Thoughts on AI checkers

[SylvanusSilvertung]

I've recently been having the experience of having most of my work flagged as AI generated - including stuff that I submitted before there was even AI to generate.

It's annoying because I don't think there's even a human looking at it. It's just an AI deciding my writing is too predictable.

I took a piece that I wrote and ran it through a "make this sound like a human" AI - which totally wrecked the tone, but I'm waiting with bated breath to see if it goes through.

 

[EmilyMiller]

I took a piece that I wrote and ran it through a "make this sound like a human" AI - which totally wrecked the tone, but I'm waiting with bated breath to see if it goes through.

You just publicly admitted to deliberately using AI in a story submission. That was probably somewhat unwise.

 

[SylvanusSilvertung]

You just publicly admitted to deliberately using AI in a story submission. That was probably somewhat unwise.

I also mentioned it in the notes to the editors. I'm getting so fed up with having my writing rejected when - as any human who read my work would note - the story building is waaay past anything AI can manage right now.

If the only way my story can make it past AI detection is using AI - then the site is broken and I'm moving.

 

[EmilyMiller]

I also mentioned it in the notes to the editors. I'm getting so fed up with having my writing rejected when - as any human who read my work would note - the story building is waaay past anything AI can manage right now.

If the only way my story can make it past AI detection is using AI - then the site is broken and I'm moving.

If the site prohibits AI-assisted stores - as it does - I’m not sure what you were hoping to achieve in telling the site you used AI. The ‘make my story more human’ tools are crap.

 

[Mastered_again]

Quick factoid (from an article I read from a very reliable source, but trying to steer clear of copyright issues)
ZeroGPT, an AI detector, flagged The Gettysburg Address as 96.4% AI generated. Lincoln received credit only for "We are met on a great battlefield of that war."

Don't you just love the precision?

 

[anthrodisiac]

Quick factoid (from an article I read from a very reliable source, but trying to steer clear of copyright issues)
ZeroGPT, an AI detector, flagged The Gettysburg Address as 96.4% AI generated. Lincoln received credit only for "We are met on a great battlefield of that war."

Don't you just love the precision?

It makes sense, given the recent discovery that Lincoln had beta access to ChatGPT 3.1.

 

[SmilingLez]

Quick factoid (from an article I read from a very reliable source, but trying to steer clear of copyright issues)
ZeroGPT, an AI detector, flagged The Gettysburg Address as 96.4% AI generated. Lincoln received credit only for "We are met on a great battlefield of that war."

Don't you just love the precision?

In another thread, a while ago, I included a screenshot of a section of the Gettysburg address that I ran through ZeroGPT and it came back as 100% AI generated.

 

[Mastered_again]

It makes sense, given the recent discovery that Lincoln had beta access to ChatGPT 3.1.

Don't forget FDR's Fireside ChatGPT

 

[Mastered_again]

In another thread, a while ago, I included a screenshot of a section of the Gettysburg address that I ran through ZeroGPT and it came back as 100% AI generated.

All from an algorithm (for lack of a better term) whose criteria are proprietary - owners of ZetoGPT declined to be interviewed for the article.

Edit to add: Makes me think that they don't fully understand what's going on in their "validator."

 

[SmilingLez]

All from an algorithm (for lack of a better term) whose criteria are proprietary - owners of ZetoGPT declined to be interviewed for the article.

Same could be said for literotica, the owners have declined to disclose what tools they use to reject a work for any reason

 

[anthrodisiac]

In another thread, a while ago, I included a screenshot of a section of the Gettysburg address that I ran through ZeroGPT and it came back as 100% AI generated.

I can understand how that could happen, since the Gettysburg address is almost certainly in all LLM training data, and ZeroGPT probably has it in its training data too. So you get a verbatum quote like that, which appears in full in the training set, it'll probably flag it. I imagine that's likely true of other works, stories, and things that the models trained on.

 

[Mastered_again]

I can understand how that could happen, since the Gettysburg address is almost certainly in all LLM training data, and ZeroGPT probably has it in its training data too. So you get a verbatum quote like that, which appears in full in the training set, it'll probably flag it. I imagine that's likely true of other works, stories, and things that the models trained on.

Wouldn't that be plagiarism, not AI generated?

 

[anthrodisiac]

Wouldn't that be plagiarism, not AI generated?

It is, but that doesn't mean it wouldn't be flagged as AI-generated. AI-checkers are (probably) pattern-recognition algorithms. And if it finds a long pattern that matches something from its training set, it's likely to flag it. When that flag is raised, it's likely just passing back how much of the text matched its training data/patterns, not what type of pattern it matched. And since the tool is for explicitly for checking AI generated text, not plagiarism, what's likely happening is that the percentage is the only variable returned by the check. So when you see the message: "X% of this text is AI generated." "of this text is AI generated" is probably hardcoded, so the only variable that changes in the message is the percent value.

Granted, that's all speculation. But it seems like the most likely reason why historical texts, and any other texts that were tossed in as part of the training data, would be flagged as AI-generated: because the tool isn't equipped to handle the distinction between AI-generation and plagiarism. I bet if you ran it through a dual AI-generation/plagiarism checker, it would pop out the message that it's plagiarized instead of AI-generated. But ZeroGPT isn't a plagiarism tool, so it just pops back the message for the "question" it was asked: how much of this is AI?

 

[SkyBubble]

Quick factoid (from an article I read from a very reliable source, but trying to steer clear of copyright issues)
ZeroGPT, an AI detector, flagged The Gettysburg Address as 96.4% AI generated. Lincoln received credit only for "We are met on a great battlefield of that war."

Don't you just love the precision?

Apparently, Shakespeare used it too.