison, donkeys, etc to offer relatively similar eating experience and nutritional value, environmental conditions considered equal. I won't bother with soy as you should now be able to extract and apply the underlying logic yourself.
And on furniture you're going even further out from the context; you must have a lot of straw at hand. Hand-made furniture primarily costs a lot more than machine-made furniture because of the difference in human effort. Secondarily because machine-made furniture tend to look similar, and humans usually place high value on uniqueness. But, as before, you're off in the weeds building straw effigies because I wasn't talking about machine-made furniture at all. I was talking about a carpenter, etc using a traditional screwdriver (you know, the ones you have to turn the wrist to drive) vs a powered screwdriver (press a button to make it go). And the outcome of using either being indistinguishable, similar to how award was given not because the author used AI or not to write the novel (that hasn't been settled AFAIK), but because the content itself passed the award-worthy bar.
Agreed. It’s not like a particular word is any different when it’s made by ai.
Would we care if someone used a typewriter vs a computer to write it.
This looks like a weird snobbishness. The fact that it’s by Europeans makes it more funny since they have been left out of the AI race.
In the end you have a final product. Can critique it on its own. He used ai to help write it and you couldn’t tell before, says more about the product and perhaps you don’t like what it says about you .
An account that has only existed for 4 months, whose only source for whenever something is plagiarised is a single company whose bread and butter is telling everybody what they want to hear, where most of the tweets barely gets a hundred likes should not be getting the level of coverage it has right now.
I guess you can pick between that and the propaganda provided by the AI peddlers, they surely have nothing to gain telling everybody what they want to hear
He is also a finalist for the most prestigious annual French language awards, the Goncourt and the Renaudot, set to be issued soon.
If I had a book that was hoping to win this award, anonymous accusations like this seem like a pretty easy way to shift the odds. It's impossible for the author to disprove the allegations and the award committee will want to 'avoid controversy'
He's been revealed as a flagrant plagiarist [1] and there's a lot of evidence that his writing before 2023 had very low pangram scores, and the other authors nominated did not have high pangram scores.
This isn't the first high profile case of a black person being reported with accusations of using AI this year[^1][^2]. Once or twice, sure, whatever. But a third time now? It's hard to think this isn't anything but targeted, especially given this particular case the group making the "claims" are entirely anonymous.
Is it that, or is it because marketing "content" by a minority author with an "interesting life story" is an industry not an art, and they're less likely to care how the "content" was produced?
Pangram is making the "claims", not some group. It is known to have a very low false positive rate, especially on long texts.
And there is now also evidence of plagiarism, rather than just AI writing, in his other work (see down thread).
The longer a text, the more entropy, the easier it is to classify. Novels are pretty long. So I suspect the false positive rate for books is far lower than for articles. And, as I said, we also have corroborating evidence now in form of cases of plagiarism discovered in his past work. Someone who has no qualms about committing plagiarism is more likely to also not have inhibitions about using LLMs to write books.
“It was just a coincidence that most American slaves were black”
I don’t think anyone is saying skin color conveys immunity from accusation. But I hope you can see that the odds that a dramatically higher per capita rate of accusations against people with a particular skin color should at least give pause.
How many non black people have been accused though?
Why would they target these specific 3 black authors? I assume there are hundreds/thousands of successful black authors, 3 "this year" seems very small to call it a targeted attack
Depends on how you define «they», but the public discourse in Norway the past two weeks have been defined by the Norwegian news outlet Manifest Media publishing a series of articles about AI “meat proxying”, starting whith an analysis of commentators employed by the Norwegian Broadcasting Corporation (NRK). The one found guilty was a white woman, and among the four non-guilty was one non-white man: https://www.manifestmedia.no/artikler/slik-testet-vi-nrk-for... A follow-up article targetting (white male) professor Asle Toje found that he had probably used AI to write a much-discussed opinion piece: https://www.manifestmedia.no/artikler/pangram-tojes-omstridt...
Race has not been a part of the public conversation. Not sure if it would have been that way if the guilty person was non-white (but I'm slightly optimistic, given that the first person found guilty was a woman and gender was not part of the discussion).
It's hard to think this isn't anything but targeted
By whom?
A simpler (and easy to think of) explanation is normal systemic racism. Black writers are likely underrepresented in the training data begin with, so the models assume Black writing can't be real.
How would the model know the author is black? People have also looked at pangram at other Haitian authors in French and it doesn't give them high AI scores.
It is possible that the works of black people are tested more, possibly by racist people systemically putting them through Pangram in hope of getting a hit.
If that's the case, unfortunately, they're bound to find legitimate fraud eventually. Given the subsequent hard evidence of plagiarism, it seems like this is one such case, and we can't exactly ignore it.
I remain convinced the whole notion of AI-detection algorithms with any semblance of precision and accuracy is a con.
Critics say the author almost exclusively relied on AI to create his work, based on Pangram’s findings.
Pangram says it “is trained specifically to detect the latest AI models from frontier labs like OpenAI, Anthropic, Google, Meta, DeepSeek, xAI, and more. [They] benchmark against 26 models (so far)…”.
I can probably list more than 26 models off the top of my head. Granted, some models have been distilled against others and will be “detectable” but that’s NOWHERE near a large enough sample size to be reliable in 2026.
You or I could get a flag by that detector in our first attempt, and it wouldn’t take long for a false positive against enough attempts of truly original thought either.
The only thing I’d actually believe is things like Claude’s text watermarking system. However, those systems will never be present for EVERY model, so reliable blanket detection remains an impossibility.
The entropy in written text of meaningful length is simply too great for detection to be reliably possible in 2026, especially considering the changes the author can make.
"Experimenting, I’ve also found that I can pretty reliably get Pangram to declare large sections of text to be AI-written or human-written depending on the larger textual context I place those sections in - that is, I can take AI-generated text and embed it in human-generated text and have Pangram declare it 100% human, and I can take human-generated text and embed it in AI-generated text and have Pangram declare it 100% AI. I’ve also found that it’s pretty easy to induce false positives, that is, to write texts myself that Pangram identifies as AI-written. All of this makes me hesitant about the Pangram tool, and in particular the way many people use it, as a one-shot gotcha machine."
Sounds like it may be very easy for someone motivated enough to find sections, even longer sections, that turn out as 100% AI written.
Sounds like it may be very easy for someone motivated enough to find sections, even longer sections, that turn out as 100% AI written.
It's probably as easy as prompting Claude, "Take text from this markdown/PDF file and run it through Pangram tool, figure out a way to, and find, the longest passage from the text that Pangram flags as 100% AI generated", and leaving it spinning in Auto mode for an hour.
The impact from a model not being included in the training set is increased false negatives for text generated by that model, not increased false positives for text written by humans.
The author behind the twitter account has been discovered [1] (source in french and behind paywall). It's team leaded by a french essayist and journalist at "Le Figaro" named Samuel Fitoussi.
There's important extra context on this case in that the author has been discovered to be a flagrant plagiarist. He won an award in 2014 for a work which was almost entirely ripped from a French novel in the 60s.
This was discussed a lot on the news here. Couple of authors feed their pre AI books to pangram and also got false positive. Maybe it is AI, maybe not. But using AI to accuse someone else of using AI is ironic.
Can you post some links? I know some authors said that, but had used a different tool than Pangram, and I don't expect non-technical people to understand the difference. Pangram is reputedly the best of them and personally I couldn't get a single false positive on my own work (both French and English).
* Media ran the entire novel through Pangram and got a 94% AI score [1], whereas all other Goncourt finalists, Haitian/African novels and French classics they tested came out as 100% human. I recall they also tested works specifically from before and after the AI boom and found a shift, but I don't remember where I saw that.
* People on social media, as well as traditional media, have then found cases of blatant plagiarism in Orelien's work a decade ago (this includes one prized short story and several posts to blogs and media) [2]
I think we have to be careful not to assume that tools like Pangram are completely accurate and the word of God, but we can't dismiss them outright, and this particular author's history of plagiarism makes it much harder to give him the benefit of doubt.
It was relevant in this context because part of Grasset's defence (Grasset being the publisher) was that the Bible and other classics were sometimes flagged as AI by these tools and therefore they are unreliable. So they actually tested it with the same tool and found that it wasn't an issue.
Comparing to other Haitian authors was also relevant, because another defence from the author was that his "caribbean style" could have tripped up the detector.
But yes, other than that, that's a good point. My anecdotal evidence is that my own unpublished writing, both in French and English, contains no false positives, but even in that case, that's only evidence that my style isn't AI-adjacent. I'd be curious to see concrete examples of false positives, but it's hard to fully trust these if they don't have proven timestamps.
Any Francophones, does the book sound like AI? I find that in English a lot of the time that something is accused of being AI-written and people cast aspersions on the quality of AI detectors, that when you read the text it’s immediately obvious as AI: chock full of em-dashes, not-x-but-y constructions, unnatural metaphors, and so on.
The first two chapters can be read online. The text is full of metaphors (many nonsensical), staccato sentences, and X not Y constructions. The rule of three comes up every other sentence, often combined with a zeugma. No em-dashes though!
Pangram 100% is notoriously difficult for humans to do, even with intent.
On how much text? There's a fair few reports that it's quite inaccurate when given smaller samples of text, and if you go cherry-picking the odds will likely go up even further.
It's in the article - looks like they ran it on a book.
No, they say they ran "passages" of the book through it.
My neighbours regularly remind me that AI is quite different in French - despite being Francophone they tend to use it in English for what they think are better results.
It's entirely plausible detection is totally different in different languages.
Sounds familiar
Note 'Balance ton Claude' is an oblique reference to 'Balance ton porc', what's commonly considered the French equivalent of the US #MeToo cultural event
What's the big deal? If the book is good, it is an excellent use of AI.
People like to admire the author, to see a better version of themselves in the author.... If they find out that image in their head is a lie, they get angy
Well a lot of my favorite authors (and musicians) are terrible people, but I like to admire part of the author, yes.
I didn't have it in my head at the time but it's just made me think of "the salt path"
I agree.
I don't like the unnecessary defensiveness by the author to pangram proof.
Perhaps the same reason people get upset when a gold medalist is discovered using enhancers. The goal isn't just to publish a great story.
Us non pod-people feel betrayed when we’ve been lied to.
Feel free to read it then.
Like it or not, people expect significant AI-use to be disclosed. When only human is credited, they feel cheated.
Excellent use of AI? Disclose it and avoid the fuss.
For the few pages I saw on X it's actually tiresome to read.
Lots of short sentences. Lots of weird metaphors. Many infamous "it's not x, it's y".
Tell us so we can make informed decisions?
"Why do you care is there is horse meat in your beef burger patties if it tastes good" vibe
What's wrong with horse meat?
It's not beef. What if I put a 65hp engine in the next car you buy instead of whatever was advertised?
Not comparable if what was advertised performs significantly differently from 65hp (assuming beef and horse meat taste similar). And if AI-assisted writing is really at the point where it can honestly win prizes, then there's no problem here, unless the prizes in question are specifically for "wholly human-created works". No need to tell furniture shoppers if the screws in the tables were driven by traditional screwdriver or a machine-powered one.
Cat or human meat, or even soy would perform the same as beef, so why bother right? Why label things? Why bother about knowing what you consume?
Why do you think hand made furnitures cost 10x more than ikea then? "BuT ItS JuSt a FuRnItUrE", I swear you slop enjoyers are either completely devoid of critical thinking or don't know anything other than bad faith arguments.
I see you building that straw man but I'll address it anyway. Did you notice my mention of taste? Cat meat is unlikely to taste like beef because they are carnivores, while cows (and horses) are herbivores. Similarly, humans are omnivores (and, tangentially, we - along with most other species - have the evolved instinct not to cannibalize each other as that's a road to extinction). I would expect cows, horses, bison, donkeys, etc to offer relatively similar eating experience and nutritional value, environmental conditions considered equal. I won't bother with soy as you should now be able to extract and apply the underlying logic yourself.
And on furniture you're going even further out from the context; you must have a lot of straw at hand. Hand-made furniture primarily costs a lot more than machine-made furniture because of the difference in human effort. Secondarily because machine-made furniture tend to look similar, and humans usually place high value on uniqueness. But, as before, you're off in the weeds building straw effigies because I wasn't talking about machine-made furniture at all. I was talking about a carpenter, etc using a traditional screwdriver (you know, the ones you have to turn the wrist to drive) vs a powered screwdriver (press a button to make it go). And the outcome of using either being indistinguishable, similar to how award was given not because the author used AI or not to write the novel (that hasn't been settled AFAIK), but because the content itself passed the award-worthy bar.
The writing is pretty terrible and shlocky but the topic is what the committee is looking for.
Agreed. It’s not like a particular word is any different when it’s made by ai.
Would we care if someone used a typewriter vs a computer to write it.
This looks like a weird snobbishness. The fact that it’s by Europeans makes it more funny since they have been left out of the AI race.
In the end you have a final product. Can critique it on its own. He used ai to help write it and you couldn’t tell before, says more about the product and perhaps you don’t like what it says about you .
It's terrible writing to be honest. He's also Haitian-Canadian and not European.
The uproar is in France. Sounds like the bigger uproar could have been the shitty writing.
Then again. Europeans have ret-conned genericized names into terroir so it's not exactly surprising.
This story smells odd; why does anybody care about the accusations of an anonymous twitter account? A claim seemingly without evidence no less.
Because it's an anonymous twitter account dedicated to scanning and identifying AI usage in published works?
An account that has only existed for 4 months, whose only source for whenever something is plagiarised is a single company whose bread and butter is telling everybody what they want to hear, where most of the tweets barely gets a hundred likes should not be getting the level of coverage it has right now.
https://x.com/Pangramed
By all means test it out with some non-AI content and report back
I guess you can pick between that and the propaganda provided by the AI peddlers, they surely have nothing to gain telling everybody what they want to hear
If I had a book that was hoping to win this award, anonymous accusations like this seem like a pretty easy way to shift the odds. It's impossible for the author to disprove the allegations and the award committee will want to 'avoid controversy'
Well, he has been evicted from the Goncourt.
He's been revealed as a flagrant plagiarist [1] and there's a lot of evidence that his writing before 2023 had very low pangram scores, and the other authors nominated did not have high pangram scores.
[1] https://www.lapresse.ca/arts/litterature/2026-09-23/thelyson...
This isn't the first high profile case of a black person being reported with accusations of using AI this year[^1][^2]. Once or twice, sure, whatever. But a third time now? It's hard to think this isn't anything but targeted, especially given this particular case the group making the "claims" are entirely anonymous.
Anyway, here's the thread from twitter if people are interested: https://x.com/Pangramed/status/2102383484027171306?s=20
[^1]: https://www.independent.co.uk/arts-entertainment/books/featu...
[^2]: https://www.thetimes.com/uk/education/article/jason-arday-re...
Are the accusations untrue? The evidence that i have seen seems pretty compelling.
Is it that, or is it because marketing "content" by a minority author with an "interesting life story" is an industry not an art, and they're less likely to care how the "content" was produced?
Pangram is making the "claims", not some group. It is known to have a very low false positive rate, especially on long texts. And there is now also evidence of plagiarism, rather than just AI writing, in his other work (see down thread).
What’s the false positive rate multiplied by number of books in print? This seems like a fraught use of odds / stakes.
The longer a text, the more entropy, the easier it is to classify. Novels are pretty long. So I suspect the false positive rate for books is far lower than for articles. And, as I said, we also have corroborating evidence now in form of cases of plagiarism discovered in his past work. Someone who has no qualms about committing plagiarism is more likely to also not have inhibitions about using LLMs to write books.
They tested passages. It is absolutely trivial to find passages that pangram classifies as 100% ai gen even if it's not the case.
I don't think it's trivial. I'm also pretty sure many people have tested the whole book.
"He is black, you cannot accuse him of anything, you racist."
“It was just a coincidence that most American slaves were black”
I don’t think anyone is saying skin color conveys immunity from accusation. But I hope you can see that the odds that a dramatically higher per capita rate of accusations against people with a particular skin color should at least give pause.
Three black people in a year across the literal entire planet’s population and you are convinced it’s a racist conspiracy?
How many non black people have been accused though?
Why would they target these specific 3 black authors? I assume there are hundreds/thousands of successful black authors, 3 "this year" seems very small to call it a targeted attack
Depends on how you define «they», but the public discourse in Norway the past two weeks have been defined by the Norwegian news outlet Manifest Media publishing a series of articles about AI “meat proxying”, starting whith an analysis of commentators employed by the Norwegian Broadcasting Corporation (NRK). The one found guilty was a white woman, and among the four non-guilty was one non-white man: https://www.manifestmedia.no/artikler/slik-testet-vi-nrk-for... A follow-up article targetting (white male) professor Asle Toje found that he had probably used AI to write a much-discussed opinion piece: https://www.manifestmedia.no/artikler/pangram-tojes-omstridt...
Race has not been a part of the public conversation. Not sure if it would have been that way if the guilty person was non-white (but I'm slightly optimistic, given that the first person found guilty was a woman and gender was not part of the discussion).
Maybe the AI companies used human feedback from Haiti people to train their models?
bro just won hn like it was nothing im in awe
It's 2026. This defence is not working anymore.
It's hard to think this isn't anything but targeted
By whom?
A simpler (and easy to think of) explanation is normal systemic racism. Black writers are likely underrepresented in the training data begin with, so the models assume Black writing can't be real.
How would the model know the author is black? People have also looked at pangram at other Haitian authors in French and it doesn't give them high AI scores.
XXCancel: https://xxcancel.com/Pangramed and https://xxcancel.com/Pangramed/status/2102383484027171306?s=...
It is possible that the works of black people are tested more, possibly by racist people systemically putting them through Pangram in hope of getting a hit.
If that's the case, unfortunately, they're bound to find legitimate fraud eventually. Given the subsequent hard evidence of plagiarism, it seems like this is one such case, and we can't exactly ignore it.
People are now also finding cases of plagiarism (not AI writing) in his past published articles. https://ca.rollingstone.com/culture/thelyson-orelien-author-...
I remain convinced the whole notion of AI-detection algorithms with any semblance of precision and accuracy is a con.
Critics say the author almost exclusively relied on AI to create his work, based on Pangram’s findings.
Pangram says it “is trained specifically to detect the latest AI models from frontier labs like OpenAI, Anthropic, Google, Meta, DeepSeek, xAI, and more. [They] benchmark against 26 models (so far)…”.
I can probably list more than 26 models off the top of my head. Granted, some models have been distilled against others and will be “detectable” but that’s NOWHERE near a large enough sample size to be reliable in 2026.
You or I could get a flag by that detector in our first attempt, and it wouldn’t take long for a false positive against enough attempts of truly original thought either.
The only thing I’d actually believe is things like Claude’s text watermarking system. However, those systems will never be present for EVERY model, so reliable blanket detection remains an impossibility.
The entropy in written text of meaningful length is simply too great for detection to be reliably possible in 2026, especially considering the changes the author can make.
best blog post I know about this: https://freddiedeboer.substack.com/p/i-wouldnt-say-pangram-i...
"Experimenting, I’ve also found that I can pretty reliably get Pangram to declare large sections of text to be AI-written or human-written depending on the larger textual context I place those sections in - that is, I can take AI-generated text and embed it in human-generated text and have Pangram declare it 100% human, and I can take human-generated text and embed it in AI-generated text and have Pangram declare it 100% AI. I’ve also found that it’s pretty easy to induce false positives, that is, to write texts myself that Pangram identifies as AI-written. All of this makes me hesitant about the Pangram tool, and in particular the way many people use it, as a one-shot gotcha machine."
Sounds like it may be very easy for someone motivated enough to find sections, even longer sections, that turn out as 100% AI written.
It's probably as easy as prompting Claude, "Take text from this markdown/PDF file and run it through Pangram tool, figure out a way to, and find, the longest passage from the text that Pangram flags as 100% AI generated", and leaving it spinning in Auto mode for an hour.
Pangram is unreliable on short text, but much more reliable on long form text.
The impact from a model not being included in the training set is increased false negatives for text generated by that model, not increased false positives for text written by humans.
Source?
The author behind the twitter account has been discovered [1] (source in french and behind paywall). It's team leaded by a french essayist and journalist at "Le Figaro" named Samuel Fitoussi.
[1] https://www.lefigaro.fr/culture/qui-se-cache-derriere-le-com...
I love learning about an alleged "uproar" in my own country from a BBC link shared on HN.
There's important extra context on this case in that the author has been discovered to be a flagrant plagiarist. He won an award in 2014 for a work which was almost entirely ripped from a French novel in the 60s.
https://www.lapresse.ca/arts/litterature/2026-09-23/thelyson...
This was discussed a lot on the news here. Couple of authors feed their pre AI books to pangram and also got false positive. Maybe it is AI, maybe not. But using AI to accuse someone else of using AI is ironic.
Can you post some links? I know some authors said that, but had used a different tool than Pangram, and I don't expect non-technical people to understand the difference. Pangram is reputedly the best of them and personally I couldn't get a single false positive on my own work (both French and English).
To give a bit more information:
* Media ran the entire novel through Pangram and got a 94% AI score [1], whereas all other Goncourt finalists, Haitian/African novels and French classics they tested came out as 100% human. I recall they also tested works specifically from before and after the AI boom and found a shift, but I don't remember where I saw that.
* People on social media, as well as traditional media, have then found cases of blatant plagiarism in Orelien's work a decade ago (this includes one prized short story and several posts to blogs and media) [2]
I think we have to be careful not to assume that tools like Pangram are completely accurate and the word of God, but we can't dismiss them outright, and this particular author's history of plagiarism makes it much harder to give him the benefit of doubt.
[1] (French) https://ici.radio-canada.ca/rci/fr/nouvelle/2286476/thelyson...
[2] (French) https://www.lapresse.ca/arts/litterature/2026-09-23/thelyson...
Wouldn’t the French classics have been part of pangram’s training set? That doesn’t seem like strong evidence at all.
It was relevant in this context because part of Grasset's defence (Grasset being the publisher) was that the Bible and other classics were sometimes flagged as AI by these tools and therefore they are unreliable. So they actually tested it with the same tool and found that it wasn't an issue.
Comparing to other Haitian authors was also relevant, because another defence from the author was that his "caribbean style" could have tripped up the detector.
But yes, other than that, that's a good point. My anecdotal evidence is that my own unpublished writing, both in French and English, contains no false positives, but even in that case, that's only evidence that my style isn't AI-adjacent. I'd be curious to see concrete examples of false positives, but it's hard to fully trust these if they don't have proven timestamps.
Any Francophones, does the book sound like AI? I find that in English a lot of the time that something is accused of being AI-written and people cast aspersions on the quality of AI detectors, that when you read the text it’s immediately obvious as AI: chock full of em-dashes, not-x-but-y constructions, unnatural metaphors, and so on.
The first two chapters can be read online. The text is full of metaphors (many nonsensical), staccato sentences, and X not Y constructions. The rule of three comes up every other sentence, often combined with a zeugma. No em-dashes though!