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.
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.
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?
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.
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.
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.
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.
10 comments
[ 0.26 ms ] story [ 10.2 ms ] threadThese kinds of things do enormous damage to people.
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.
Judging by the given remarks I see nothing here which proves AI use.
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.