Universities offered software to sniff out ChatGPT-written essays Turnitin, best known for its anti-plagiarism software used by tens of thousands of universities and schools around the world, is building a tool to detect text generated by AI. Large language models have gained traction since the commercial release of OpenAI's GPT-3 in 2020. Now multiple companies have built their own rival …
One problem is that a tool like this will only give you a prediction of whether GPT was used. It can say that such a thing is probable, but not that it is certain. If a student denies it and you have a tool which says "computer says yes" but you can't verify it, what will you do? If you catch someone using it to generate work or can prove it with certainty, then treating it as a violation of policies against cheating makes sense, but there will have to be some plan for cases where it's in doubt whether it happened.
I wonder if one might present the results using colour to indicate the probability/confidence of GPT-3 having been used to generate the text? Reprint the essay but with coloured characters, not black on white... The greener, the better; the more red on the page, the greater the concern the teacher might have (other colours are available).
Plagiarism is easier to establish, since you can actually track down (internet) sources that closely match submitted material; this is pretty much what Turnitin already does. Chat GPT use will be much harder to establish, as it doesn't (at least in my experience) simply quote internet sources verbatim.
On the other hand, Chat GPT will also reliably generate rubbish essays, which will be marked accordingly; so students may find it's not such a good cheat.
> Not that they care; they just blunder through and take the software's as the truth.
If by "they" you mean university teaching staff, I am in that game, and can honestly say that my colleagues are admirably conscientious about marking, not so naive as to take "software" as an ultimate truth, and have little difficulty identifying poor-quality work. Besides which, students generally have a right of appeal (at least at my institution).
The key word is caught.
This implies either:
The first would have to be accountable and trustworthy and not just go on hearsay and accusations but be a recorded incident. The second is subjective as there is nothing stopping student's own writing being similar to ChatGPT output and vice-versa.
Therefore rather going further down the route of backwards thinking, it's better to consider that firstly ChatGPT is a tool and secondly that the capabilities of such systems will improve.
ChatGPT output is far from perfect, however how about consider its use as a tool? One can generate the core of a document, the structure as such, and then go through the document and rewrite and restyle this in one's own language and understanding overlaid on top of it? There is little different in this compared to a student taking previous examples on the same topic and using them as templates for their own work. This is different to plagiarism, this is research - both in the subject and ways to present the subject (which is almost as important). Also consider that it's known that many students pay other students or people to write their work for them, therefore the emphasis on ChatGPT is ridiculous.
In the end, the only way to really test a student is going to be do so in a controlled environment. This does not necessarily have to be the pressure of a tightly controlled exam, but could be a much longer process in a controlled environment. Students who relied on something or someone else for their earlier work will not be able to perform, students who used something or someone else to assist them with their work will be discernable by either their understanding or lack of.
>ChatGPT output is far from perfect, however how about consider its use as a tool? One can generate the core of a document, the structure as such
It would be interesting to know just how good ChatGPT is at structure, I suspect this requires coding to embed both the concepts of an argument and a viewpoint from which to argue. Then it would need to know how to structure the delivery (ie. “Tell them what you’re going to tell them, tell them and tell them what you’ve told them”), naturally the exact presentation will depend upon whether the material is to be spoken or written and place of publication. I expect all of these to likewise require coding. Otherwise it would require ChatGPT to 'understand' the meaning and apply the content, ie. That the book "The Pyramid Prinicple" is about structure and thus apply the concepts it contains to the format of its output.
I suspect the real value of ChatGPT would be gain ideas and coverage, ie. when doing research there always is at least one book/source you haven't read, ChatGPT reduces that with respect to stuff published on the Internet, but you will still have to visit a library or two to gain access to primary sources.
So, if the makers of CheatYourWayToADegreeGPT want to evade this, they have to make their language model pick less obvious words. But a lower probability word is more likely to be wrong. And so the tool becomes less useful.
Is this game over? Or can the language models pick rare enough words that they look human to the detectors while still making enough sense for them to be worthwhile to use? Or could you run a filter over the output that substitutes synonyms and fuzzes it to make the language model output look more human?
I guess, we're about to find out!
It might be harder than that...
A question has to be how individual the responses ChatGPT gives are. If they are not personalised we can expect at some point several students on the same course etc. will hand in work that is broadly identical.
Naturally, if a savvy lecturer submitted the question to ChatGPT before they issued it to the students, they would have a reference that will assist them in detecting the use of ChatGPT...
So CheatYourWayToADegreeGPT is going to have to take the ChatGPT results and personalise them - by training it on a user's previous essays?
"anyone with half a brain would only enter what they want in bits and then go over it and make some changes before submitting it"
There's a very simple solution. Universities used to have personal tutorials at which the student had to read their essay and discuss it with the lecturer. If you can't do that, it's not really yours (even if you actually typed it). We've got stuck on checking the writing, whereas we should be concentrating on checking the thinking.
More generally, anyone who's teaching college composition in a manner that's amenable to doing a significant amount of the work using an LLM is way behind on current pedagogy, and frankly deserves to be fooled. Time to start teaching like it's the 21st century. Assigning 5-page essays for homework went out decades ago.
they have to make their language model pick less obvious words
This is so far from being correct, it's not even really wrong. It's just wildly irrelevant.
Transformer LLMs are so far from being bag-of-words models that no BoW (unigram) model will have significantly better-than-random performance at detecting their output.
Given ChatGPT can generate output good enough to fool recruiters:
https://news.sky.com/story/recruitment-team-unwittingly-recommends-chatgpt-for-job-interview-12788770
[The linked articles:
https://news.sky.com/story/the-ultimate-homework-cheat-how-teachers-are-facing-up-to-chatgpt-12780601
https://news.sky.com/story/chatgpt-we-let-an-ai-chatbot-help-write-an-article-heres-how-it-went-12763244
are also worth a read.]
I would want to test this tool against such output.
However, I suspect this will be an ongoing arms race as ChatGPT steadily improves, so the only natural way forward will be for Universities to increase "contact time" with on-going in-person viva style investigation of understanding and research.
This is certainly an effective countermeasure for would-be cheaters, if they have access to detectors. And restricting access to detectors will be very hard indeed. How many starving grad-student TAs would be willing to let the frat chapter next door use their account for a little extra pocket money? A whole bunch of them. (We already know many existing paper-mill staff are graduate students and adjunct faculty, who don't make enough to live on otherwise.)
It is something I have considered... if the grammar (and spelling) is predicated on that which might be found on the internet *in general*, then surely one might expect such a model to mimic those spelling/grammar errors?
Unless someone has carefully curated the information from which ChatGPT has been taught? And if so, who, and how?
It's better than that of several humans, in fact. Other humans can best it. Did you have a particular human in mind?
More seriously, "grammar" isn't an interesting criterion here. What you likely mean is "usage and mechanics", and even that isn't a particularly useful metric, because English usage and mechanics aren't standardized. Audiences judge them according to a wide range of expectations, including not just their own personal expectations but contextual factors such as occasion, tone, and speech community. As with other matters of prose style, LLMs are generally trained on large corpora so they can generate sequences in a large variety of styles, and for a specific output style will be heavily influenced by the prompt, or succession of prompts if the user is making a decent effort.
Turnitin just isn't that great. But if they and the universities just go, 'computer says no', then they can brute force the sense that something is being done.
That there are only so many combinations of words and sentences one can have for a given topic, and there are new university students every year, then increasingly more essays and papers are going to be marked as 'plagiarism'. Some will be plagiarism, some won't.
"The Turnitin results show where the suspected plagiarization comes from, along with links in many cases"
That's only detecting the crudest and most incompetent form of plagiarism. The more sophisticated form is the use of other people's ideas without acknowledgement. AI ain't going to find that very easily.
If we've come down to the assumption that university students are habitually copying and pasting content from third parties instead of thinking for themselves, education is truly deceased.
>all we got was "Computer says no". Not sure if that's Turnitin's fault or the Uni
"When you use Grammarly’s free online plagiarism check to detect plagiarism, you’ll see an instant report that tells you whether or not plagiarism was found
...
Grammarly’s Premium plagiarism checker flags specific sentences and provides reference information about the source"
I assume Turnitin have a similar free and subscription services, hence suggest someone was not paying...
Personally, I would expect the Uni to have the subscription version just so that they have the evidence to support their judgement.
Now this smells like snake oil. Turnitin and co. are taking education institutions for a ride.
Detecting plagiarism is one thing (and having seen Turnitin results, they aren't even perfect at that). Detecting a language learning model that is designed to sound natural? Yeah good luck without penalising innocent people.
And to avoid it, all someone needs to do is read the output and put in a few edits.
Education institutions do need to adapt to this. The solutions are all available in-house and don't even have to cost extra (so do, but this ain't it). Many of those institutions won't like making the necessary changes (do they ever?), but they will have to. How much they waste on services like Turnitin's is up to them.
I suspect that to reduce the match to completely non-suspicious you'd have to do more than a few edits; you'd have to reword a substantial fraction of the sentences. At which point, assuming the GPT-generated result was not full of wrong, your learning objective may have been achieved. But are students using GPT output likely to check it for correctness?
That really depends on the course. Though one could argue that where LLMs would would work best will also likely remove many jobs in those areas and make degrees in them a lot less useful (other than personal knowledge enrichment).
LLMs do take a lot of the (tedious) work out the equation. To do well, you'd still need to have knowledge in the area and spend time checking any output, plus adding what you need. You'd also need to craft your responses.
The point is that some (many?) education institutions seem to be taking a rather simplistic approach of 'lets ban it' or something close to a ban. When LLMs aren't going to go away, that's just a waste of energy and money. LLMs could instead be integrated into courses and in the process teaching (haha, universities teaching) how to identify problems with them and the morals of using them. And they could actually assess students properly for a changeb something only really PhD students get the luxury of.
Checking for plagiarism is still necessary even if they also often use abhammer to enforce it.
I was chatting to a friend the other day who works in IT about this. They said that although ChatGPT is a concern, it's not a total dealbreaker for them. A key point of essays isn't getting the right answer but how you reach your conclusion: The sources you cite, the way you structure your essay and link all the ideas together.
And as if my magic, Sixty Symbols has just done a video on this exact topic: www.youtube.com/watch?v=GBtfwa-Fexc.
For those who can't be bothered to watch the video: ChatGPT fails to understand physics concepts.
So here is a silly thought, what if the people creating the ChatGPT stuff and the people writing the TurnitIn detector stuff are the same people? They get paid to create the essay and paid to scan it and then they adjust the detection results to maximise their income. One year they lean towards a higher detection rate and the next year they offer an upgraded ChatGPT which squeaks more essays through. Then the next year the detector is upgraded which catches more ChatGPT essays...
They could probably run that for decades before anyone bothered to check
But you can't complain that ChatGBT is able to create study-schedule, pick the right literature, tell you all the right answers and explains it all to you, It is a powerfull tool and it's dunb to underestimate it. If you're looking for something written by humans I advise you look somewhere else.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
