I've been researching AI detection (and also made a game to see if humans can quickly detect AI writing just based on style in a short passage: https://prose-or-con.com/ as a way to gather data and to sate my own curiosity about whether I can actually detect AI writing reliably...I can, but not as reliably as I thought).

Storyscope (https://github.com/jenna-russell/storyscope) is a really interesting concept, and one I'm in the process of reproducing and expanding on, even though it is less accurate on direct detection than something like BERT or Binoculars. It is interesting because the usual obfuscation tactics, even having a human transcribe and rephrase the story don't work; it looks at the story itself, rather than word use and grammatical quirks, etc. Plot, agents, temporal structure.

So far, AI writing is detectable to a pretty high degree, though I think it'll be a war of attrition that AI eventually wins. And, all the detectors are the same tools one could use to create undetectable AI writing; AI loves to iterate in a loop, seems like iterating to rewrite to be undetectable is a soluble problem (though models currently probably would end up writing worse and worse to avoid detection, and the current best models have begun integrating watermarking, pushing back the defeat of human writers for some time).

This was an interesting test, thank you for posting the link. I also found the AI writing in it harder to spot than I expected, but I suppose I'm mostly tuned in to the low-quality Claude-style slop that's most commonly posted to HN. The coding-oriented models seem to be optimised for feedback-driven code creation at the expense of the writing style (which stinks and is quite easy to spot).