I’m a teacher that requires students to write, so I’m always on the lookout for reliable AI detection. Haven’t anything close to reliable yet, but Pangram definitely seemed to be better than anything else in my (very limited) tests.
That said, it did incorrectly flag something that I personally wrote as AI, which baffled me. I spent a while playing around with the text and feeding it back into Pangram, trying to figure out what the issue was. Turns out, I had a section in the text with three bullet points. I removed the bullets (literally just the bullets themselves, no actual text) and it passed as human.
While it seems to perform better than anything else, I’m a bit skeptical of their claimed success rate.
It's interesting that it cares about formatting choices like bullet points.
On the one hand, it makes the system more brittle and arguably overfit, since it's not making a decision on the content of the writing itself. Most people would say that AI writing is still AI writing even if bullet points are removed, and vice versa.
On the other hand, we know that AI writing does has a very specific formatting signature (e.g. em-dashes, bullet points) so it seems unwise to completely ignore it.
> It's interesting that it cares about formatting choices like bullet points.
Since AI I have noticed a lot of sudden bulleted lists appearing in pretty benign forum posts. That along with "You are right" "You are correct" "you're right" as the response.
The tools will never be perfectly reliable. It’s good that people like you are aware that it’s a statistical tool, not a perfect oracle.
For student work I would think it’s reasonable to look for patterns across each student’s work. A single Pangram positive across 10 assignments shouldn’t be enough to bring suspicion upon a student, but when 8 out of 10 of someone’s assignments are coming up as AI it’s time to talk.
The trend I’m hearing from my teacher friends is that students are generating AI papers and then putting them into tools like ZeroGPT and Pangram before submission. They do what you did: Remove sections, rewrite things, and make adjustments until the AI score goes down enough.
I heard that one classroom had developed a trick of adding spelling mistakes to ChatGPT output to try to make it look human written. So the teacher was getting pure AI slop but every paragraph had a couple typos inserted.
As for the 3 bullet points: That was a huge tell for one of the recent popular models. Almost all writing output would get sections of plus or minus 3 bullet points in a certain style. The newer frontier models do it less but that was a big LLM pattern for a while. Unfortunate for anyone who liked to write with 3 bullet points commonly. They’re feeling the pain of people who were emdash enjoyers before 2023
Interesting methodology. The results for Pangram are surprisingly good.
These tools are definitely not 100% perfect, which is the primary complaint used to dismiss them. However the error rate is also getting impressively low.
In most cases I see socially and online, there is a high suspicion that the content is AI generated before someone thinks to submit it to Pangram. It’s used on-demand as a tool to confirm suspicions. I have seen several cases where Pangram had some false negatives where the content was judged to be likely human written but the author later admitted it was written by an LLM.
Pangram is very interesting in the context of Substack because the platform was a target for lazy AI newsletters. People realized they could start 10 (or maybe many more) newsletters and spend only a few minutes getting ChatGPT to write posts for them. Starting an AI generated substack and trying to get paid subscribers for it was becoming one of the popular ways to use AI to try to get a little cash. Having a tool that makes it a little bit harder, at least until the LLMs get good enough to evade it, was important for the platform.
I think outside of a few niche uses, e.g. schools/universities, services like Pangram will struggle.
Currently people care because so much of AI output is still poor quality. But given I personally know of someone with published AI-generated works where the reviews are gushing and positive across the board, I doubt people will care whether works are AI generated for very long vs whether the text is good.
A way of certifying a work as above a certain literary quality would likely have far more longevity, as I doubt the market of people who care whether a given work is AI or not irrespective of quality will be particularly large.
I actually disagree, I think B people care a lot if a created work is an intended artistic thing vs produced by a machine. It’s like the difference between getting someone a generic hallmark cards vs actually writing something inside. Theres a place for the hallmarks but a plain one with no human touch is still considered tacky
Survey people about who created a given work, and you'll find a whole lot of people who like something without having the faintest clue who made it, or anything about why, before even considering the people who can even name the people involved. It's hard to believe that suggest people are so quick to care about the artistic merit.
As a concrete example relating to books, the success of the Stratemeyer Syndicate in packaging up book series ghostwritten by "nameless" authors under single pseudonyms is an example. Stratemeyer series like Nancy Drew has sold more than 80 million copies. At one point 98% of American children listed a Stratemeyer title as their most popular books.
How many of those do you think saw those works - most sold as a work for hire for next to nothing and churned out by the hundreds - as "an intended artistic thing"?
People who are a lot about literature tends to seriously overestimate how readers engage with writing. As it stands it's hard to get people to care enough to read, much less to get them to care about artistic intent.
I think people who bet on people caring about that as a means to stem the tide of AI content will be very disappointed.
> I have seen several cases where Pangram had some false negatives
Yep, Pangram says they think false positives are way worse than false negatives, so if their AI isn't sure enough they prefer a "not ai" output to an "is ai" output.
> Pangram boasts a false positive rate of 1 in a 10,000. That is, if Pangram says a block of text is AI there is only a one in ten thousand chance that it was written by a human.
That'd be if they had a false discovery rate of 1/10,000.
If for instance:
* 100,000 samples are tested
* 100 of which are AI-generated, the rest human-written
* Pangram flags 50 of the AI-generated samples (true positives)
* Pangram also flags 10 human-written samples (false positives)
Then the FPR is 1 in 10,000, but the chance that a flagged sample isn't actually AI (FDR) is 1 in 6.
The classic. If there is one concept people should learn from stats it should be base rates and the Bayes update. It's very difficult to teach though. The mind just jumps to that interpretation, and even if you teach it, people just go right back to interpreting it like that anyway, because they think "yeah yeah technically something something math, I know, but basically in practice it kinda just means that anyway". Even plenty of educated professionals in science do not understand this and continue to get it wrong each day.
(Somewhat-plausible-to-me explanation: It's looking for various stylistic features; older writing very rarely has the most AI-like features, or perhaps almost always has some non-AI-like features that outweigh whatever signs of AI-ness might be there. Present-day writers are more likely to get wrongly flagged as AI.)
Pangram is witchcraft to me. The way they can correctly detect AI writing from small samples, with so little statistical signal, is crazy. I've seen it correctly detect AI writing even when people use those "humanizer" rewriting skills that remove the hallmarks of AI writing and make the text indistinguishable from human text. But not to Pangram.
Consider that they are in effect detecting whether text has been generated by one of a few dozen entities that has produced more text for them to analyze than any given human author.
Now imagine if they put the same effort into detecting if a given text was produced by one of a few dozen super-prolific human writers. I'd imagine they'd get pretty good at that too.
The main limitation of those "humanizer" rewriters is that most of them focus on making the next less detectable to humans, by making them read better. There's likely to be plenty of signal left that isn't affected by trying to make the text read better the same way human writers have plenty of idiosyncrasies despite being human.
I agree. I ran it on a small sample last night, basically three paragraphs of AI text that I thought read 100% human, except for em dashes, which I use in writing anyway. It flagged the first and last paras as "high" AI and the middle para as "high" human. I have no idea what it is seeing that sets it off.
I find it's surprising that people doubt the effectiveness of Pangram when I feel like its job has gotten easier ever since Claude models got RLHF to death with its bizarre punchy negatives and tortured analogies, making AI writing much more noticeable than the GPT 4o era.
Pangram is amazing. I hope they eventually provide something like a browser plugin that scans text for being AI generated as you read stuff. I’d pay good money for that.
If you just try to minimize the share of false positives, you are only maximizing 1.
The measure which maximizes all four values is the binary Pearson correlation, also called "phi coefficient" or "MCC". It is 1 if the classifier is always correct, 0 if it guesses randomly (its predictions are statistically independent of the facts) and -1 if it always predicts the opposite of the true value.
That said, it did incorrectly flag something that I personally wrote as AI, which baffled me. I spent a while playing around with the text and feeding it back into Pangram, trying to figure out what the issue was. Turns out, I had a section in the text with three bullet points. I removed the bullets (literally just the bullets themselves, no actual text) and it passed as human. While it seems to perform better than anything else, I’m a bit skeptical of their claimed success rate.
On the one hand, it makes the system more brittle and arguably overfit, since it's not making a decision on the content of the writing itself. Most people would say that AI writing is still AI writing even if bullet points are removed, and vice versa.
On the other hand, we know that AI writing does has a very specific formatting signature (e.g. em-dashes, bullet points) so it seems unwise to completely ignore it.
Since AI I have noticed a lot of sudden bulleted lists appearing in pretty benign forum posts. That along with "You are right" "You are correct" "you're right" as the response.
For student work I would think it’s reasonable to look for patterns across each student’s work. A single Pangram positive across 10 assignments shouldn’t be enough to bring suspicion upon a student, but when 8 out of 10 of someone’s assignments are coming up as AI it’s time to talk.
The trend I’m hearing from my teacher friends is that students are generating AI papers and then putting them into tools like ZeroGPT and Pangram before submission. They do what you did: Remove sections, rewrite things, and make adjustments until the AI score goes down enough.
I heard that one classroom had developed a trick of adding spelling mistakes to ChatGPT output to try to make it look human written. So the teacher was getting pure AI slop but every paragraph had a couple typos inserted.
As for the 3 bullet points: That was a huge tell for one of the recent popular models. Almost all writing output would get sections of plus or minus 3 bullet points in a certain style. The newer frontier models do it less but that was a big LLM pattern for a while. Unfortunate for anyone who liked to write with 3 bullet points commonly. They’re feeling the pain of people who were emdash enjoyers before 2023
These tools are definitely not 100% perfect, which is the primary complaint used to dismiss them. However the error rate is also getting impressively low.
In most cases I see socially and online, there is a high suspicion that the content is AI generated before someone thinks to submit it to Pangram. It’s used on-demand as a tool to confirm suspicions. I have seen several cases where Pangram had some false negatives where the content was judged to be likely human written but the author later admitted it was written by an LLM.
Pangram is very interesting in the context of Substack because the platform was a target for lazy AI newsletters. People realized they could start 10 (or maybe many more) newsletters and spend only a few minutes getting ChatGPT to write posts for them. Starting an AI generated substack and trying to get paid subscribers for it was becoming one of the popular ways to use AI to try to get a little cash. Having a tool that makes it a little bit harder, at least until the LLMs get good enough to evade it, was important for the platform.
Currently people care because so much of AI output is still poor quality. But given I personally know of someone with published AI-generated works where the reviews are gushing and positive across the board, I doubt people will care whether works are AI generated for very long vs whether the text is good.
A way of certifying a work as above a certain literary quality would likely have far more longevity, as I doubt the market of people who care whether a given work is AI or not irrespective of quality will be particularly large.
As a concrete example relating to books, the success of the Stratemeyer Syndicate in packaging up book series ghostwritten by "nameless" authors under single pseudonyms is an example. Stratemeyer series like Nancy Drew has sold more than 80 million copies. At one point 98% of American children listed a Stratemeyer title as their most popular books.
How many of those do you think saw those works - most sold as a work for hire for next to nothing and churned out by the hundreds - as "an intended artistic thing"?
People who are a lot about literature tends to seriously overestimate how readers engage with writing. As it stands it's hard to get people to care enough to read, much less to get them to care about artistic intent.
I think people who bet on people caring about that as a means to stem the tide of AI content will be very disappointed.
Yep, Pangram says they think false positives are way worse than false negatives, so if their AI isn't sure enough they prefer a "not ai" output to an "is ai" output.
That'd be if they had a false discovery rate of 1/10,000.
If for instance:
* 100,000 samples are tested
* 100 of which are AI-generated, the rest human-written
* Pangram flags 50 of the AI-generated samples (true positives)
* Pangram also flags 10 human-written samples (false positives)
Then the FPR is 1 in 10,000, but the chance that a flagged sample isn't actually AI (FDR) is 1 in 6.
https://www.graphpad.com/guides/prism/latest/statistics/comm...
(OP isn't exactly talkigh about p-value, because it's not comparing means, but the main mathematical structure of the Bayesian error is the same)
Of course, it's unclear if the article is using the wrong term, or the wrong definition, or if Pangram is just lying in the first place.
(Somewhat-plausible-to-me explanation: It's looking for various stylistic features; older writing very rarely has the most AI-like features, or perhaps almost always has some non-AI-like features that outweigh whatever signs of AI-ness might be there. Present-day writers are more likely to get wrongly flagged as AI.)
Now imagine if they put the same effort into detecting if a given text was produced by one of a few dozen super-prolific human writers. I'd imagine they'd get pretty good at that too.
The main limitation of those "humanizer" rewriters is that most of them focus on making the next less detectable to humans, by making them read better. There's likely to be plenty of signal left that isn't affected by trying to make the text read better the same way human writers have plenty of idiosyncrasies despite being human.
https://www.pangram.com/solutions/chrome-extension
(disclaimer, i am a pangram research scientist)
1. P(actually true | predicted true) (= true positive rate, recall, sensitivity)
2. P(actually false | predicted false) (= true negative rate, specificity)
3. P(predicted true | actually true) (= positive predictive value, precision)
4. P(predicted false | actually false) (= negative predictive value)
If you just try to minimize the share of false positives, you are only maximizing 1.
The measure which maximizes all four values is the binary Pearson correlation, also called "phi coefficient" or "MCC". It is 1 if the classifier is always correct, 0 if it guesses randomly (its predictions are statistically independent of the facts) and -1 if it always predicts the opposite of the true value.
https://www.pangram.com/research/model-card/pangram-3-3
Recall = 1 - FNR