You know that it's not possible to do worse than a coin flip, right? If you're getting it 100% wrong, I'll just do the opposite of what you say, and have a 100% correct predictor.
The threshold isn't 50% because the distribution of human and AI written cases isn't naturally 50-50. So a coin flip will underperform always guessing the more frequent class. Where it gets interesting is if the base is unknown or variable over time or between application domains. Like, since AI written text is being generated faster than the human kind, soon guessing AI every time will be 99% accurate. That doesn't mean such a detector is useful.
When we say "coin flip" in these situations we mean "chance", ie the prior distribution. Otherwise a predictor of the winning lottery numbers that's "no better than a coin flip" would mean it wins the jackpot half the time.
Yup! My point is that the 'coin flip baseline' model that's as good as chance isn't actually trivial to create, for an unbalanced and time varying underlying distribution.
