It's a moot point if you are talking about such systems integrating into a platform like Steem. You could train it, you can have multiple algos and tweak their parameters. The problem you speak of comes from their systems being untrustworthy and opaque. It's a powerful tool, but it should be fully in the hands of the user. You could even have data sharing of decision algorithm parameters (neural net maps and such) so that you could even see how other people's systems are running, and even integrate them into your own. A peer to peer sharing system could be created for this. Not the data that is sorted, but the configuration the user sets to scrape their data and use it to help you find resources for you. Not just your own, but you could add and weight the data of those you follow, for example, maybe even be able to query automatically another person's configuration with a query and get the result they would get (or automatically compile multiple weighted by confirmation feedback).

It could be bigger and more important than Google, and competitive, and diverse. This is much the sort of thing discussed in Charlie Stross' book Accelerando. It can learn to anticipate your movements, adapt to your moods, and basically act as a filter that helps you more quickly locate information and spend less time looking for it.

But we must control it, and teach it, or it's a potential psywar weapon.