Some insights from the report:
* The study uses donated data from mozilla's browser plugin [1]. This means it almost surely has self-selection bias and does not offer representative usage data. But that's a fair tradeoff to get realistic measurements of recommendations.
* The focus is on harmful content, so we don't know how much this is out of the entire exposure people get. (But there are a couple of representative studies out there).
* Out of all reported videos, 12% were considered actually harmful. That's not a terrible rate in my opinion.
* 70% of harmful videos came from recommendations - that's why the authors demand changes to the YT recommender system.
* Data quality: Reports came from ~37k users of the browser extension, but only ~1600 submitted a report. Total N of reports is ~3300, so pretty low. The actual analysis was performed on a sample of ~1100 out of these (why?). Harmfulness was analyzed by student assistants which is ok.
* Non-english languages seem to have higher prevalence of harmful videos (which is to be expected given the attention and resources spent).
[1] https://foundation.mozilla.org/en/campaigns/regrets-reporter...