He dives into why it’s such a hard problem to fix and explains how the root cause is more financial than ideological. It’s worth a listen.
Consider a situation in which you are selling a gadget, or software.
Do you want your ad run up next to a guy using the fgt term to describe other people? Even in 'good spirits' or even 'in context'?
Maybe that's too abstract, given that you perhaps might not be selling stuff.
I do.
Frankly, I think it's all overblown, I think people should speak their minds and we should all be tolerant.
That said I don't want my products anywhere near your controversy. No, way. We put a lot of effort in what we do, we're trying to communicate a message, we 100% do not want that anywhere near our message.
When we get up and go to work and try to do our jobs, a lot of this stuff seems academic and ridiculous. For the same reason we don't want our ads next to porn, we don't want it near ugly language. I love stand up comedians, but much of their (live) work is too off colour for many advertisers.
So, the decision by YouTub etc to pull ads, but allow the content to remain is actually quite reasonable.
They claim that this is backed by machine learning, but I have a hard time imagining how. I've seen many machine learning systems, but never one whose entire model whip-saws around with the schizophrenic rapidity of YouTube. Watch just one single video about something like a conspiracy theory or anti-vaccination nonsense out of cursiousity and their "machine learning system" will immediately conclude that you want to watch large quantities of those kinds of videos and your recommendations will be filled with them for quite awhile. It seems obvious to me that they're not training any kind of model to predict the viewers desires. Not even very naive Bayesian models would predict that 1 data point outweighs every single other action by a user and translates into a massive shift in interests. There seems to be a strong tendency for their systems to ignore individual preferences as much as possible in favor of funneling users into broad categories that the system knows better. I wouldn't be surprised if the way things work were that it took something like a conspiracy video, groups it with other conspiracy videos, and then just sends any viewer in that direction while ignoring the users own history, subscriptions, like/dislike status, and all other individualized signals. Given the scale, maybe YouTube simply can't afford actual personalized recommendation systems?
Back to advertising potential, imagine some 10sec no-context dashcam video of some racially ambiguous dude who allegedly did something getting run down by a cop car. Whether an ad for a Ford Explorer is appropriate in that context depends totally on the tribe viewing the video.
Such things have existed for quite awhile. The difference over time is that they are becoming more common and more frequent, while disseminating them got easier and easier. Still, we only rarely saw quite the same kinds of viral common conflagrations of outrage as we have today versus what we had in the cell-phone camera and viral emails era. The structure of the social media of today is vastly different and omnipresent.
https://www.youtube.com/watch?v=rE3j_RHkqJc
Technologists recognized the addictive power of games, applying those dynamics to social media. We tried our hardest to increase engagement, and that drew technologists into combining gamification and viral discovery to amplify outrage on social media.
Now we're in 2019. Looking back on it all, is it really any surprise? Now we're in 2019, and it seems like society is ready to let multinational corporations who have access to tremendous amounts of information about us, to monitor, censor, and censure us for "wrongthink." What could possibly go wrong?
Otherwise it'd just be hypocrisy and double standards.
We don't make peace with our friends, we make peace with our enemies.