Here's the thing: many of these public, human-trained experiments are being intentionally misled as an act of trolling by those interacting with it, the most obvious example being Tay. That's a very different result than ambiently picking up on some sort of latent racism. Not to mention that having the web as its only source will naturally lead to a more unfiltered and ambiguously accurate source.

true, but that doesn't apply to some of the examples given in the article - the police AI for example

That opens a different discussion, but I'll give one point of note on that: Bayes' Theorem.

statistics are only as good as the data you use

Wow, crazy:

(defn reduce
  ([f [sf & sr]]
   "var acc = run(f, runtime::first(sr), sf);
    var r = runtime::rest(sr);
    for (auto const& i : runtime::range(r))
     acc = run(f, acc, i);
    __result = acc;")
  ([f acc coll]
   "__result = acc;
    for (auto const& i : runtime::range(coll))
     __result = run(f, __result, i);"))