4 ms·

The problem is that the study's design makes it impossible to measure the magnitude of the issue. Any recommendation system operating on the scale of YouTube's

The problem is that the study's design makes it impossible to measure the magnitude of the issue. Any recommendation system operating on the scale of YouTube's is going to have a non-zero error rate, and this study effectively just confirmed this known fact. The relative uselessness of the design makes me think this is more about ammo for advocacy efforts than genuine interest in studying this issue. The question I'd be interested in is: how frequently does a typical YouTube viewer have actually harmful content recommended to them? Looks like it's pretty rare, if even the self-selected group that was interested in this effort and reported harmful content saw an average of (0.12 * 3300)/1600 = 0.2475 over the entire duration of the study. And that's a very generous upper bound, assuming that all of the people who didn't submit reports just forgot about the extension rather than failing to see any potentially harmful content.

uniqueuid 5y ago
Right, that's the issue here: We don't get any meaningful base rate. But the problem is: It's incredibly hard to get representative measurements of personalized recommendations. How do you want to collect those? You'd need a representative sample in the upper four digits (costs five to six figures), and track their browsing including on mobile phones. This is expensive and technically difficult, because you need to essentially MITM the connection (SNI doesn't give you the videos, only the domain). Of course, Youtube knows the precise extent of the issue, but they can't/won't/don't give precise data to anyone (problematic but understandable, given that every totalitarian government in the world would immediately send a wishlist).
- pseudo0 5y ago
  It would be more expensive, but I don't think it would be infeasible for an organization like Mozilla. Focus just on the desktop use to start with (extracting recommendations from the YouTube app would likely be challenging) and pay a representative set of users to install a browser extension that records the videos they are recommended. Then go through some subset of the videos and evaluate if they are harmful or not. It would definitely be more work, but at least that would give some useful new information on this issue, rather than confirming what was previously known.
  - uniqueuid 5y ago
    Well, I've done some of these things in other contexts. Market rate compensation for participation in such studies is ~$1 per day. You'd need at least 10'000 users to have menaningful confidence intervals. Say you're tracking for one month, then you pay 300k for the sample alone. If you have a single multi-purpose researcher doing study design, programming, logistics, analysis, outreach and all, that's perhaps 50k-100k a year (highly dependent on country and institution). Next you hire a bunch of people to do the content analysis. Either mechanical turk with a 2-3 factor overprovisioning or actually trained people who are 2-3 times as expensive. At roughly one minute per video, you'd need one hour for 20 videos, i.e. approximatelly 50 cents to 1 dollar (ballpark) per video. Say you argue that the long tail is irrelevant, and you want to code the top 100k videos (danger, arbitrary cutoffs are bad), then that's another 100k. That's half a million for something that isn't mozilla's core mission and is already being done by publicly funded researchers. Would be nice if they did it, but it sounds like a bad investment to me!
- gipp 5y ago
  YT started reporting Violative View rate a few months ago: https://blog.youtube/inside-youtube/building-greater-transparency-and-accountability/ https://blog.youtube/inside-youtube/building-greater-transpa...

uniqueuid 5y ago

Right, that's the issue here: We don't get any meaningful base rate. But the problem is: It's incredibly hard to get representative measurements of personalized recommendations. How do you want to collect those? You'd need a representative sample in the upper four digits (costs five to six figures), and track their browsing including on mobile phones. This is expensive and technically difficult, because you need to essentially MITM the connection (SNI doesn't give you the videos, only the domain). Of course, Youtube knows the precise extent of the issue, but they can't/won't/don't give precise data to anyone (problematic but understandable, given that every totalitarian government in the world would immediately send a wishlist).

pseudo0 5y ago
It would be more expensive, but I don't think it would be infeasible for an organization like Mozilla. Focus just on the desktop use to start with (extracting recommendations from the YouTube app would likely be challenging) and pay a representative set of users to install a browser extension that records the videos they are recommended. Then go through some subset of the videos and evaluate if they are harmful or not. It would definitely be more work, but at least that would give some useful new information on this issue, rather than confirming what was previously known.
- uniqueuid 5y ago
  Well, I've done some of these things in other contexts. Market rate compensation for participation in such studies is ~$1 per day. You'd need at least 10'000 users to have menaningful confidence intervals. Say you're tracking for one month, then you pay 300k for the sample alone. If you have a single multi-purpose researcher doing study design, programming, logistics, analysis, outreach and all, that's perhaps 50k-100k a year (highly dependent on country and institution). Next you hire a bunch of people to do the content analysis. Either mechanical turk with a 2-3 factor overprovisioning or actually trained people who are 2-3 times as expensive. At roughly one minute per video, you'd need one hour for 20 videos, i.e. approximatelly 50 cents to 1 dollar (ballpark) per video. Say you argue that the long tail is irrelevant, and you want to code the top 100k videos (danger, arbitrary cutoffs are bad), then that's another 100k. That's half a million for something that isn't mozilla's core mission and is already being done by publicly funded researchers. Would be nice if they did it, but it sounds like a bad investment to me!
gipp 5y ago
YT started reporting Violative View rate a few months ago: https://blog.youtube/inside-youtube/building-greater-transparency-and-accountability/ https://blog.youtube/inside-youtube/building-greater-transpa...