Twitter optimises for engagement. That's not optimising for controversy, but it's close. Shiri's scissor statements[0] aren't the fault of the people writing the training data; they're the fault of the people making them.
Twitter doesn't synthesise things, but it fosters an environment where people are driven to write more controversial, outrageous, engaging things by tight-loop-feedback classical conditioning. It's not magic – the people posting such things are partly to blame – but Twitter wouldn't have half as many problems if it just showed random tweets to people. Instead, it shows people what it thinks will keep people on Twitter; short term, that works, but long-term it destroys Twitter's value (and value of everything Twitter touches, as a side effect).
Twitter doesn't synthesise things, but it fosters an environment where people are driven to write more controversial, outrageous, engaging things by tight-loop-feedback classical conditioning. It's not magic – the people posting such things are partly to blame – but Twitter wouldn't have half as many problems if it just showed random tweets to people. Instead, it shows people what it thinks will keep people on Twitter; short term, that works, but long-term it destroys Twitter's value (and value of everything Twitter touches, as a side effect).
[0]: https://slatestarcodex.com/2018/10/30/sort-by-controversial/