How much fact checking do you do on the research AI does?

I have the same problem in this project as I do at work: how do you verify the output?

Human review is the bottleneck at my office. I think software will be more and more AI written but also more and more AI validated.

For my nonfiction series, I have attempted adversarial agent fact checking, and I rerun it with newer models as they come out.

So: little human review of citations. I didn't expect this thread to shine on me favorably, but perhaps it's interesting to some.