Do AI detectors work? Students face false cheating accusations

However programming a computer to do this isn't easy. In a previous job I had dealing with plagiarism detectors and soon realised how garbage they were (and also how easily fooled they are - but that is another story). The staff soon realised what garbage these tools are so if a student accused of plagiarism decided to argue back then the accusation would be quietly dropped.

Easy to catch people that aren't trying in the slightest not to get caught, right? I could instead feed a corpus of my own writing to ChatGPT and ask it to write in my style.

I assume most people would make least amount of effort and simply prompt chat interface to produce some text, such text is rather detectable. I would like to see some experiments even for this type of detection though.

The point of these detectors is to spot stuff the students didn’t research and write themselves. But if the corpus is your own written material then you’ve already done the work yourself.

Student don't have to perform research or study for the given task, they need to acquire an example of text suitable for reproducing their style, text structure, to create an impression of being produced by hand, so the original task could be avoided. You have to have at least one corpus of your own work for this to work, or an adequate substitute. And you still could reject works by their content, but we are specifically talking about llm smell.

I was talking about the task of detecting llm generated text which is incredibly hard if any effort is made, while some people have an impression that it's trivially easy. It leads to unfair outcomes while giving false confidence to e.g. teachers that llms are adequately accounted for.

I ask myself when the time comes that some student will accuse the stuff of libel or slander becuase of false AI plagiarism accusations.

So far. Unless there is a new generation of teachers who are no longer able to learn on non-AI generated texts because all they get is grammatically corrected by AI for example...

Even I am using Grammarly here (as being non-native), but I usually tend to ignore it, because it removes all my "spoken" style, or at least what I think is a "spoken style"

So do humans. Many people have pet phrases or words that they use unusually often compared to others.

(Makes me worry I'm not paying enough attention, that I can't).

It would also identify random 5-6 word phrases and attribute them to different random texts on completely different topics where those same 5 words happened to appear.

I forgot the name of the guy, who said it, but he was some big philosophy lecturer at Harvard and his view on the matter ( heavy reading course and one student left a course review - "not reading assigned reading did not hurt me at all") was ( paraphrased):

"This guy is an idiot if he thinks the point of paying $60k a semester of parents money is to sit here and learn nothing.'

//

And no, those were not generally followed by a real comment.

It seems likely that a suitably disciplined student could look a lot like ChatGPT and the cost of a false accusation is extremely high.

Low light scenarios are just a thing, you would need more expensive hardware do deal with it.

That’s the problem. It is trivially easy, 99% of the time. But that misses the entire point of the article.

If I got 99% on an exam I’d say that was trivially easy. But making one mistake in a hundred is not ok when it’s someone else’s livelihood.

Could it be mostly just be..reality? More expensive hardware doesn't somehow make a darker surface reflect more energy in the visible spectrum. "Low light" is not the same condition as "dark surface in well-lit environment."

Leaving the visible spectrum is one possible solution, but it's substantially more error-prone and costly. This is still not the same solution as classical CV with "more expensive hardware."

To get an LLM to generate content in your own writing, there’s going to be no substitute for training it on your own corpus. By which point you might as well do the work yourself.

The whole point cheating is to avoid doing the work. Building your own corpus requires doing that work.

Also, this could have been handled much better in the cases that came up in the media if there had been proper human review of all cases before prosecuting the students.

But as a second language learner, you notice that people get stuck on particular words during writing sessions. If I run into a very unusual (and unnecessary) word, I know they're going to use it again within a page or two, maybe once after that, then never again.

I blame it on the writer remembering a cool word, or finding a cool word in a thesaurus, then that word dropping out of their active vocabulary after they tried it out a couple times. There's probably an analogue in LLMs, if just because that makes unusual words more likely to repeat themselves in a particular passage.

You're saying it's not good enough at mimicking styles. others saying it's good enough. I think if it's not good enough today it'll be good enough tomorrow. Are you betting on it not becoming good enough?

Teachers will notice if students writing style shifts in one piece compared to another.

Nobody disputes that you can get LLMs to mimic other people. However it cannot mimic a specific style it hasn’t been trained on. And very few people who are going to cheat are going to take the time to train an LLM on their writing style since the entire point of plagiarism is to avoid doing work.

I could go on like this, but I won't. Each of us has a choice how we play the cards we are dealt.

I accept your point, but this point reinforces a perspective I heard from my accountant family member, who clearly can identify price, but has a hard time not equating it with value. I hesitate to use the word wrong, because it is pragmatic, but it is also rather wasteful ( if not outright dumb ).

"You asked us to write an essay on the Civil War. The first thing I did was ask an AI to explain it to me, and I asked the AI some follow-up questions. Then I did some research using other sources and wrote my paper."

It might even be a true story, and in such a case it's not surprising that the student would repeat words they encountered while studying.

I've seen exactly what you describe and worse *, and I've also seen them keep to one style until I got bored of prompting for new features to add to the project.

* one standard test I have is "make a tetris game as a single page web app", and one model started wrong and then suddenly flipped from Tetris in html/js to ML in python.

This was at a long-established mall shop that specialized in photography products and services. The same photographer had taken suitable photos of some other people in line ahead of us rather quickly.

The studio area was professional enough, with a backdrop, with dedicated photography lighting, with ample lighting in the shop beyond that, and with an adjustable stool for the subject to sit on.

The camera appeared to be a DSLR with a lens and a lens hood, similar enough to what I've seen professional wedding photographers use. It was initially on a tripod, although the photographer eventually removed it during later attempts.

Despite being in a highly-controlled purpose-built environment, and using photography equipment much better than that of a typical laptop or phone camera, the photographer still couldn't take a suitable photo of this particular woman, despite repeated attempts and adjustments to the camera's settings and to the environment.

Was the photographer "racist"? I would guess not, given the effort he put in, and the frustration he was exhibiting at the lack of success.

Was the camera "racist"? No, obviously not.

Sometimes it can just be difficult to take a suitable photo, even when using higher-end equipment in a rather ideal environment.

It has nothing to do with "racism".

I don't think anyone is saying that the universities or the software companies have some kind of secret agenda to keep black people out. As far as I can tell there's good evidence they're mostly trying to get more black people in (and in some cases to keep Asians out, but that's another story). I also don't think anyone here was acting out of fear or hatred of black people.

What I am claiming is that the universities in question ended up with a proctoring product that was more likely to produce false positives for students with darker skin colors, and did not apply sufficient human review and/or giving people the benefit of the doubt to cancel out those effects. It is quite likely that whatever model-training and testing the software companies did, was mostly on fair-skinned people in well-lit environments, otherwise they would have picked up this problem earlier on. This is not super-woke Ibram X Kendi applied antiracism, this is doing your job properly to make sure your product works for all students, especially as the students don't have any choice to opt out of using the proctoring software beyond quitting their college.

To me it's on the same level as having a SQL injection vulnerability: maybe you didn't intend to get your users' data exposed - about 100% of the time when this happens, the company involved very much did not intend to have a data breach - but it happened anyway, you were incompetent at the job and your users are now dealing with the consequences.

And to the extent that those consequences here fall disproportionately on skin colors (and so, by correlation, on ethnicities) that have historically been disadvantaged, calling this a type of racism seems appropriate. It's very much not the KKK type of racism, but it could very well still meet legal standards for discrimination.

LLMs already discriminates against African-American English. You could argue a human grader would as well, but all tested models were more consistent in assigning negative adjectives to hypothetical speakers of that dialect.

Dodging the detector is done regardless of whether or not one has used AI to write that paper.

Imagine the teacher saying "this is not your style it's too good" to a student who legit tried killing any motivation to do anything but cheat for remaining life

If the student always uses LLMs then it would be pretty obvious by the fact that they’re failing at the cause in all bar the written assessments (ie the stuff they can cheat on).

> Also do you expect that student's style is fixed forever

Of course not. But people’s styles don’t change dramatically on one paper and reset back afterwards.

> teachers are all so invested that they can really tell when the student is trying something new vs use an LLM that was trained to output writing in the style of an average student?

Depends on the size of the classes. When I was at college I do know that teachers did check for changes in writing styles. I know this because one of the kids on my class was questioned about his changes in his writing style.

With time, I’m sure anti-cheat software will also check again previous works by the students to check for changes in style.

However this was never my point. My point was that cheaters wouldn’t bother training on their own corpus. You keep pushing the conversation away from that.

> Imagine the teacher saying "this is not your style it's too good" to a student who legit tried killing any motivation to do anything but cheat for remaining life

That’s how literally no good teacher would ever approach the subject. Instead they’d talk about how good the paper was and ask about where the inspiration came from.

Every major system in the US academic system is aimed to reducing Asian population. It often comes in the guise of DEI with a very wide definition of "Diversity" that rarely includes Asian.

These systems will use subtle features to blackbox racism. They may just be overt and leak over metadata to achieve it, or get smart and using writing styles.

The issue is that, for most people, the term "racism" connotes a moral failing comparable to the secret agendas, fear and hatred, etc. Specifically, an immoral act motivated by a deliberately applied, irrational prejudice.

Using it to refer to this sort of "disparate impact" is at best needlessly vague, and at worst a deliberate conflation known to be useful to (and used by) the "super-woke Ibram X Kendi" types - equivocating (per https://en.wikipedia.org/wiki/Motte-and-bailey_fallacy) in order to attach the spectre of moral outrage to a problem not caused by any kind of malice.

If you're interested in whether someone might have a legal case, you should be discussing that in an appropriate forum - not with lay language among laypeople.

  For a human who deals with student work or reads job applications spotting AI generated work quickly becomes trivially easy.

> Despite being in a highly-controlled purpose-built environment

Frankly it sounds like the environment was not purpose-built at all. It was built to meet insufficient standards, perhaps.

There's nothing stopping students from generating an essay and going over it.

>Of course not. But people’s styles don’t change dramatically on one paper and reset back afterwards.

Takes just a little effort to avoid this.

>With time, I’m sure anti-cheat software will also check again previous works by the students to check for changes in style.

That's never going to happen. Probably because it doesn't make any sense. What's a change in writing style ? Who's measuring that ? And why is that an indicator of cheating ?

>However this was never my point. My point was that cheaters wouldn’t bother training on their own corpus. You keep pushing the conversation away from that.

Training is not necessary in any technical sense. A decent sample of your writing in the context is more than good enough. Probably most cheaters wouldn't bother but some certainly would.

This then comes back to my original point. If they learn the content and rewrite the output, is it really plagiarism?

> Takes just a little effort to avoid this.

That depends entirely on the size of the coursework.

> That's never going to happen. Probably because it doesn't make any sense. What's a change in writing style ? Who's measuring that ? And why is that an indicator of cheating ?

This entire article and all the conversations that followed are about using writing styles to spot plagiarism. It’s not a new concept nor a claim I made up.

So if you don’t agree with this premise then it’s a little late in the thread to be raising that disagreement.

> Training is not necessary in any technical sense. A decent sample of your writing in the context is more than good enough. Probably most cheaters wouldn't bother but some certainly would.

I think you’d need a larger corpus than the average cheater would be bothered to do. But I will admit I could be waaay off in my estimations of this.

Who said anything about rewriting? That's not necessary. You can have GPT write your essay and all you do is study it afterwards, maybe ask questions etc. You've saved hours of time and yes that would still be cheating and plagiarism by most.

>This entire article and all the conversations that followed are about using writing styles to spot plagiarism. It’s not a new concept nor a claim I made up.

>So if you don’t agree with this premise then it’s a little late in the thread to be raising that disagreement.

The article is about piping essays into black box neural networks that you can at best hypothesize is looking for similarities between the presented writing and some nebulous "AI" style. It's not comparing styles between your past works and telling you just cheated because of some deviation. That's never going to happen.

>I think you’d need a larger corpus than the average cheater would be bothered to do. But I will admit I could be waaay off in my estimations of this.

An essay or two in the context window is fine. I think you underestimate just what SOTA LLMs are capable of.

You don't even need to bother with any of that if all you want is a consistent style. A style prompt with a few instructions to deviate from GPT's default writing style is sufficient.

My point is that it's not this huge effort to have generated writing that doesn't yo-yo in writing style between essays.

NB: I am not African-American, nor did I grew up on an African-American community, and I performed very well on all sorts of verbal tests. Yet, even I made the all intensive purposes mistake until well into adulthood. Probably a Midwestern thing.

Maybe. But I think we are getting too deep into hypotheticals about stuff that wasn’t even related to my original point.

> The article is about piping essays into black box neural networks that you can at best hypothesize is looking for similarities between the presented writing and some nebulous "AI" style. It's not comparing styles between your past works and telling you just cheated because of some deviation. That's never going to happen.

You cannot postulate your own hypothetical scenarios and deny other people the same privilege. That’s just not an honest way to debate.

> My point is that it's not this huge effort to have generated writing that doesn't yo-yo in writing style between essays.

I get your point. It’s just your point requires a bunch of assumptions and hypotheticals to work.

In theory you’re right. But, and at risk of continually harping on about my original point, I think the effort involved in doing it well would be beyond the effort required for the average person looking to cheat.

And that’s the real crux of it. Not whether something can be done, because hypothetically speaking anything is possible in AI with sufficient time, money and effort. But that doesn’t mean it’s actually going to happen.

But since this entire argument is a hypothetical, it’s probably better we agree to disagree.

Your characterization of an adult child does not seem fair. What makes someone an adult? If it's academic discourse, then why is it valuable?

I mean, if you're into it, more power to you. Go nuts with finally figuring out what makes a human. Just don't claim it's more virtuous than anyone else's hobby, unless you can find a reason.

To call it "wasteful" says that something of "value" is being squandered, but the value is perceived by each of us differently.