One thing that's struck me which I assume only reddit has, is the social graph data which holds all the relationships trends on who votes on what/who. There's limits on that based on how pseudoanonymous it is and the ease of making new accounts, but that seems valuable (at least to an outsider), although possibly in a "fighting yesterday's war" way as 'true' social networks like Facebook get value out of posts, which would be different to the way ML training may value it.

I think the problem is that the votes, not necessarily representative or anything “true” either.