Viliam's Shortform

Viliam

Viliam's Shortform — LessWrong

Comment Permalink

I checked three of my recent blog articles in Pangram and they were all evaluated as "100% human". (1, 2, 3)

I am reporting this because I definitely would have reported if the result was the opposite.

I was originally skeptical about detecting AI generated content, because the first implementations had lots of false positives, and fueled the hysteria about kids using AIs to do their homework. This implementation seems better.

There are a few more things I would like to check (when I get more free credits tomorrow), specifically:

AI generated content with a nontrivial prompt that tells the AI how to write (to check whether it detects "AI writing" or only "AI writing with default settings")
human generated, AI translated content
human generated, AI redacted content (e.g. "make this sound as if a native English speaker wrote it")

Maybe I will get some more ideas.

Showing 3 of 7 replies (Click to show all)

4RobertM5mo

I can assure you that LessWrong did not switch from Sapling to Pangram because of a "well-funded PR campaign", but because Pangram was in fact just better (in reducing both false positives and false negatives). You are welcome to look at the list of posts rejected for being LLM-written at lesswrong.com/moderation. [...] This seems like an extremely confused claim. In practice, the distribution of LLM-generated text is extremely different from the distribution of human-written text, which is why it's historically been pretty easy to tell when something is LLM-written, if you've seen a lot of writing by the relevant model. This is not a claim that there is some foolproof way to guarantee detection (short of the labs themselves cooperating and enforcing that their LLMs generate outputs with private statistical watermarks that they can verify after the fact); merely that this is not actually a hard problem to solve 99.9% of the time for things that have actually happened in real life.

14gwern5mo

So, then, it should be easy for you to give 3 examples of past handwritten text that triggered Pangram as 100% AI. Speaking just for myself, I haven't seen 'countless' examples of this on Twitter. (I've seen countless examples of the opposite, Pangram failing to detect AI text as AI, but not the other way.)

lilkim20254mo10

Are you claiming that it is impossible for a human to write in the voice of an LLM? That's a take I genuinely haven't heard yet. I'm willing to give it a quick go if this is actually what you believe, but I'd encourage you to try it yourself, or think about this claim a bit more, if that's really your position and I haven't misread.

See in context