Overactive AI Filter?

[XerXesXu]

Depends on the reward function.

If it's actually being checked--that is, if there is training data where the human operator knows which text is AI-generated and which isn't, and can mark LLM guesses as good or bad--there would in fact be a "benefit" to the LLM for guessing right.

--Annie

The idea is to reward the LLM more, for saying 'I don't know', or 'there is no real-world answer to your question', than for guessing. At present they've all been trained using a system which rewards guessing.

LLM's will have to be retrained, at great expense, to achieve this. As a stop-gap, LLMs are being equipped with AI Experts, which limit the training data it will examine to answer factual questions. No more reddit or other social media posts. The latest Chinese release which currently tops the scores, KIMI K2, has 384 experts.

 

[Bramblethorn]

But in this case, if it actually gives them hints that allows them to fix his story to sound less AI, so that it and every other story they write can get posted without flagging?

"If."

Imagine there was a site whose "AI detection" system was simply tossing a coin. Heads, your story goes through; tails, your story bounced.

You submit your story; the coin comes up tails; your story bounces. You go to the trusty Magic 8-Ball in your cupboard, and ask it "should I make [whatever] change to get the story past the AI filters?" You shake the 8-Ball up, and it says "yes".

You make the change accordingly, and resubmit the story. This time the Coin of AI Detection comes up heads, and your story is accepted. Huzzah!

Does that mean the Magic 8-Ball actually knows how to get a story through the AI filter? Nope. Just means that you consulted the 8-Ball and then, unrelated, you tried again and this time you were lucky.

I don't suggest that this is an exact parallel to what's happening here, but it's close enough to be relevant.

 

[StillStunned]

Do you really think someone like StillStunned (an editor by profession) leaves grammatical or punctuation mistakes in his stories?

Yes . Or so @StillStunned tells me.

It's pretty much impossible to eliminate all typos, but my writing generally has fewer mistakes than AI-generated texts do.

 

[millamomud]

Here is another angle that might help contextualize things.
In the German BDSM section we used to see around three to six new submissions per day before the pending pandemonium began. Right now it is six stories in the last seven days.

There is also an extensive thread in the German authors' hangout about the AI-rejection issues, and frankly, I find it hard to believe that 80 percent of German authors suddenly switched to AI or that 80 percent simply stopped publishing because they discovered new hobbies.

From what I can see, the slowdown seems far more closely tied to the current approval process than to any actual "exodus".

That said, the site may very well continue to thrive without its German authors. In the larger scheme of things we are a fairly small community.

 

[firsttimecaller75]

From what I can see, the slowdown seems far more closely tied to the current approval process than to any actual "exodus".

This is sad to read! My rejected stories are available on MCStories and AO3, if anyone wants to decide for themselves. They involve CNC, mind control, betrayal, and medical abuse, though, so you may not want to do that--YKINMK, etc.

"Overactive AI detection software" as suggested by my editor still feels plausible to me, even if we don't really know the truth. Makes me especially concerned about the state of plagiarism- and AI-detection in universities. From what I can tell so far, ChatGPT won't even write such things....

 

[Bramblethorn]

This is sad to read! My rejected stories are available on MCStories and AO3, if anyone wants to decide for themselves. They involve CNC, mind control, betrayal, and medical abuse, though, so you may not want to do that--YKINMK, etc.

"Overactive AI detection software" as suggested by my editor still feels plausible to me, even if we don't really know the truth. Makes me especially concerned about the state of plagiarism- and AI-detection in universities. From what I can tell so far, ChatGPT won't even write such things....

It's not supposed to write erotica, but there are ways around that, and there are alternative products without that restriction.