Plagiarism detection for GPT LLMs is a bad idea. Change my view.

Plagiarism detection for GPT LLMs is a bad idea. Change my view.

Matt Bury -
回帖数：9
Particularly helpful Moodlers的头像 Plugin developers的头像

Plagiarism detection for GPT LLMs is a bad idea. Change my view.

So far everything I've heard about AI generated text detectors has been that firstly, they're wildly unreliable & easy to defeat, e.g. simply change the first sentence & you get a much lower score, & more importantly, they give a lot of false positives & there have been cases where students have been falsely accused of academic misconduct. In the anglophone world, there also seems to be a bias in false positives towards students who's first language isn't English, upon which public universities are now relying more heavily than ever to make up for shortfalls in govt. funding.

Have I got this wrong? Is there some evidence that isn't being reported in the media? Can anyone here change my view?

P.S. This post was provoked by news that Moodle.com is now promoting an AI plagiarism detection tool: https://moodle.com/news/check-similarities-and-ai-generated-texts-with-advacheck/

平均分： Coolest thing ever! (2)
回复Matt Bury

Re: Plagiarism detection for GPT LLMs is a bad idea. Change my view.

Carles Aguiló -
Moodle HQ的头像 Particularly helpful Moodlers的头像

Hi Matt,

Thanks for your post.

I believe that the concerns that you raise are valid and Moodle doesn't suggest any other practice than what other experts have publicly shared: if plagiarism concerns are raised, they should be discussed with the student rather than disciplinary action taken.

This said, plagiarism detection works as a deterrent as much as an actual check, and we see it implemented in an array of situations. In my opinion, both generative AI and plagiarism detection should be included in the projects and curriculum of the learning experience, in the sense of asking an LLM as a part of what the actual task for the learner is.

When Moodle certifies an integration, we are not suggesting that any piece of software should be used, but rather that if you want to integrate a tool covering this functionality, a Certified Integration has been vetted in terms of code quality, user experience and accessibility, so the integrator can trust that these tools verify a high standard when it comes to compliance with the coding guidelines at https://moodledev.io/ and Moodle's expectation in terms of user experience and accessibility.

Best, Carles

平均分： -
回复Matt Bury

Re: Plagiarism detection for GPT LLMs is a bad idea. Change my view.

Eduardo Kraus -
Particularly helpful Moodlers的头像 Plugin developers的头像 Testers的头像

Ah, Matt, what a great discussion you’ve brought up!

In fact, many of the current systems are still inaccurate and can generate false positives, which is a big problem for students and teachers. But does that mean we should simply abandon any attempt to identify AI-generated texts?
I’d say no! 🤖

If an AI detector is treated as an absolute and infallible tool, with decisions based solely on the results it generates, then we indeed run a huge risk of injustice. Another important point is that these tools are constantly evolving. What may be inaccurate today could be much more refined tomorrow. We are already seeing detectors that not only analyze linguistic patterns but also take metadata and context into account, reducing the chance of errors.

And do you know what the biggest mistake is? Asking ChatGPT itself whether a text was generated by AI—this simply doesn’t work! ChatGPT has no memory of the texts it generates and cannot accurately determine whether something was written by it or not. I tested newspaper articles from before AI existed, and in some cases, ChatGPT claimed they were AI-generated. Also, ChatGPT was not trained to detect AI-generated content.

Eduardo Kraus
Innovation and new products

平均分： Very cool (1)
回复Eduardo Kraus

Re: Plagiarism detection for GPT LLMs is a bad idea. Change my view.

Matt Bury -
Particularly helpful Moodlers的头像 Plugin developers的头像

I wouldn't say abandon. As you say it's early days. Wouldn't it be better to wait until the AI detectors work within acceptable degrees of accuracy before implementing them on students? Much the same way we'd like our dentists to make sure a new procedure has a minimum probability of success before trying it out on us?
平均分： Very cool (2)
回复Matt Bury

Re: Plagiarism detection for GPT LLMs is a bad idea. Change my view.

Eduardo Kraus -
Particularly helpful Moodlers的头像 Plugin developers的头像 Testers的头像

Hello Matt,

Generative AIs, like ChatGPT and Gemini, are evolving so quickly that any AI detector that seems promising today could be completely obsolete tomorrow. In the past, it was easy to identify AI-generated text: short sentences, always a conclusion structured the same way… But now? I recently tested ZeroGPT, and to my surprise, it classified almost all texts generated by ChatGPT and Gemini as written by humans! 

The big problem is that these detectors work based on statistical patterns, trying to guess if a text was generated by AI based on predictability. But what happens when AI models start varying their writing, making it more and more human-like? We're already seeing that happen! So, is it worth relying on a system that could unfairly harm students and teachers with false positives? 🤔 Instead of relying on flawed detectors, we can teach students to use AI ethically and responsibly, as a tool to enhance learning, not to cheat. Just like calculators in math and Wikipedia in research, AI is here to stay – and we need to learn to live with it!

But there’s a much more effective approach than relying on these flawed systems: investing in e-proctoring with AI 💡! If the goal is to prevent the misuse of AI in assessments, it makes much more sense to create mechanisms that directly prevent cheating. Implementing systems that prevent students from leaving the screen, disable copy/paste, and block the possibility of inserting externally generated text could be a much more efficient solution and stop them from copying data to transfer to the AI. That way, we ensure the integrity of the test without relying on questionable detectors!

Instead of trying to block the advancement of AI, we can focus on smart strategies to adapt our teaching environment and ensure fair assessments! Facebook, for example, has a strange implementation that doesn’t allow copying the post date (why is that???). Imagine this in the exam questions where the student copies the question, but when pasting, it pastes millions of irrelevant characters along with it:

Captura%20de%20Tela%202025-02-04%20a%CC%80s%2009.39.51.png
PS: Screenshot in potuguese, but you can understand that I right-clicked and clicked "search on Google".

Eduardo Kraus
Innovation and new products

平均分： -
回复Eduardo Kraus

Re: Plagiarism detection for GPT LLMs is a bad idea. Change my view.

Matt Bury -
Particularly helpful Moodlers的头像 Plugin developers的头像
My understanding is that although the models are getting larger, they're not improving much. The last I heard was that they had increased the model size by an order of magnitude but only observed marginal improvements. I also read from some researchers, I can't remember who, that GPT LLMs had pretty much reached their theoretical limits.

Anyway, that is all theoretical & speculative as are predictions of "breakthroughs" in models being just around the corner. Personally, I'll believe it when I see it. Until then, how about we give poor students a break & stop falsely accusing them of plagiarism based on unreliable evidence?
平均分： -
回复Matt Bury

Re: Plagiarism detection for GPT LLMs is a bad idea. Change my view.

Marcus Green -
Core developers的头像 Particularly helpful Moodlers的头像 Plugin developers的头像 Testers的头像
"stop falsely accusing them of plagiarism based on unreliable evidence?"

Step 1 may be to get people to understand exactly how unreliable the evidence is. People can believe that something is true because the computer says so.

This has been illustrated by the "Post Office scandal" in the UK, which was partly driven by 

"In 1999, the presumption was introduced into law on how courts should consider electronic evidence. The rule followed a Law Commission recommendation that courts should presume a computer system has operated correctly unless there is explicit evidence to the contrary."

https://www.computerweekly.com/news/366618322/Review-of-legal-rule-on-computer-evidence-long-overdue-say-Post-Office-scandal-victims#:~:text=IT%20scandal%20exposes%20legal%20rule,to%20subpostmasters%20being%20wrongly%20convicted.

This is aggravated by the "black box" nature of AI and where unpredictability/unreliability is baked in.


平均分： -
回复Eduardo Kraus

Re: Plagiarism detection for GPT LLMs is a bad idea. Change my view.

Matt Bury -
Particularly helpful Moodlers的头像 Plugin developers的头像
More effective proctoring & ending open book assignments & exams is becoming more common, as far as I understand it. It'd seem that preventing students from being able to submit GPT LLM generated text is broadly seen as an effective strategy.

It's also my understanding that many university curricula under-prepare & overload students with work to a point where they have to make choices between what to study & what to ignore. See Graham Gibbs' short essay on this topic: https://web.archive.org/web/20160405133040/http://seda.ac.uk/resources/files/publications_175_26%20Students%20are%20selectively%20negligent,%20and%20successful%20students%20neglect%20the%20right%20stuff.pdf In other words, many students turn to academic malpractice as a survival strategy. I doubt any amount of persuasion or guidance will convince such students not to take what they see as necessary shortcuts.

Is it that GPT LLMs are exposing deeper issues in higher education, which have always been there & have always provided advantages for those who can provide ways to game the system in their favour? I very much doubt that the solutions will be found in yet more algorithms.
平均分： Coolest thing ever! (2)
回复Matt Bury

Re: Plagiarism detection for GPT LLMs is a bad idea. Change my view.

Visvanath Ratnaweera -
Particularly helpful Moodlers的头像 Translators的头像
The Magic Bag. It’s a familiar storytelling device. Ask the bag for something, anything, whatever you might want—and poof, out it pops. We find variations in myths and fables, in jokes, in numerous Twilight Zone episodes. The genie in a bottle. The monkey’s paw. The holodeck on the USS Enterprise. The moral of the story, as often as not, turns out to be: Be careful what you wish for. By giving us what we thought we wanted, the Magic Bag instructs us on the danger of having one’s desires fulfilled, reminding us that it is often better to want than to get.
 
I Used to Teach Students. Now I Catch ChatGPT Cheats
I once believed university was a shared intellectual pursuit. That faith has been obliterated
https://thewalrus.ca/i-used-to-teach-students-now-i-catch-chatgpt-cheats.

(Found in: https://news.slashdot.org/story/25/03/09/1929235/i-used-to-teach-students-now-i-catch-chatgpt-cheats.)
 
My hope is that the article is not written by a chat bot. From what I hear about the wonders of these tools, there must be a button "Is this written by ChatGPT?"
平均分： -