back
18 comments
My butcher is still bitter because forty years ago he studied very hard and passed latin but the teacher accused him of cheating and failed him in the grounds that a loutish thuggish kid like him couldn't possibly have studied that hard or learned latin. He still tells the story and still gets very angry while telling it.

These kinds of things do enormous damage to people.

We were warned very early on, by people more prescient than me, of all the secondary harms to society like this one that were going to come with LLMs. We can't be surprised.
From article: > The marker's notes - seen by the BBC - raise concerns about the "polished/uniform tone" of the work, noting the "consistent flawless grammar and similar sentence structure with no variation in style". The notes also mention a "formulaic paragraph structure that follows a typical AI pattern of 'general claim, the example/result, and conclusion'." Other observations point to repetitive phrasing and a lack of "nuanced critique" as well as a high usage of "em" dashes, a type of punctuation thought to be favoured by AI models.

Judging by the given remarks I see nothing here which proves AI use.

Would you say there is any way to definitively prove a text is AI generated?

I believe it's impossible to prove that some text was generated by an LLM simply due to the limitations of the medium. At the end of the day, text is just a sequence of bytes, and a keyboard can create a byte-identical sequence to an LLM generated one.

Of course you can run various statistical tools, judge a text by gut feeling, or look for specific clues that LLMs usually generate more than humans (e.g. em dashes). But none of those will ever be a proof. Just indications, which together could add up to a very high likelihood.

IMO, the only way to prove an LLM generated some text is to see the original chat session.

AI “detection” tools should be banned. They have the potential to cause as much harm as AI plagiarism itself. Non technical people don’t understand the shortcomings of such tools, they treat the verdict as truth.
People also treat genAI outputs as truth. At least that mistake goes both ways.
If we're going to ban tools that are unreliable but people are starting to treat them like truth...
18 years ago I was accused of plagiarizing an unnamed online monograph (I was never told which one) on an English literature exam that I wrote with pen and paper inside a classroom. It was a very weird experience.
At some point, people will figure out that this doesn't matter.

The more I use AI, and especially the more I have my team deliver work and documents to me using AI, the more I realize that the skill of producing high-quality written materials is intact and still is a human characteristic.

If you can produce a high-quality dissertation, then you can produce one. That means that you can tell the difference between what the AI draft and great final output, because I'm real sure that AI can't produce something great already.

And when it can produce great by today's standards, then raise the standards. I assume the standards must have been raised around the time that access to research libraries and then later online research tools became available.

If I was a student in today’s world I’d probably be tempted to just make my own elaborate surveillance setup to just record myself as I do all the work just in case my school administrators were this boneheaded.
I've seen people on HN mistake genuine writing for LLM output as well. It seems to be most common with Aspie writing, presumably because of their learning style. I think I've seen it happen two or three times?
> It seems to be most common with Aspie writing

That's quite a claim, do you have any proof of this?

The 'reliable heuristics' for identifying LLM output tend to pattern-match on technical lexicons, adherence to syntax at the cost of parsability, and atypically rigid punctuation and literary contrivances. Typically these form the same rulesett historically used to demarcate the aneurotypical - clear, direct and precise to the point of coming across as mechanical.

Anybody familiar with ASD would wholly support the claim, and even the most cursory search validates it.

https://www.reddit.com/r/autism/comments/1q4vh0z/autistic_da... https://www.reddit.com/r/AutisticAdults/comments/1n8hrf6/aut... https://www.reddit.com/r/autism/comments/15sw3ei/have_you_be... https://www.reddit.com/r/autism/comments/1oh1kk9/update_just... https://www.reddit.com/r/autism/comments/1vgboem/autistic_wr... https://www.reddit.com/r/autism/comments/1f1n98g/does_anyone...

Given a large curated corpus of Reddit posts, it was shown that those posted in autism-centric subreddits—presumably written by autistic people—were more likely than posts from other subreddits to be flagged as AI-generated by OpenAI’s GPT-2 detector. https://arxiv.org/pdf/2607.14729 https://www.researchgate.net/publication/410300817_The_Miscl...

https://medium.com/illumination/when-my-writing-was-mistaken...

https://www.bloomberg.com/news/features/2024-10-18/do-ai-det...

I am specifically talking about Aspie and not other places on the autistic spectrum. (You would be one; I am not.) But otherwise, yes you are correct. However, note that it is not just the mechanical feeling of perfectly executed prose, but the mechanical feeling especially of imperfect prose that tends to come off as LLM-ish to some people. Specifically, imperfect Aspie prose is somewhat distinct and also may feel somewhat similar to LLM prose to those who don't have tells to distinguish the two, because the mechanical structure becomes more obvious, and one of the tells of LLM content is sloppy or awkward phrase usage or word choice, which is common from Aspies that are still learning. I tend to see that kind of prose like Scratch blocks of phrases they have learned. Once you know how to spot it, it is fairly obvious and distinct from pure LLM, even if LLMs have an otherwise mechanical output. The reason it's specifically Aspie writing that tends to get mistaken this way is because even other types of autistics don't typically learn to template whole phrases and sentences this way, so they tend to produce different kinds of flaws that are not as LLM-ish feeling.
Just show them your hand-written initial draft
That could have been produced after the fact, and even if i.e. a scanned version were provided with an eIDAS-compliant qualified electronic timestamp, it can still be argued that the student just wrote LLM output down.
It could, but it wouldn't, LLM cheats are lazy, and that sounds too much like work.
It's less work than doing it over again for sure. People lazy enough to be efficient in the first place will typically do some work to avoid even more work.