Funnily enough, though, many ML engineers and data scientists I know (even those at Google, etc., who depend on human-annotated datasest) aren't familiar with these kinds of errors. At least in my experience, many people rarely inspect their datasets -- they run their black box ML pipelines and compute their confusion matrices, but rarely look at their false positive/negatives to understand more viscerally where and why their models might be failing.
Or when they do see labeling errors, many people chalk it up to "oh, it's just because emotions are subjective, overall I'm sure the labels are fine" without realizing the extent of the problem, or realizing that it's fixable and their data could actually be so much better.
One of my biggest frustrations actually is when great engineers do notice the errors and care, and try to fix them by improving guidelines -- but often the problem isn't the guidelines themselves (in this case, for example, it's not like people don't know what JOY and ANGER are! creating 30 pages of guidelines isn't going to help), but rather that the labeling infrastructure is broken or nonexistent from the beginning. Hence why Surge AI exists, and we're building what we're building :)