Can you tell whether the #ideas in a piece of writing came from a human or an AI?
As AI writing becomes common, policies on AI use focus on who came up with the ideas in the text rather than who wrote it.
AI detectors answer the latter, not the former.
We release #IdeaLens, a detector of AI ideas.
As AI writing becomes common, policies on AI use focus on who came up with the ideas in the text rather than who wrote it.
AI detectors answer the latter, not the former.
We release #IdeaLens, a detector of AI ideas.
October 6, 2026 at 5:18 PM
For a limited time, we are releasing a demo of #IdeaLens and #ProseLens, so try it out!
ideadetector.ai
ideadetector.ai
IdeaLens: is this text human or AI?
We check a text's ideas and prose separately.
ideadetector.ai
October 6, 2026 at 5:18 PM
For a limited time, we are releasing a demo of #IdeaLens and #ProseLens, so try it out!
ideadetector.ai
ideadetector.ai
Here's the paper that accompanies the demo. The researchers find that IdeaLens can fairly accurately flag (when Pangram can't) stories that are written by people based on "AI-generated plans."
IdeaLens: Detecting AI Ideas in Long-form Writing
While modern AI detectors identify who wrote the words, emerging policies on AI use increasingly hinge on a different question: who came up with the ideas? We introduce IdeaLens, a detector that ident...
arxiv.org
October 6, 2026 at 4:58 PM
Here's the paper that accompanies the demo. The researchers find that IdeaLens can fairly accurately flag (when Pangram can't) stories that are written by people based on "AI-generated plans."
IdeaLens can detect varying levels of human input that went into the writing.
For documents generated from a very detailed human outline, IdeaLens flags only 6.8% as AI, compared to ProseLens's 86.5% and Pangram's 92.4%.
For documents generated from a very detailed human outline, IdeaLens flags only 6.8% as AI, compared to ProseLens's 86.5% and Pangram's 92.4%.
October 6, 2026 at 5:18 PM
IdeaLens can detect varying levels of human input that went into the writing.
For documents generated from a very detailed human outline, IdeaLens flags only 6.8% as AI, compared to ProseLens's 86.5% and Pangram's 92.4%.
For documents generated from a very detailed human outline, IdeaLens flags only 6.8% as AI, compared to ProseLens's 86.5% and Pangram's 92.4%.
This is a really fascinating research demo, from the lab of Mohit Iyyer at the University of Maryland: It predicts the likelihood that the ideas contained in a piece of text are AI-generated versus human-generated. A couple of examples in this thread.
IdeaLens: is this text human or AI?
We check a text's ideas and prose separately.
ideadetector.ai
October 6, 2026 at 4:58 PM
This is a really fascinating research demo, from the lab of Mohit Iyyer at the University of Maryland: It predicts the likelihood that the ideas contained in a piece of text are AI-generated versus human-generated. A couple of examples in this thread.
IdeaLens can also successfully detect human writing based on AI ideas.
October 6, 2026 at 5:18 PM
IdeaLens can also successfully detect human writing based on AI ideas.
There's some crazy stuff in the ablations. The order of elements has almost no significance. The "discourse roles" they infer? Not load-bearing. You can remove them. On the other hand, it's possible to get AUC 0.913 on the discourse roles alone.
October 7, 2026 at 4:50 PM
There's some crazy stuff in the ablations. The order of elements has almost no significance. The "discourse roles" they infer? Not load-bearing. You can remove them. On the other hand, it's possible to get AUC 0.913 on the discourse roles alone.
On conventional AI detection benchmarks, where prose and ideas come from the same source, IdeaLens correctly flags 91.1% as AI at 0.6% FPR, outperforming all open-source detectors.
ProseLens does even better, flagging 98.9% at 0.7% FPR.
ProseLens does even better, flagging 98.9% at 0.7% FPR.
October 6, 2026 at 5:18 PM
On conventional AI detection benchmarks, where prose and ideas come from the same source, IdeaLens correctly flags 91.1% as AI at 0.6% FPR, outperforming all open-source detectors.
ProseLens does even better, flagging 98.9% at 0.7% FPR.
ProseLens does even better, flagging 98.9% at 0.7% FPR.
IdeaLens can also detect AI ideas in languages other than English despite never seeing them during training.
This also works in low-resource languages such as Marathi or Khmer.
The model flags 94- 96% of AI-ideated documents & only 0.6-0.8% of AI-written documents.
This also works in low-resource languages such as Marathi or Khmer.
The model flags 94- 96% of AI-ideated documents & only 0.6-0.8% of AI-written documents.
October 6, 2026 at 5:18 PM
IdeaLens can also detect AI ideas in languages other than English despite never seeing them during training.
This also works in low-resource languages such as Marathi or Khmer.
The model flags 94- 96% of AI-ideated documents & only 0.6-0.8% of AI-written documents.
This also works in low-resource languages such as Marathi or Khmer.
The model flags 94- 96% of AI-ideated documents & only 0.6-0.8% of AI-written documents.
I checked an essay my dad wrote and had ChatGPT significantly edit. Pangram calls it 100% AI or AI-assisted. IdeaLens calls the prose 98% likely to have been AI-generated but gives the ideas just a 19% chance of having been AI-generated.
October 6, 2026 at 4:58 PM
I checked an essay my dad wrote and had ChatGPT significantly edit. Pangram calls it 100% AI or AI-assisted. IdeaLens calls the prose 98% likely to have been AI-generated but gives the ideas just a 19% chance of having been AI-generated.
Det har gått ett helt år sedan jag skrev den här korta texten om Gaza som en mer definitiv slutpunkt för efterkrigstiden. Blir glad om ni vill läsa idag. esbati.blogspot.com/2024/04/gaza...
Gaza och efterkrigstidens slut
Publicerat i norsk översättning Klassekampen den 2 april. Illustrationen som syns på fotot: Knut Løvås. Idealens kollaps I Gaza angrips ef...
esbati.blogspot.com
April 3, 2025 at 9:41 AM
Det har gått ett helt år sedan jag skrev den här korta texten om Gaza som en mer definitiv slutpunkt för efterkrigstiden. Blir glad om ni vill läsa idag. esbati.blogspot.com/2024/04/gaza...
and of course you get all the usual:
Paper: arxiv.org/abs/2610.06778
Models & Datasets: hf.co/collections/...
Github: github.com/RishanthRaje...
PyPi: pypi.org/project/idea...
+ You can test both detectors on Tinker API for as low as $1 for 1K documents.
Paper: arxiv.org/abs/2610.06778
Models & Datasets: hf.co/collections/...
Github: github.com/RishanthRaje...
PyPi: pypi.org/project/idea...
+ You can test both detectors on Tinker API for as low as $1 for 1K documents.
IdeaLens: Detecting AI Ideas in Long-form Writing
While modern AI detectors identify who wrote the words, emerging policies on AI use increasingly hinge on a different question: who came up with the ideas? We introduce IdeaLens, a detector that ident...
arxiv.org
October 6, 2026 at 5:18 PM
and of course you get all the usual:
Paper: arxiv.org/abs/2610.06778
Models & Datasets: hf.co/collections/...
Github: github.com/RishanthRaje...
PyPi: pypi.org/project/idea...
+ You can test both detectors on Tinker API for as low as $1 for 1K documents.
Paper: arxiv.org/abs/2610.06778
Models & Datasets: hf.co/collections/...
Github: github.com/RishanthRaje...
PyPi: pypi.org/project/idea...
+ You can test both detectors on Tinker API for as low as $1 for 1K documents.
We represent documents as outlines of their ideas and train #IdeaLens to fit Pangram 3 labels after paraphrasing the outlines to remove stylistic cues from the source.
As a control, we train #ProseLens directly on the raw documents.
As a control, we train #ProseLens directly on the raw documents.
October 6, 2026 at 5:18 PM
We represent documents as outlines of their ideas and train #IdeaLens to fit Pangram 3 labels after paraphrasing the outlines to remove stylistic cues from the source.
As a control, we train #ProseLens directly on the raw documents.
As a control, we train #ProseLens directly on the raw documents.
IdeaLens, l'outil expérimental qui détecte si le fond d'un texte a été généré par IA. Premier essai : concluant.
ideadetector.ai. le preprint du projet : arxiv.org/pdf/2610.067...
ideadetector.ai. le preprint du projet : arxiv.org/pdf/2610.067...
October 8, 2026 at 6:44 AM
IdeaLens, l'outil expérimental qui détecte si le fond d'un texte a été généré par IA. Premier essai : concluant.
ideadetector.ai. le preprint du projet : arxiv.org/pdf/2610.067...
ideadetector.ai. le preprint du projet : arxiv.org/pdf/2610.067...
The award-winning Commonwealth Prize story in Granta that Pangram flags as 100% AI comes out in IdeaLens as having a more than 99% chance of being AI for both prose and ideas. granta.com/the-serpent-...
The Serpent in the Grove
‘The ring of stone lay there where cocoa gave to bush, its mouth boarded with ply and chance.’ Fiction by Jamir Nazir.
granta.com
October 6, 2026 at 4:58 PM
The award-winning Commonwealth Prize story in Granta that Pangram flags as 100% AI comes out in IdeaLens as having a more than 99% chance of being AI for both prose and ideas. granta.com/the-serpent-...
The critic @roncharles.bsky.social put AI-generated material based on an interview transcript in his Substack (deliberately, to make a point). The AI detector Pangram flagged it as 77% AI. IdeaLens says the prose has a 60% chance of being AI but the ideas 14%. roncharles.substack.com/p/will-human...
October 6, 2026 at 4:58 PM
The critic @roncharles.bsky.social put AI-generated material based on an interview transcript in his Substack (deliberately, to make a point). The AI detector Pangram flagged it as 77% AI. IdeaLens says the prose has a 60% chance of being AI but the ideas 14%. roncharles.substack.com/p/will-human...
Jag skrev i Klassekampen igår om Gaza, vad det gör med efterkrigstidens ideal och hur det skakar till när palestinska liv värderas till så nära noll, att tio, femton tusen staplade barnlik inte når upp till entydiga fördömanden. Här på svenska. Läs gärna! esbati.blogspot.com/2024/04/gaza...
Gaza och efterkrigstidens slut
Publicerat i norsk översättning Klassekampen den 2 april. Illustrationen som syns på fotot: Knut Løvås. Idealens kollaps I Gaza angrips ef...
esbati.blogspot.com
April 3, 2024 at 7:53 AM
Jag skrev i Klassekampen igår om Gaza, vad det gör med efterkrigstidens ideal och hur det skakar till när palestinska liv värderas till så nära noll, att tio, femton tusen staplade barnlik inte når upp till entydiga fördömanden. Här på svenska. Läs gärna! esbati.blogspot.com/2024/04/gaza...
Rishanth Rajendhran, Minjoon Choi, Jenna Russell, Ramya Namuduri, Deniz B\"ol\"oni-Turgut, Marzena Karpinska, John Wieting, Mohit Iyyer: IdeaLens: Detecting AI Ideas in Long-form Writing https://arxiv.org/abs/2610.06778 https://arxiv.org/pdf/2610.06778 https://arxiv.org/html/2610.06778
October 6, 2026 at 6:41 AM
Rishanth Rajendhran, Minjoon Choi, Jenna Russell, Ramya Namuduri, Deniz B\"ol\"oni-Turgut, Marzena Karpinska, John Wieting, Mohit Iyyer: IdeaLens: Detecting AI Ideas in Long-form Writing https://arxiv.org/abs/2610.06778 https://arxiv.org/pdf/2610.06778 https://arxiv.org/html/2610.06778
De är uppfyllda av de liberala idealens kraft och stillsamma inre protest.
February 13, 2025 at 3:52 PM
De är uppfyllda av de liberala idealens kraft och stillsamma inre protest.
Skrev detta efter Mishras essä, med reflektioner om folkmordet som en katalysator för fördjupad och existentiellt omskakande rasism också här på hemmaplan esbati.blogspot.com/2024/04/gaza...
Gaza och efterkrigstidens slut
Publicerat i norsk översättning Klassekampen den 2 april. Illustrationen som syns på fotot: Knut Løvås. Idealens kollaps I Gaza angrips ef...
esbati.blogspot.com
March 26, 2025 at 1:06 PM
Skrev detta efter Mishras essä, med reflektioner om folkmordet som en katalysator för fördjupad och existentiellt omskakande rasism också här på hemmaplan esbati.blogspot.com/2024/04/gaza...
@isobelhk.bsky.social @astrojoachim Kul. En fundering: till och med astronauter på väg till månen kände sig osäkra på om de räckte till. Det skrattar vi åt och går vidare. Men om flickor och kvinnor tvivlar på sitt utseende, då är det osunda idealens fel, och alla måste tänka på hur de uttrycker […]
Original post on mastodon.green
mastodon.green
August 22, 2025 at 1:54 PM
@isobelhk.bsky.social @astrojoachim Kul. En fundering: till och med astronauter på väg till månen kände sig osäkra på om de räckte till. Det skrattar vi åt och går vidare. Men om flickor och kvinnor tvivlar på sitt utseende, då är det osunda idealens fel, och alla måste tänka på hur de uttrycker […]
*social rörlighet, menade egentligen idéernas, idealens och föreställningarnas rörlighet. Att värden som tett sig som så självklara och uppenbara snabbt kan överges till förmån för andra, objektivt osunda idéer. Man kommer ju aldrig säga "hur kunde det hända" igen.
December 20, 2025 at 12:00 PM
*social rörlighet, menade egentligen idéernas, idealens och föreställningarnas rörlighet. Att värden som tett sig som så självklara och uppenbara snabbt kan överges till förmån för andra, objektivt osunda idéer. Man kommer ju aldrig säga "hur kunde det hända" igen.
