Hacker News Reader

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).

View on HN

When instructed to emulate a style (like "travel writing" or "an encyclopedia style article") the best models are very difficult to detect in short passages. Longer work does tend to out even the best ones, though. But, I couldn't make a game of "read these ten pages ten times" and expect anyone to play it. Even now, it's not exactly a rip roaring good time, more of an interesting experiment that curious folks do once or twice just to find out how well-calibrated their built-in AI detector is.

I'd like to come up with a more fun game loop, but nothing has revealed itself to me yet. Shorter passages are easier to gamify, but shorter passages are much harder to detect.

Reply on news.ycombinator.com