People expect a binary response, is it AI generated yes or no. But it's more complicated than that. For example, if you see an emdash, it's probably AI generated. But it can also mean the author used it for fixing grammar or tenses. LLMs can't help but try to help. The same for it's not X, but Y. Sure it's a known pattern, but it's not like people don't use this trope all the time.
In my experience, Pangram is great for detecting an author who is trying to pass someone else's work as theirs, or if they are tackling a subject they have little to no knowledge in.
I’m surprised at the current sentiment in the comments. Pangram is amazing and has really interesting engineering too. I would have guessed that reliably identifying LLM generated text was not possible without watermarks.
> I would have guessed that reliably identifying LLM generated text was not possible
It depends what you mean by "reliably." If you mean, "we should be comfortable relying on this kind of tool at scale to identify and punish students, professionals, and writers who may have used AI," absolutely not.
If you take "reliably" to mean "1 in 200 false positive rate" as they disclose on their front page, absolutely that is possible (they are doing it today!). If you think there are more than 200 assignments turned in over a given year at university, you probably do not consider a tool like this fit for purpose. It's an open question whether those procuring said tool are aware of this
Unfortunately their marketing is really insisting on the former, and trying to push it into the zeitgeist that detection of AI-generated or edited text is reliable-type-1 now and long-term. They fail to make it clear that this is merely a tool that strongly suggests text follows patterns known to us at the present time of known LLMs. However, that fingerprint will drift over time, as LLMs get better, human writing style evolves, and the line between human and "smart autocorrect" becomes even blurrier (does speech-to-text push the model into "AI assisted" mode, because it tidied up your punctuation, for example?)
> If you take "reliably" to mean "1 in 200 false positive rate" as they disclose on their front page
Pangram claims a 1 in 10,000 false positive rate (rate at which human-authored texts are incorrectly classified as AI-generated). 1 in 200 sounds like the false negative rate (rate at which AI-generated texts are classified as human-authored), or perhaps a rate for a specific category of text.
Good point of precision, I read their site too quickly. I don't think my argument materially changes with that number instead however. There were over 40,000 students enrolled in my university alone. Generously assuming they only turn in one assignment per year, having 4 of them go through the "computer says you cheated, and as you can see, it's 99.9% accurate" gauntlet is not a price we should be willing to pay for... the marginal benefit of this tool over more classical ways to proctor and assess pupils
they also publish great tech reports. their founder is so confident in their model that he's regularly on social media offering bounties for false positives
Their philosophy is that false positives are much worse than false negatives (where "positive" means AI generated), so one would expect high precision and only medium recall.
Based on their methodology it looks like the accuracy figures (99.82% for Opus 5) are the true positive rate rather than a combined metric that factors in the false positive rate as well. They claim 1 in 10,000 but it would be nice if we had a per-model breakdown for that specific test.
Tools like Pangram can detect whether text is written by AI or a human with a probability non-negligibly different from 0.5. Therefore, the Turing test in its strictest form (that there must not exist such a distinguisher) has not been passed.
Well, the Turing test involved a _human_ evaluating whether it was talking to a machine or not. In this case, another AI is necessary, which means, in its strict form, the Turing test has indeed been beaten.
I didn't want to shell out $20 a month for the general thing, so I spent $90 on data collection and built my own for code comments specifically. It runs locally in your browser with a relatively small classification model trained on old-school stylometric features. You can try that before turning to Pangram for uncertain cases, if you wish.[1]
It's easy to get high accuracy numbers if you're testing on long (50+ words) texts. Much harder when the documens are short, as code comments tend to be.[2]
Heh some obvious AI generated image made with NovelAI V5 with transparent background enabled are being detected as no AI, so there goes one way to bypass it.
I don't work for Pangram, but I do a lot of writing at work. API docs, blogs, code, tweets, linkedin, all that.
If you've tried AI detectors a couple years ago, they're basically in the same position that coding agents were a few years ago, where everyone was skeptical at first, but the tech has gotten a lot better. Give it a shot, it's quite good. They are slightly tuned a bit towards classifying things as AI, but I imagine that's deliberate.
The only thing is that their models are pricey, but, very useful.
People expect a binary response, is it AI generated yes or no. But it's more complicated than that. For example, if you see an emdash, it's probably AI generated. But it can also mean the author used it for fixing grammar or tenses. LLMs can't help but try to help. The same for it's not X, but Y. Sure it's a known pattern, but it's not like people don't use this trope all the time.
In my experience, Pangram is great for detecting an author who is trying to pass someone else's work as theirs, or if they are tackling a subject they have little to no knowledge in.
> …if you see an emdash it’s probably AI generated…
Or, and—hear me out—the writer typed two successive hyphens on their iDevice.
Microsoft Word converts dashes to emdashes if you type "<word> - <word>".
Outlook used to do this. I'm using a Mac, and Outlook on the Mac is a different animal, so Outlook on Windows may still do this conversion.
Small correction:
To get an en dash in Word, type "<word> - <word>" (word space hyphen space word) followed by spacebar.
To get an em dash, type "<word>--<word>" (word hyphen hyphen word) followed by spacebar.
Which has been happening for over twenty years on taoofmac.com, for instance. It’s a reflex by now.
I’m surprised at the current sentiment in the comments. Pangram is amazing and has really interesting engineering too. I would have guessed that reliably identifying LLM generated text was not possible without watermarks.
> I would have guessed that reliably identifying LLM generated text was not possible
It depends what you mean by "reliably." If you mean, "we should be comfortable relying on this kind of tool at scale to identify and punish students, professionals, and writers who may have used AI," absolutely not.
If you take "reliably" to mean "1 in 200 false positive rate" as they disclose on their front page, absolutely that is possible (they are doing it today!). If you think there are more than 200 assignments turned in over a given year at university, you probably do not consider a tool like this fit for purpose. It's an open question whether those procuring said tool are aware of this
Unfortunately their marketing is really insisting on the former, and trying to push it into the zeitgeist that detection of AI-generated or edited text is reliable-type-1 now and long-term. They fail to make it clear that this is merely a tool that strongly suggests text follows patterns known to us at the present time of known LLMs. However, that fingerprint will drift over time, as LLMs get better, human writing style evolves, and the line between human and "smart autocorrect" becomes even blurrier (does speech-to-text push the model into "AI assisted" mode, because it tidied up your punctuation, for example?)
"What color are your bits" is good reading today as it was 20 years ago: https://ansuz.sooke.bc.ca/entry/23
> If you take "reliably" to mean "1 in 200 false positive rate" as they disclose on their front page
Pangram claims a 1 in 10,000 false positive rate (rate at which human-authored texts are incorrectly classified as AI-generated). 1 in 200 sounds like the false negative rate (rate at which AI-generated texts are classified as human-authored), or perhaps a rate for a specific category of text.
Good point of precision, I read their site too quickly. I don't think my argument materially changes with that number instead however. There were over 40,000 students enrolled in my university alone. Generously assuming they only turn in one assignment per year, having 4 of them go through the "computer says you cheated, and as you can see, it's 99.9% accurate" gauntlet is not a price we should be willing to pay for... the marginal benefit of this tool over more classical ways to proctor and assess pupils
they also publish great tech reports. their founder is so confident in their model that he's regularly on social media offering bounties for false positives
https://www.pangram.com/blog/pangram-4-technical
https://pangram-public.s3.us-east-1.amazonaws.com/pdf/pangra...
https://www.pangram.com/blog/introducing-pangram-image-detec...
Gotta be honest. Pangram is not that reliable. It flagged my texts as 100% human made whilst at best it was 50%.
Their philosophy is that false positives are much worse than false negatives (where "positive" means AI generated), so one would expect high precision and only medium recall.
I had the same experience for my writing that was AI generated.
Based on their methodology it looks like the accuracy figures (99.82% for Opus 5) are the true positive rate rather than a combined metric that factors in the false positive rate as well. They claim 1 in 10,000 but it would be nice if we had a per-model breakdown for that specific test.
Really we'd want a false discovery rate but that would depend on the prior.
Tools like Pangram can detect whether text is written by AI or a human with a probability non-negligibly different from 0.5. Therefore, the Turing test in its strictest form (that there must not exist such a distinguisher) has not been passed.
Well, the Turing test involved a _human_ evaluating whether it was talking to a machine or not. In this case, another AI is necessary, which means, in its strict form, the Turing test has indeed been beaten.
I didn't want to shell out $20 a month for the general thing, so I spent $90 on data collection and built my own for code comments specifically. It runs locally in your browser with a relatively small classification model trained on old-school stylometric features. You can try that before turning to Pangram for uncertain cases, if you wish.[1]
It's easy to get high accuracy numbers if you're testing on long (50+ words) texts. Much harder when the documens are short, as code comments tend to be.[2]
[1]: https://xkqr.org/aicomment
[2]: https://entropicthoughts.com/better-ai-comment-classifier
Heh some obvious AI generated image made with NovelAI V5 with transparent background enabled are being detected as no AI, so there goes one way to bypass it.
I detect SPAM.
I don't work for Pangram, but I do a lot of writing at work. API docs, blogs, code, tweets, linkedin, all that.
If you've tried AI detectors a couple years ago, they're basically in the same position that coding agents were a few years ago, where everyone was skeptical at first, but the tech has gotten a lot better. Give it a shot, it's quite good. They are slightly tuned a bit towards classifying things as AI, but I imagine that's deliberate.
The only thing is that their models are pricey, but, very useful.
> An AI detector that actually works.
I detect Claude.
Interestingly checking good chunks of their site, it detects as 100% human
I mean, that's just the most direct way to say what they're saying? Doesn't smell Claudy to me.
Like, if I had an AI detector that actually worked—and all the other ones didn't, which they don't—I'd put it exactly the same way.
If I were your editor, I'd strike the adverb. It weakens the sentence.
[dead]
[dead]
Negative? What?
[flagged]