I spoke to several academics who had mentioned it, none had read the original - they all relied on someone else's referencing of it.
See https://shkspr.mobi/blog/2021/06/where-is-the-original-overv... for more details.
You find a lot of citations of Metropolis, Metropolis, Rosenbluth, and Teller’s paper on MCMC but I bet 1 in 10 have read it.
Ulf Grenander’s massive book on Pattern Theory is another, or Kolmogorov’s 1933 monograph on probability. Often cited but come on, did you really read it?
Another I’m aware of is references to the notion of VC dimension in learning theory. Many papers mention a hard paper by Shelah as an original reference to the concept…I saw the same reference given in a book by mathematician David Pollard with a remark along the lines of, “people say this is the same as VC dimension but I confess this paper is impenetrable to me so I really don’t know.” I appreciated the honesty.
I actually read that. Its definition of probability is pretty nice: it allows non-impossible events to have zero probability even in discrete case, e.g. a coin landing on its edge.
For example, when I started working in NLP, I did some things with the CYK algorithm (https://en.wikipedia.org/wiki/CYK_algorithm), which was defined in three papers by three separate authors. While I learned the algorithm from secondary sources, I of course cited the original papers (because one has to acknowledge the original author of the algorithms one uses). At that point, I wanted to read the original papers, even if just for curiosity, but they were impossible to find (I think one of them was available, but behind an outrageous paywall).
Now those papers are easier to find, but anyway I don't think I'd encourage a grad student to read them except for curiosity, the algorithm has been explained much better and in more accessible ways in more modern papers. Which doesn't mean that we shouldn't acknowledge the discoverers.