Post

Bloody hell. That's frightening.

Essentially yes, because both GPT and these detectors rely on LLMs, so it's the same tech.

Of course. Definitely maths from the Large LANGUAGE Model. Trustworthy maths. With real numbers. Not approximations in the output. Never that.

Maybe some of the scammy ones work like that. The correct way to do this is to train your own classification model, which is what pangram does. It uses many of the same techniques as training an LLM, but it's not the same as asking ChatGPT at all.

In the wake of Windows simplifying en dashes and em dashes, I guess everyone has become AI.

The perils of putting a BS machine to work as a BS detector.

Could you try it in Pangram? Interested in the results.

No I don't believe Pangram has any association with them! www.pangram.com

Textguard is very bad, as I found out here: www.nature.com/articles/d41...

Per other comments here, would be curious about if you encounter any of that with them. They are certainly not perfect but I have not seen any evidence of false positives on creative work especially; their specific emphasis on avoiding false positives even at expense of false negatives seems real.

Surely I'm hopelessly naive, but if you're already a price winning author, why are they checking this at all? I fear to learn how common AI use might be among established writers. Please tell me it isn't so.

I'm sorry, what the ACTUAL fuck???

Right now we have the honors system and it's failing. I don't know if there's a way to tech your way out of this. The only ideas I can think of rely on good faith. Leave a penny take a penny. Anyone operating in bad faith can circumvent.

I find if I use either a semi-colon or an em-dash, I get flagged as AI.

I used one of these sites to try out their reliability for art by feeding a piece I did into it. Hand drawn on paper and scanned, it said 80% chance it was AI and definitely digitally enhanced. Only 1 site said it was human made. They're absolutely broken.

Nah, because rejection letters likely *are* being written (or at least edited) by AI. People are massively abandoning email-writing.

These detectors have been debunked for a while now. I think that false positives are the point, actually. To further muddy the waters.

There is something horrifically ironic about being given ai generated suggestions to avoid ai checkers

(Your response, that is)

As someone who uses like a lot (I think), this makes me cringe. Yeah AI slop machine, feed me more easily digestible text, in case my readers are stupid. I dunno, that suggestion feels insulting.

Do you have specific examples of agents or publishers using it? I'd love to share with anyone in the industry my specific experience with informing Originality about serious and ridiculously broken false positives and them doubling down on them.

Are ai detectors just a way for ai companies to get fed new material legally, without needing an author's consent?

It puts in to question other serious questions. Academic credentials nowadays should have an Asterix with people that earn their degrees before AI or after especially when there was a thesis involved. 

Building these on the gateway side, this tracks. Detectors read surface stats, never intent. So they flag a human who writes in fragments and pass a model parroting a clean template. The real lever isn't 'human vs AI', it's verifiability: sources + a trace you can audit.

AI RANT (cont’d) AI slop machines excel at pattern matching, but what they can’t do for you is analyze the author’s _intent_. You hit them with too many sentence fragments or em dashes, and they’ll claim you’re AI; they have zero idea whether you really _meant_ it or not. 🧵5/5

Plenty of people who are otherwise aware of just how fallible generative AI is just completely forget that when they're asking generative AI if something is generative AI. -.-

the AI 'suggestion' doesn't even make any damn sense and loses all emotional context. like damn.

And leave one with uninteresting sponsors reading (not the OP’s writing, of course)

"A ship's bell AND diver's lantern is terrible"--maybe. It's a WIP. I didn't want to use any of my published works so as not to feed the detector something it had quite probably been trained on.

Being accused of using AI when you didn't is like a 21st century witch accusation. You can't disprove it. Damn bro, you could have just said it was bad.

What a nightmare.

I guess authors are going to have to start recording their screens as they write, as many students have started doing when writing their papers.

Exactly. AI *can* help, but the criteria of detection has to be precise, not just "there's X% probability this is A". That will only bite everyone back. Looking forward to watermarking.

Yes, it will. Its primary purpose will be to sift out the low-effort scammers, which are the majority flooding the market. Any watermarking can be overcome with sufficient effort, to which I make a counterpoint: why not put that effort into the actual art?

Alternative?

So analyzing things based on minuscule details, which is all a GPT could possibly do, will never be a useful AI detector, because the AI stole all those details from human writing. The only effective "AI detector" is a human reading a piece holistically, looking for intent.

pangram does actually seem to be doing something different and not just asking an llm if text is ai or not

I mean, using AI to detect AI already seems absurd on its face, since you can always then train it to defeat the detection. To say nothing of how fast the models (and probably more importantly tooling) is still evolving.

I'm surprised there isn't more trust in you as a known author. That sucks.

I think pangram is a strong argument against your claim. Their FPR is very low, and all the examples I've seen so far are very understandable. False negatives are more common, but IMO a much less severe problem. At this point I would only be swayed by several examples of odd pangram false positives.

Ah! I hadn't read any of their documentation for pangram 4. Thanks for the update

I only trust Pangram as not a scam. They're good eggs, but thing is, by the time you're reaching for it suspecting writing is ai-bad... You already decided it isn't for you. The algorithm notes bounce back when you immediately click off. Why give bad text more of your care anyway?

I 100% support your AI rant. Can't wait for the bubble to collapse.

Someone wrote alt-text for it here :) bsky.app/profile/book...

Please please please, do not use "AI detectors," they are all scams. Of all the content type "detecting" ai writing is considered the most difficult because of how subjective it is. They look for style like it's habitually uses em-dash because of it's training. This is NOT proof of AI

Damn, I almost wish I’d known I could make bank selling software that just did a search for the existence of an em dash.

I try to do it with my 2009 PhD thesis but i failed to do it for free, what's the best way ? (Nb : im pro AI as soon as i do it for free (yes im greedy as f))