4 ms·

How exactly do you prove that someone plagiarized via ChatGPT? In my experience, my students simply deny deny deny, even when I have them dead-to-rights. I me

How exactly do you prove that someone plagiarized via ChatGPT? In my experience, my students simply deny deny deny, even when I have them dead-to-rights. I mean, I can understand a professor have a strong suspicion of plagiarism, but proving that a student used ChatGPT seems difficult.

Barrin92 4y ago
the same way you prove that someone plagiarized anyone else. You demonstrate that the text was reproduced, in this case probably by checking the assignment with ChatGPT themselves and comparing it to student answers. ChatGPT has a very verbose, uncanney valley style to it that likely sounded like nothing the student had organically submitted before, it's not that hard to tell.
- busyant 4y ago
  > ChatGPT has a very verbose, uncanney valley style to it that likely sounded like nothing the student had organically submitted before, it's not that hard to tell. I've had 2 students threaten me and my university with lawsuits over accusations of cheating (not plagiarism, fwiw). And I had far more compelling evidence than an uncanny valley of verbosity. I think it's more likely that this story is bullshit (as someone else in this forum has also speculated) than the likelihood that a professor would be willing to go to the mattresses with only a stylistic argument for plagiarism. > it's not that hard to tell. It might not be that difficult to tell, but proof is another matter. edit: The students were bluffing with their threats of lawsuits, but it still made for two unpleasant semesters.
chrsstrm 4y ago
In some alternate timeline an enterprising person at OpenAI has written a memo in which they requested a new service to be implemented. This service hashes all generated responses before they are sent to the end user and saves each hash in a db somewhere. After OpenAI’s products have become popular enough that plagiarism is a valid concern, OpenAI releases their (paid) generated response history API which allows you to provide a hash and get a response as to whether or not it exists in their database. This puts a stop to 85% of all low effort plagiarism and earns OpenAI a decent additional revenue stream.
- neodymiumphish 4y ago
  You'd have to have the hashes based on small portions of the responses, in order to catch copies where the recipient made minor changes throughout.
  - gruez 4y ago
    Sounds like a reason to partner with turnitin.
  - comboy 4y ago
    Shazam but for text. You read a good article so you wonder which bot made this. You've heard it here first.
  - g_p 4y ago
    This sounds like a good opportunity for fuzzy hashes like SSDEEP. This ought to give enough flexibility to detect minor tweaks to output. https://www.microsoft.com/en-us/security/blog/2021/07/27/combing-through-the-fuzz-using-fuzzy-hashing-and-deep-learning-to-counter-malware-detection-evasion-techniques/ https://www.microsoft.com/en-us/security/blog/2021/07/27/com...
  - wkat4242 4y ago
    And then you have the issue of those small parts being so common that they could match even normal English. It's not an easy problem to solve :(
- tkiolp4 4y ago
  But that would immediately scare away all the future potential OpenAI customers… why would OpenAI do that? If customers know that it’s possible to determine that a text came from ChatGPT with a 100% accuracy, then no one would use ChatGPT.
  - paulryanrogers 4y ago
    Why? This assumes the only market is for cheaters and those seeking ghost writers.
schwoll 4y ago
Someone was working on an AI that could detect if text was generated by ChatGPT. I'm sure through some smart queries it could be figured out.
naikrovek 4y ago
> proving that a student used ChatGPT seems difficult. I wonder if there is a transparent proxy in place logging access to ChatGPT when on a campus network. My employer does this, and makes no effort to hide it. Everything is logged and specific things are captured for evaluation later.
g_p 4y ago
While there are interesting comments here looking at technical solutions like fuzzy/partial hashing, the solution for a professor is likely a bit different. Rather than prove the student used Chat GPT (hard for reasons you point out), isn't this better handled as a suspected cheating scenario (collusion, candidate impersonation etc.) Many institutions have a regulation on the books to allow for oral examination to be carried out for any course at the discretion of the course leader, for the purpose of evaluating a student for award of a grade. If you suspected a student had used significant outside help, you'd presumably do the same - have an oral examination, and discuss the work, ask them to discuss the subject matter, and test their understanding of it. At that point, either the student crumbles and doesn't have any level of understanding beyond "paper thin" (like if they paid an essay mill to write the paper), or they have a good understanding. If the latter, then they meet the learning outcomes of the class and have demonstrated this satisfactorily. And if they have learned the content and show an understanding of it, it's effectively acted as a study aid and actually helped them to learn. They now have shown they know the material, so they have passed the course legitimately. I suspect in the longer term we'll see attempts to ask questions which can't be so easily answered by generative waffle, which actually should be an improvement on many types of examination. How this works in less scientific disciplines will be more challenging - I imagine they'll struggle to create "un-GPT'able" questions. But colleges adapted (or at least good professors did) to remote examinations during the pandemic - declare the exam "open book", and ask questions that require robust understanding of topics, such that sitting and reading the book will run you out of time. Give students questions with generated parameters so each student gets a slightly different variant of numerical questions, etc. The solution to language models generating believable or compelling waffle doesn't need to be technical - I suspect non technical solutions may win out. Whether they scale enough (or a professor cares enough to actually sit down with students for a few minutes) is another question, but maybe they should be doing that in the first place?
- busyant 4y ago
  > Many institutions have a regulation on the books to allow for oral examination to be carried out for any course at the discretion of the course leader, for the purpose of evaluating a student for award of a grade. This, in my opinion, is the only approach that will work. When I ask cheaters to come in to explain the work, they get cold feet.
vba616 4y ago
Why do you have to prove plagiarism? All you have to do is fail every paper of ChatGPT quality.
- g_p 4y ago
  This would be one approach. However devils advocate here - I took an essay question verbatim from a humanities assessed exercise, and very quickly (think seconds of effort) created a prompt and got Chat GPT to write text each time it was prompted, until told to stop. I copied the output (which I'd requested references to be included in), and gave it to the professor who set the question. They very quickly confirmed the language model essay was considerably better than their human students' essays. It was more coherent and readable, better answered the question set, was more "on topic", and even better structured - it followed a structure that introduced a position, then substantiated the position, gave some mention to alternative viewpoints, and compared and justified the position in relation to them. The topic wasn't hard, you could easily Google copious essays on the topic, but the output wasn't pulled from the web verbatim (even in parts), and it was better than the actual students' work! The biggest failing was the references - language models generate plausible and readable text... So it invented lots of very interesting and plausible sounding references, none of which existed. The titles sounded relevant. The authors were appropriate to the subject matter, the journals or publication venues were also appropriate and relevant. It's just that none of them actually existed!
stonogo 4y ago
If ChatGPT input and output is not being logged for future use, they're missing a trick. They should provide a tool to universities to paste in a given excerpt and see if it shows up in the output logs. Might even charge for the service!

Barrin92 4y ago

the same way you prove that someone plagiarized anyone else. You demonstrate that the text was reproduced, in this case probably by checking the assignment with ChatGPT themselves and comparing it to student answers. ChatGPT has a very verbose, uncanney valley style to it that likely sounded like nothing the student had organically submitted before, it's not that hard to tell.

busyant 4y ago
> ChatGPT has a very verbose, uncanney valley style to it that likely sounded like nothing the student had organically submitted before, it's not that hard to tell. I've had 2 students threaten me and my university with lawsuits over accusations of cheating (not plagiarism, fwiw). And I had far more compelling evidence than an uncanny valley of verbosity. I think it's more likely that this story is bullshit (as someone else in this forum has also speculated) than the likelihood that a professor would be willing to go to the mattresses with only a stylistic argument for plagiarism. > it's not that hard to tell. It might not be that difficult to tell, but proof is another matter. edit: The students were bluffing with their threats of lawsuits, but it still made for two unpleasant semesters.

chrsstrm 4y ago

In some alternate timeline an enterprising person at OpenAI has written a memo in which they requested a new service to be implemented. This service hashes all generated responses before they are sent to the end user and saves each hash in a db somewhere. After OpenAI’s products have become popular enough that plagiarism is a valid concern, OpenAI releases their (paid) generated response history API which allows you to provide a hash and get a response as to whether or not it exists in their database. This puts a stop to 85% of all low effort plagiarism and earns OpenAI a decent additional revenue stream.

neodymiumphish 4y ago
You'd have to have the hashes based on small portions of the responses, in order to catch copies where the recipient made minor changes throughout.
- gruez 4y ago
  Sounds like a reason to partner with turnitin.
- comboy 4y ago
  Shazam but for text. You read a good article so you wonder which bot made this. You've heard it here first.
- g_p 4y ago
  This sounds like a good opportunity for fuzzy hashes like SSDEEP. This ought to give enough flexibility to detect minor tweaks to output. https://www.microsoft.com/en-us/security/blog/2021/07/27/combing-through-the-fuzz-using-fuzzy-hashing-and-deep-learning-to-counter-malware-detection-evasion-techniques/ https://www.microsoft.com/en-us/security/blog/2021/07/27/com...
- wkat4242 4y ago
  And then you have the issue of those small parts being so common that they could match even normal English. It's not an easy problem to solve :(
tkiolp4 4y ago
But that would immediately scare away all the future potential OpenAI customers… why would OpenAI do that? If customers know that it’s possible to determine that a text came from ChatGPT with a 100% accuracy, then no one would use ChatGPT.
- paulryanrogers 4y ago
  Why? This assumes the only market is for cheaters and those seeking ghost writers.

schwoll 4y ago

Someone was working on an AI that could detect if text was generated by ChatGPT. I'm sure through some smart queries it could be figured out.

naikrovek 4y ago

> proving that a student used ChatGPT seems difficult. I wonder if there is a transparent proxy in place logging access to ChatGPT when on a campus network. My employer does this, and makes no effort to hide it. Everything is logged and specific things are captured for evaluation later.

g_p 4y ago

While there are interesting comments here looking at technical solutions like fuzzy/partial hashing, the solution for a professor is likely a bit different. Rather than prove the student used Chat GPT (hard for reasons you point out), isn't this better handled as a suspected cheating scenario (collusion, candidate impersonation etc.) Many institutions have a regulation on the books to allow for oral examination to be carried out for any course at the discretion of the course leader, for the purpose of evaluating a student for award of a grade. If you suspected a student had used significant outside help, you'd presumably do the same - have an oral examination, and discuss the work, ask them to discuss the subject matter, and test their understanding of it. At that point, either the student crumbles and doesn't have any level of understanding beyond "paper thin" (like if they paid an essay mill to write the paper), or they have a good understanding. If the latter, then they meet the learning outcomes of the class and have demonstrated this satisfactorily. And if they have learned the content and show an understanding of it, it's effectively acted as a study aid and actually helped them to learn. They now have shown they know the material, so they have passed the course legitimately. I suspect in the longer term we'll see attempts to ask questions which can't be so easily answered by generative waffle, which actually should be an improvement on many types of examination. How this works in less scientific disciplines will be more challenging - I imagine they'll struggle to create "un-GPT'able" questions. But colleges adapted (or at least good professors did) to remote examinations during the pandemic - declare the exam "open book", and ask questions that require robust understanding of topics, such that sitting and reading the book will run you out of time. Give students questions with generated parameters so each student gets a slightly different variant of numerical questions, etc. The solution to language models generating believable or compelling waffle doesn't need to be technical - I suspect non technical solutions may win out. Whether they scale enough (or a professor cares enough to actually sit down with students for a few minutes) is another question, but maybe they should be doing that in the first place?

busyant 4y ago
> Many institutions have a regulation on the books to allow for oral examination to be carried out for any course at the discretion of the course leader, for the purpose of evaluating a student for award of a grade. This, in my opinion, is the only approach that will work. When I ask cheaters to come in to explain the work, they get cold feet.

vba616 4y ago

Why do you have to prove plagiarism? All you have to do is fail every paper of ChatGPT quality.

g_p 4y ago
This would be one approach. However devils advocate here - I took an essay question verbatim from a humanities assessed exercise, and very quickly (think seconds of effort) created a prompt and got Chat GPT to write text each time it was prompted, until told to stop. I copied the output (which I'd requested references to be included in), and gave it to the professor who set the question. They very quickly confirmed the language model essay was considerably better than their human students' essays. It was more coherent and readable, better answered the question set, was more "on topic", and even better structured - it followed a structure that introduced a position, then substantiated the position, gave some mention to alternative viewpoints, and compared and justified the position in relation to them. The topic wasn't hard, you could easily Google copious essays on the topic, but the output wasn't pulled from the web verbatim (even in parts), and it was better than the actual students' work! The biggest failing was the references - language models generate plausible and readable text... So it invented lots of very interesting and plausible sounding references, none of which existed. The titles sounded relevant. The authors were appropriate to the subject matter, the journals or publication venues were also appropriate and relevant. It's just that none of them actually existed!

stonogo 4y ago

If ChatGPT input and output is not being logged for future use, they're missing a trick. They should provide a tool to universities to paste in a given excerpt and see if it shows up in the output logs. Might even charge for the service!