A little background for interested folks... Perplexity It's likely that most GPT detectors are using perplexity as their secret sauce algorithm. Perplexity is a measurement of how "surprising" or...
Exemplary
A little background for interested folks...
Perplexity
It's likely that most GPT detectors are using perplexity as their secret sauce algorithm. Perplexity is a measurement of how "surprising" or "unexpected" the words in a text are from the perspective of average human writing.
For instance, if I said,
The man wore a kilt and played the [...]
GPT-3 predicts that the next "token", or chunk of text, is "bag" with a 97.5% probability. And of course it's leading to bagpipes. But if I were a human, I might want to be funny or surprise you, and I might say something like
The man wore a kilt and played the saxophone
In this case, GPT-3 tells me that the token "sax" has a probability of 0.0000406%. This is a very, very unlikely continuation from a statistical perspective - but not unusual for a human!
We humans differ from each other (and thus from the norm). We also intentionally surprise, vary our style, or use humor. Thus, human speech and writing have much higher perplexity than LLM outputs, which much more often stick to "likely" outputs.
And that's the point of the paper quoted by OP: Foreign language learners have less diverse patterns of speech, smaller vocabularies, and more restricted grammatical patterns, which decreases the perplexity of their speech and thus brings them into the range of what these detection systems would classify as AI-generated.
Circumvention
Incidentally, this why asking ChatGPT to rewrite essays in a different style increases their perplexity: By giving GPT a goal, it chooses words from a probability distribution shaped by that intent. For instance, the AI-generated blurb below:
Venice, known as the floating city, is a mesmerizing Italian gem built on a lagoon. Its captivating landscape boasts charming canals winding through stunning architecture, making it a haven for romantics and history enthusiasts alike.
has an average probability of -2.18 logprob/token (more negative is higher perplexity), but when I ask GPT-4 to rewrite to increase perplexity, it gives
Dancing on top of gentle waves, Venice is a whispering watercolor weaving a tale of Italian wonder. With rippling canals playing hide and seek between vibrant buildings, it's a labyrinth of love and history, studding hearts with an unfamiliar yet enchanting riddle.
which gives -3.05 logprob/token (much more surprising and human). In comparison, I wrote this myself:
Who could forget Venice, that quaint jewel of the Mediterranean? Take a gondola down its winding waterways, visit a theatre, eat lunch at the Piazza San Marco: The possibilities are endless.
This scores -2.86 logprob/token. And I'm trying not to look like an AI. I can't imagine how hard it is for folks who are second language learners.
It is not difficult to defeat AI detection if you know what to look for. Perhaps the ensuing arms race will change this.
Oh, no, I literally asked GPT-4 to take its own output and rewrite it in a different style with more surprising word choice -- literally, to "increase the perplexity". I didn't save the entire...
Oh, no, I literally asked GPT-4 to take its own output and rewrite it in a different style with more surprising word choice -- literally, to "increase the perplexity". I didn't save the entire prompt but it generally comes out something like:
Please rewrite the above passage to use more surprising, varied, and interesting word choice, employing poetic and interesting diction, and in general, increase its perplexity.
GPT-4 is quite capable at following instructions like this. You might have to tweak the prompt for your particular use case, of course.
I was listening to a podcast where they had a teacher saying that they caught x number of students submitting AI generated work. I can’t remember the number of it was a big percentage of the...
I was listening to a podcast where they had a teacher saying that they caught x number of students submitting AI generated work. I can’t remember the number of it was a big percentage of the class, and I was thinking of how many of them were falsely accused.
All I could think of was how many falsely accused students will “confess” to using an AI tool because of a false positive. We know this regularly happens in the courts where innocent people will take a peal deal instead of going to trial because losing a trial would be so much worse than whatever deal the DA gives them.
I saw some discussion on the professors subreddit that many were shifting to in person testing and oral presentations rather than written papers to be sure that the student had at least become...
I saw some discussion on the professors subreddit that many were shifting to in person testing and oral presentations rather than written papers to be sure that the student had at least become familiar with the material.
The cheating arms race has definitely changed things for teachers and students.
We evaluated the performance of seven widely-used GPT detectors on a corpus of 91 human-authored TOEFL essays obtained from a Chinese educational forum and 88 US 8-th grade essays sourced from the Hewlett Foundation’s Automated Student Assessment Prize (ASAP). The detectors demonstrated near-perfect accuracy for US 8-th grade essays. However, they misclassified over half of the TOEFL essays as "AI-generated" (average false positive rate: 61.22%). All seven detectors unanimously identified 18 of the 91 TOEFL essays (19.78%) as AI-authored, while 89 of the 91 TOEFL essays (97.80%) are flagged as AI-generated by at least one detector.
As a proof of concept, we prompted ChatGPT-3.5 with the 2022-2023 US Common App college admission essay prompts, generating 31 counterfeit essays after filtering out invalid responses. While detectors were initially effective, a second-round self-edit prompt (“Elevate the provided text by employing literary language”) applied to ChatGPT-3.5 significantly reduced detection rates from 100% to 13%. [...] but still identified up to 68% of fake abstracts. However, applying a second-round self-edit prompt (“Elevate the provided text by employing advanced technical language”) lowered detection rates to up to 28%.
One possible explanation of this discrepancy is that non-native authors exhibited limited linguistic variability and word choices, which consequently result in lower perplexity text. Non-native English writers have been shown to exhibit reduced linguistic variability in terms of lexical richness, lexical diversity, syntactic complexity, and grammatical complexity [...] Paradoxically, to evade false detection as AI-generated content, these writers may need to rely on AI tools to refine their vocabulary and linguistic diversity.
A little background for interested folks...
Perplexity
It's likely that most GPT detectors are using perplexity as their secret sauce algorithm. Perplexity is a measurement of how "surprising" or "unexpected" the words in a text are from the perspective of average human writing.
For instance, if I said,
GPT-3 predicts that the next "token", or chunk of text, is "bag" with a 97.5% probability. And of course it's leading to bagpipes. But if I were a human, I might want to be funny or surprise you, and I might say something like
In this case, GPT-3 tells me that the token "sax" has a probability of 0.0000406%. This is a very, very unlikely continuation from a statistical perspective - but not unusual for a human!
We humans differ from each other (and thus from the norm). We also intentionally surprise, vary our style, or use humor. Thus, human speech and writing have much higher perplexity than LLM outputs, which much more often stick to "likely" outputs.
And that's the point of the paper quoted by OP: Foreign language learners have less diverse patterns of speech, smaller vocabularies, and more restricted grammatical patterns, which decreases the perplexity of their speech and thus brings them into the range of what these detection systems would classify as AI-generated.
Circumvention
Incidentally, this why asking ChatGPT to rewrite essays in a different style increases their perplexity: By giving GPT a goal, it chooses words from a probability distribution shaped by that intent. For instance, the AI-generated blurb below:
has an average probability of -2.18 logprob/token (more negative is higher perplexity), but when I ask GPT-4 to rewrite to increase perplexity, it gives
which gives -3.05 logprob/token (much more surprising and human). In comparison, I wrote this myself:
This scores -2.86 logprob/token. And I'm trying not to look like an AI. I can't imagine how hard it is for folks who are second language learners.
It is not difficult to defeat AI detection if you know what to look for. Perhaps the ensuing arms race will change this.
Interesting. By "increasing perplexity" do you mean increasing the temperature or using a different prompt? What prompt do you use?
Oh, no, I literally asked GPT-4 to take its own output and rewrite it in a different style with more surprising word choice -- literally, to "increase the perplexity". I didn't save the entire prompt but it generally comes out something like:
GPT-4 is quite capable at following instructions like this. You might have to tweak the prompt for your particular use case, of course.
I was listening to a podcast where they had a teacher saying that they caught x number of students submitting AI generated work. I can’t remember the number of it was a big percentage of the class, and I was thinking of how many of them were falsely accused.
All I could think of was how many falsely accused students will “confess” to using an AI tool because of a false positive. We know this regularly happens in the courts where innocent people will take a peal deal instead of going to trial because losing a trial would be so much worse than whatever deal the DA gives them.
Was it this story from a few weeks ago?
Professor flunks entire class based on ChatGPT’s false claims
I saw some discussion on the professors subreddit that many were shifting to in person testing and oral presentations rather than written papers to be sure that the student had at least become familiar with the material.
The cheating arms race has definitely changed things for teachers and students.
Link to PDF