|
> for every tool mankind has created - there are ways to utilize it in good VS bad outcomes Counter point: Slot machines exist. (I suppose you might find some few instances of them being used to raise money for charity or something, but I don't think that redeems the technology.) Recommendation engines are generally benign when the input content is benign. Pandora itself doesn't seem to cause trouble. Though even with benign content, content recommendation can encourage unhealthy behavior, such the content recommendation engine of netflix/etc encouraging binge consumption and couch potatoism. But what does it mean for content to be benign and how is that determination made? When you have relatively little content that changes relatively infrequently, like netflix, you can have trusted humans perform the selection. That bakes the biases of those humans into the system, which is not inherently a problem (it's not as though any newspaper or book was ever any different in this regard.) But with things like youtube or the hypothetical Pandora for News, the the depth and breadth of content is too large for trusted human curation. I think we've actually already seen some examples of this manifest on Facebook, where the recommendation engine sends people down radicalizing holes of local news particularly. There has been a lot of talk recently about the identification and suppression of disinformation and misinformation. D&M is problematic and tasking independent fact checkers to fight it may have some merit, but I think there is much more to consider. For starters, the local fact-checkers, who understand the language and cultural context of the material they are evaluating, are imposing their own potentially harmful biases onto the system. Secondly, and more importantly, some propaganda is neither misinformation nor disinformation. An example: racists love to pass around links to factually correct local news articles about crime. They will inundate their targets with news stories that support their narrative. Each of those stories, evaluated independently, might be factual and come from a reputable news organization. But the sum of those stories may mislead people by not presenting mitigating considerations, social or historical context, etc. If you create a system that recommends people the intersection of 'local news' and 'crime', you've automated the job of the racist propagandists. But if your content recommendation engine doesn't recommend such intersections, by lumping all local news into a single category, then it has little utility over simply visiting the websites of local news organizations. And if your system forbids local news entirely because the breadth of local news across the world is too much to moderate, then your system omits the news that directly impacts the lives of news readers the most. I don't see any way around the above, but what makes a recommendation engine any worse than the local newspaper itself? Simply this: a newspaper is the same for everybody who reads it, when somebody is being sent down some bizarre radical rabbit-hole, those around them can read the newspaper and see what sort of thing their friend is reading. But recommendation engines provide personalized experiences. People can be radicalized without others in their community seeing what is happening, denying that community the opportunity to effectively respond and intervene. People withdraw into their own personal realities, losing touch with those around them. The further apart people grow, the harder reconciliation becomes. |