Plagiarism-sniffing Turnitin tries to find AI writing by students – with mixed grades

Turnitin, which touts itself as a maker of plagiarism-detecting software for academia, demoed on Tuesday what we're told is a tool capable of detecting AI writing. Crucially, it boasted its machine learning software will flag up computer-generated cheating when it has at least "98 percent confidence" it is right. Be as that …

[DS999]

ChatGPT

And if you ask ChatGPT to write you something that Turnitin won't detect I imagine it will be able to do so - if via no other means than brute force method of writing 50 different versions and choosing the one that is part of the 2% Turnitin can't detect!

[Anonymous Coward]

Re: ChatGPT

ChatGPT can't learn from Turnitin, in fact ChatGPT can't learn anything, but that's another topic. Even if it could, you'd be better off feeding ChatGPT results to Turnitin via script to keep ChatGPT iterating variations as fast as possible. Even if ChatGPT could learn it would become slower and bloated by what a separate system could perform and doing that isn't exactly "intelligent".

[ahabilal]

Re: ChatGPT

The sibling model of InstructGPT, taught to follow instructions in prompts and provide thorough responses, is ChatGPT.

We are eager to launch ChatGPT in order to gather user input and understand its advantages and disadvantages. ChatGPT is available free of charge during the study preview.

Without additional information, it's challenging to pinpoint the code's flaws. Could you elaborate on what the code is meant to accomplish and what isn't going as planned? Is this a complete copy of the code, or just a portion of it?

only a portion of it; the mistake never manifests. I believe there is a problem with the service.

Without knowing more about what the code is intended to do and what occurs when it is executed, it is difficult to tell. One possible problem with the code you offered is that the resultWorkerErr channel is never ended, which means that if the resultWorkerErr channel is never written to, the code might hang. This might occur if b.resultWorker cancels before returning an error or if it never returns an error.

You could stop the resultWorkerErr channel after writing to it to resolve this problem. For instance, after the line of code that transmits the error to the channel, you could append the following line:

[Anonymous Coward]

It's not clear how Turnitin verified the accuracy of its software ...

You do what it takes to qualify for insider developer status.

[ChoHag]

> detect AI-generated text with "98 percent confidence" – but has failed to provide any evidence of this.

Did they copy someone else's work without understanding it and are now furiously backpedalling because they're in front of the class being asked questions?

> They need to be able to detect AI with very high certainty to assess the authenticity of a student's work and determine how to best engage with them

> determine how to best engage with them...

I think I know the answer to this one!

[Anonymous Cowherder]

Looks like some AI help with the headline might've helped correct the typo? Missing I in Turnitin resulting in Turntin

[Great Bu]

No, they are using a Belgian detective boy and his dog to detect cheating.

[imanidiot]

pointless

IMHO if a teacher relies solely on a written essay that can be written easily by ChatGPT it means that what they're teaching is pointless. Either do verbal exams or build exams and homework requirements such that it would be hard to get ChatGPT to do all of it.

If the essay can be fully automated by ChatGPT it likely means the entire subject can be automated (and removed from human jobs) and it's pointless to force those who would use ChatGPT to write their essays to learn it.

[Tom 38]

Re: pointless

Academic study is built on layers. My physics teacher each year would carefully explain that everything he was going to teach that year was mostly a lie - that fuller understanding of the concept was beyond us at that point in our development. You have to do the basic steps before you can get to the big steps.

So, yeah, probably 90% of secondary school education could be regurgitated by a ChatGPT bot, does that make it worthless? No. Unlike the AI, humans learn by studying. In order for those humans to reach the level above the AI, they have to study things at lower levels.

Otherwise, we're only a few years away from Brawndo: its got what plants crave

[imanidiot]

Re: pointless

I don't disagree with that, but on the other hand you're not going to get kids who don't want to do the work to retain any of the knowledge by forcing them to write an essay. IMHO essays on low level topics are just a bad way to gauge a students actual knowledge and understanding of a subject. Especially those who do not do well with the written word.

[doublelayer]

Re: pointless

Depending on the level, the essay might have two purposes: making sure they understand the topic and helping them improve their writing. They'll eventually have to write something to express their ideas. Those who do not do well with the written word probably need to learn more about how to write well, because in our world, it's rarely optional. That's not universally true, but low-level education is trying to train everyone in general skills, and writing is one of the most general ones out there.

"you're not going to get kids who don't want to do the work to retain any of the knowledge by forcing them to write an essay."

How are you going to get them to retain the knowledge? Some kids are lazy about doing the work in whatever form but understand it, and that's great, but for those who don't want to do the work and don't understand it by themselves, do you have a plan for checking whether they understand it and fixing this problem? I don't think homework in a different form is going to change things very much. You'll have students who don't want to present in front of a class, don't want to have one-on-one discussions with teachers about the subject, and don't want to do the homework assignments. That's a problem with not wanting to do the work, not an indication that we need to change what the work is. There are cases where the essay is poorer than alternative methods, and change makes sense in that case, but not just to deal with students who don't want to do it.

[veti]

Re: pointless

Writing school essays is nothing to do with "retaining knowledge". Nobody cares what the student can or can't remember a few weeks later about the migration of the bar-tailed godwit, or whatever.

It's about developing the student's ability to collect, collate and present information. That's a skill that needs to be practised.

[nijam]

Re: pointless

> Writing school essays is nothing to do with "retaining knowledge".

Too right! As an aside, I maintain that the popularity of "Game Of Thrones" was down to it being very much like actual history [1] [2], but without the pointless drudgery of essays.

[1] Bloodthirsty, backstabbing, etc, etc.

[2] OK, very few history textbooks mention the dragons...

[tekHedd]

Why are you in school again?

This objection implies (perhaps unintentionally) that the point of writing the essay is to get a grade, that the *grade* is the point of taking a class. The point is learning, the grade is incidental. You have to grade to show that somebody did the work, but if they don't want to learn, they won't. People are excellent at gaming the system. Hack the world! If you fail and get caught, that's also a valuable life lesson. If you do it effectively and learn nothing, well, that's your loss.

[doublelayer]

Re: pointless

"If the essay can be fully automated by ChatGPT it likely means the entire subject can be automated (and removed from human jobs)"

Yep, definitely. Writing an essay describing how the cleaning process of a historical oil spill went wrong can be automated, so let's just let the AI figure out how to do the next one. The point of the work isn't writing it up, but understanding what didn't work, being able to explain why it didn't work, and being able to use that to figure out what will work later. Why does C4 explode? Easy to automate, but future chemists need to know that not so they can answer a quiz question, but so that their new compound that hasn't been described online yet doesn't explode unless they wanted it to. Other examples are available.

It's been that way no matter what tool was used. A simple calculator can automate all the primary school arithmetic assignments as long as you can explain to a student that you read from left to right, push all the buttons that have the symbols, then write down the thing the calculator says. If you let them use a more complex calculator, it can handle the next few levels as well without much human input. This would be great for jobs that just involve adding up numbers, but spreadsheets eliminated most of those. We still teach it so that the people know what numbers and operations to put into the spreadsheets, because numerical jobs don't involve a list of formulas people just want solved.

[Mike 137]

A rather important question

What remains to be answered is why lecturers (who are supposedly experts in the relevant subjects) aren't trusted any more to identify plagiarism. The most common characteristic of chatbot generated text is that it's bland and lacking in originality (for obvious reasons -- it's a collage of existing material). Is that the current expectation of students' essays? If so, there's something far more wrong than students' use of chat bots -- it's that the fundamental purpose of education itself (to facilitate individual growth of understanding) has been abandoned.

[imanidiot]

Re: A rather important question

Have you read the average student essay? For lower level subjects it's very unlikely they'll be anything but indistinguishable from AI generated "collage of existing materials" because for the most parts that's how a human student writes too.

[nijam]

Re: A rather important question

> ... because for the most parts that's how a human student writes too.

In turn, because that's how they are taught, dare I suggest?

[MrBanana]

Re: A rather important question

Identifying plagiarism is a non-trivial task. When the possible subject matter space is large, how can a lecturer be able to read and remember every possible specific text that may have been copied and pasted by a student. Turnitin has been a good tool for for identifying such obvious plagiarism - sadly, many students are just too lazy to present the copied text as a referenced citation. AI is a new problem. It may be possible to identify AI generated text, but it can easily be used as a template and just modified so that it is not recognised as AI output.

This is being likened to calculators destroying maths exams. I don't think it is quite the same, but it does require some rethinking of the way teaching assessments will be judged for essay style answers.

[doublelayer]

Re: A rather important question

"the fundamental purpose of education itself (to facilitate individual growth of understanding) has been abandoned."

When you're growing individuals' understanding, they start by not understanding something that everyone in the field will need to understand. You need to teach them the basics and then verify that they understand them. Let's take computer science as an example. When you're doing early CS education, the basic code the students write isn't groundbreaking, because they're demonstrating that they understand basic concepts. One student's recursive factorial function will look a lot like everyone else's. Even at higher levels, students doing the same assignment will end up producing similar results. I graded a bunch of students' basic HTTP servers in C. Yes, they all looked a little different, but nobody came up with a radical reinterpretation of how to implement a protocol as simple as HTTP. They were demonstrating that they could write a program that implemented a spec, manage network resources to deal with a bunch of incoming connections, and didn't use up too many resources in doing so. Had they gone to any of the other schools that has an HTTP server as an assignment, they could have gotten someone else's code and I wouldn't have been able to instantly detect it. That's it.

Essays can be similar. Not all of them are new original research. Some of them are intended to check that the student knows about the topic they're talking about, has correctly interpreted what they've read, and can get results from that. Eventually, they will need to do more original work than that, but not at the earlier levels where demonstrating understanding and the ability to describe what they know is being tested. As things get more complex, sending a prompt through GPT will become worse and worse. Even for earlier essays, GPT is liable to end up spitting out wrong answers and causing the cheaters to submit flawed essays. That isn't guaranteed, though, which is why chatbot-generated essays are still an issue.

[An_Old_Dog]

Recognizing AI/ML writing

I recently did a Google search to learn what GaN electronic components were. I read the first hit, and got the feeling, "This feels like artificial writing." Judge for yourself:

https://www.electronics-notes.com/articles/electronic_components/fet-field-effect-transistor/gallium-nitride-gan-fet-hemt.php

I went to the second hit, a whitepaper by Renesas, and in reading it, quickly got the answer I was looking for.

Some people probably don't have a "feel" (mental pattern-matching) for non-human writing, just as some people are horrible at inventing clear, meaningful names. (2DEG is a recent worst: Two-Dimensionsal Electron Gas is not two-dimensional, nor is it a gas ...)

[DrBobK]

ChatGPT isn't good enough yet

My experience, from asking chatGPT to generate answers to various exam questions I have set, is that chatGPT answers would get 2ii marks at best, and that it is pretty obvious that they were not written by a human, or at least someone with a good grasp of structuring an argument and use of evidence. I've also tried software intended to detect machine generated text. It always classified the chatGPT output as more the 97% likely to be machine generated. If, however, I rewrote the first two or three sentences (so a very small portion of the essay) the ratings changed to 2 or 3% likely to be machine generated (this wasn't Turinitin, but it was meant to be the best chatGPT detector at the time). Personally, I think if students want to use chatGPT they should be allowed to. They'd have to do so much work editing the chatGPT output that a good answer would still demonstrate a student's understanding of a topic and their use and knowledge of appropriate evidence in supporting an argument. Yours truly, an old Professor.

[yetanotheraoc]

Re: ChatGPT isn't good enough yet

"Personally, I think if students want to use chatGPT they should be allowed to. They'd have to do so much work editing the chatGPT output..."

Bit of a disconnect here. What is the meaning of "allowed"? Are you suggesting they are "allowed", and yet somehow need to do some work to evade the detection? If not the second part, then they wouldn't have to do *any* work editing the ChatGT output.

[doublelayer]

Re: ChatGPT isn't good enough yet

I think the suggestion is that, because the raw GPT output wouldn't receive a good grade because it is too bad, that students could start with that and modify it into a good essay. The modification would be equivalent to the work required to write the essay from scratch. This works well as long as the GPT output is really that bad, because any student who knows what they're doing could create a valid essay and would probably realize the GPT output wasn't helpful. It breaks down if GPT can generate working answers for some questions without requiring the student to do a complete rewrite, which might mean that the proposed policy is only useful in the classes the poster teaches and not necessarily anyone else's.

[doublelayer]

Re: ChatGPT isn't good enough yet

"My experience, from asking chatGPT to generate answers to various exam questions I have set, is that chatGPT answers would get 2ii marks at best, and that it is pretty obvious that they were not written by a human, or at least someone with a good grasp of structuring an argument and use of evidence."

I suggest that this may be due to a small sample set. If you are setting more advanced exams, your questions may be the kind that a chatbot can't answer plausibly. This will affect the lower level classes first, because the answers in those cases are simpler to generate without really knowing the content. Any student that takes something sufficiently difficult will find that GPT is no longer useful, but it could create problems for those taking prerequisites where the questions are easier. I'd expect this to be faced by teachers of younger students before they get to you.

[Simon Harris]

Fail

“the features now integrated into its existing anti-plagiarism products can detect AI-generated text with "98 percent confidence" – but has failed to provide any evidence of this.”

As someone who will soon be receiving a pile of student dissertations in his inbox to mark (which, incidentally, are screened by TurnItIn), if such a statement appeared in their dissertations without supporting evidence, that would be a fail from me.

[Version 1.0]

Re: Fail

You can avoid AI problems by just requiring that everyone writes with a quill, that will give them some excellent physical abilities plus plenty of time to think about what they are writing.

[Simon Harris]

Re: Fail

I will insist on nothing less than illuminated manuscript from my students, to include humorous satire. I will, of course, take care not to lick my fingers while turning the pages.

[veti]

Re: Fail

Next challenge?

[An_Old_Dog]

Writing "Manually"

No need to resort to the quill and inkpot. Give them VT-100s connected to an RPi, running some flavor of Unix. They will have a choice of editors: vi, vim, emacs, and ed. (Ed is the standard editor.)

[dajames]

Doesn't add up!

To maintain a less than one percent false positive rate, we only flag something when we are 98 percent sure it is written by AI ...

It's hard to be sure with marketroid BS like that, but I think they're saying that 98% plus 1% is 100% ???

[doublelayer]

Re: Doesn't add up!

I think it's two independent statistics. Their model probably provides a confidence number based on how likely it thinks the text is to be AI-generated. That number is generated by an opaque process and isn't necessarily correct. They've looked at how well the model is predicting this, but not to try improving the results. Instead, they just ask themselves how high their threshold needs to be to have a marketable statistic for accuracy.

They need to do this because, if their false positive rate is too high, they'll have universities and students alike protesting at their offices. Students told that their self-written essays have been rejected because a computer decided it might be AI-generated and there's no way to prove that it wasn't are not going to be in a charitable mood, and they would be right about that. Universities don't want students to get that angry. To keep this rate down, it's likely that this software is letting all sorts of AI-generated stuff through on the theory that it's better to let a bunch of guilty students through than to punish more than 1% of innocent ones. I suppose I have to agree that false positives are a lot worse than false negatives, but I'm dubious that this software is great on either.

[Innominate Chicken]

Re: Doesn't add up!

As someone who's submitted work using Turnitin, I can confirm that it's an extremely rudimentary program. It appears to simply compare text with no consideration of context.

Turnitin usually gives you a report showing exactly what text it's flagged, and I've seen it take issue with things like standard equations (which have a generally accepted way of writing them), quotes (which are obviously copied because it's a quote) and citations (other people cited the same work this is clearly a match.)

I can forgive the quotes since that's less obvious to detect, but typed maths is very good obvious, and citations are usually in a format that should be able to be picked up and ignored.

[lafnlab]

They must not read El Reg

Otherwise they would have learned there is no reliable way of detecting AI-Generated text

https://www.theregister.com/2023/03/21/detecting_ai_generated_text/

.

[TheMaskedMan]

Easy. If it's grammatically correct and contains no typos, it's probably chatGPT. Problem solved:)

I do see the point, though. As a student I never saw the point in cheating - I wanted to learn, and I wanted to know how well - or otherwise - I was progressing. Some of my colleagues took an alternative view :(

People do learn through study and writing about what they studied - I still remember a few useless facts from essays I wrote at school. And if they want an academic career, it's all about writing papers - might as well get the hang of it as soon as possible.

That said, there is a learnable technique for writing essays - and even short, saleable articles - without having to become an expert in the subject. Just because a student turns in a good, original essay doesn't mean they have a deep understanding of the subject.

It's a tricky problem. Perhaps you could work around it by setting a multiple choice quiz with fairly in depth questions, then using that as the basis for an essay. That would take extra time, though. Or give them a prewritten essay and ask them to find and correct the factual inaccuracies in the form of a new essay. All of which sounds complicated - I'm glad I don't have to solve this problem for a living!

[Electric Panda]

The key to this is the use of simple linguistic flair: you must have the skills to reword what the AI tells you in a way that sounds plausibly human. Use a thesaurus, change the word order, the punctuation, the voice, the tenses (if appropriate), contractions, possessives etc.

This - and simply referencing the legitimate sources at the bottom of the articles - was how we got around the Wikipedia 'ban' at university.

[An_Old_Dog]

Saving Time??

Did doing all that really save you time??!

[doublelayer]

Re: Saving Time??

It might have, since it removes the part of the process requiring original thought. I can't say I've done something like this, but the closest I came was in a literature class I took as a young student. The teacher of this class was not very good, and whenever they assigned a literary analysis assignment, original thought was not recommended. You could come up with an original idea for something contained in a book, rigorously defend it with quotations, comparative analyses of other books by the author, verifiable or likely references contained in the source material, but if it wasn't the interpretation the teacher thought of you would still score poorly. Meanwhile, if I listened to what the teacher said the theme was, paraphrased it, and backed it up with the first quotes that came to my mind, I'd get a good grade. Either that or I'm still deluding myself that some of my interpretations had value and I was always terrible at it.

In that case, I was saving time as a side-effect of writing to get the highest grade rather than the other way around, but it was more efficient to figure out what the simple answer was and write for that rather than to think of a novel one and find the information to demonstrate it. The work was reduced from thinking, proving, and writing to just writing. I was less happy about the result, but it took less time to get it.

[Falmari]

AI or Human Author?

Lets say this software is able to accurately detect essays that are written by AI. That's not the same as being able to detect if the student wrote the essay.

When academic qualifications rely in some part on course work then the grader needs to have confidence that it is all the students own work, in other words be able to recognise the students work. If software is needed to check for AI generated work then that is not the case.

The question is not, is this essay the work of AI but is it the work of the student. Any academic course that can be cheated by the student using AI to generate essays has always had that problem, the essay might not be the students own work.

[spold]

Same old....

Add some text, make the font tiny, and white text on white background. Just as you would do for a job application that is going to be processed by some moronic screening application (copy all the job requirements and paste them in invisibly).

[plrndl]

About Time

This is a clear demonstration of the pointlessness of testing students' ability to regurgitate what they have been taught.

Maybe it's time to start testing their ability to do something useful with what they have learned.

Unfortunately this may result in academics having to do some real work.

[MachDiamond]

TestGPT

Profs could take the work submitted by the students and have an AI system generate individual exams based on those papers. It doesn't prevent students from using AI to help with their papers, but they better understand the material (which is the whole point) or they'll get slaughtered at test which will call into question their assignments. Of course exams will also have questions thought up by the teacher that covers material in the text and lectures to go with that. I don't see that this would be too hard to implement. There's some work up front to set it up, but once in place it would scale to any number of students and any subject that uses a similar teaching method.