I slightly disagree - it's the verification and integration of knowledge. The integration part - that is, reading papers and assessing them compared to all the others you've read, is quite hard, but it performs the useful function of creating a weighted average in your mind of what is real, and where the boundary is. This is part of graduate training, but it's basically what review papers do, so there is a time lag built in. AI for this would be huge, and it seems like it's not far off. And I want it to cite sources and point me to the real references, please.