It's not flawless, but for this task it works very well. Admittedly there's slight survivorship bias, as I do skim the results and discard anything that stands out as immediately bad.

That's actually pretty rare though, and as soon as the ChatGPT API is available, I'd be reasonably comfortable hooking it up live. I would want to experiment with some post-filtering to try and minimize the amount of weirdness.