Профиль

(and that is super important in terms of protecting not just children but everyone being abused by monsters like Musk and the users of his platform?)

Given you helped pass it I'm now super confused on the nature of the law. CSAM AI was already illegal in many states, but adult "nudification" (aka deepfake porn/abuse) was the thing that made this a first-in-the-nation for MN, no?

This isn't just for children, right? It's for adults too?

From Yrsa Daley-Ward's book, Bone: bookshop.org/a/862/9780143132615 #poem #booksky #writing

From Yrsa Daley-Ward's book, Bone: bookshop.org/a/862/9780143132615 #poem #booksky #writing

Yeah, he generally is pretty responsive, which I like; I think the responses to him are also valid criticisms of the parts he's ignoring and that the percent system as it functions is clearly broken.

Right, the scale is certainly a point in favor of it being useful unless it has a consistent form of error, but what about the questions about them in general and what they do differently? Also, do you have any larger write-up about your general detection maximalism I could look at? Thanks!

(Sorry, didn't mean for that to be a multipost rant, just... I feel like almost no one is talking about this the ways it needs to be talked about. So thank you!)

Meanwhile, originality.ai for example will flag a random chapter as 90% likely AI... then if you remove 4 em dashes from 1,000+ words, change that to 50-60% likely AI (still garbage). They have literally made an em-dash counter that they're selling as an "industry leading/most accurate AI detector"

But dear god some of the others. I and a coauthor have a novel out on sub right now. Pangram, thankfully, flags any tested chapter as 100% human (accurate, never touched by any genAI software period).

That, for the reasons you're saying, is potentially even worse, because it should not be trusted, per for example: freddiedeboer.substack.com/p/i-wouldnt-...

Been engaging in this all day after the Atlantic piece and Tuhin Chakrabarty's study, relying on Pangram (even though it's for a good purpose, the aid of copyright lawsuits against AI). I hate seeing any checker seemingly become super trusted; Pangram does somehow seem leagues above the others.

And this is a vital read about their failures: freddiedeboer.substack.com/p/i-wouldnt-...

@tuhinchakr.bsky.social Note particularly concerns like this, for example: freddiedeboer.substack.com/p/i-wouldnt-...

This, though, is a good criticism: freddiedeboer.substack.com/p/i-wouldnt-...

Vs. for example originality(dot)ai which will flag 100% human text as 90% likely AI, and then if you randomly like the three em dashes in that thousand-word text to commas, will lower it to 50% likely. Aka a glorified em-dash detector.

Yeah, I have significant skepticism over any text detector; at the same time, it does pass every test I give it that every other one fails (properly flags AI that all the previous supposedly best detectors miss; most importantly haven't yet seen it do anything but 100% human for human text).

Same. I really want to know what they're supposedly doing so differently that they both have a completely different kind of result (amount AI generated vs. likelihood AI generated that all others have). I will say their results ARE impressive vs. lots of bullshit ones, but still.

If you want people to take this seriously, an understanding of what Pangram is doing/how they are somehow completely different than all the other bullshit out there (particularly originality.ai for example) feels really important!

As I mentioned in another thread, I've been impressed by their results and the false negatives and more importantly false positives from other detectors they avoid, but people are understandably skeptical and I think dealing with those nuances is important.