Machine learning as a way to hate on the Internet?

In the era of ubiquitous hatred and more and more often organized social campaigns to fight it (e.g. regarding beauty hat, organized by the Rimmel, HejtStop or Stop Funding Hate brands), it is natural that a certain position is also expected from important and influential personalities or big companies. Therefore, the Google group decided to react and decided to launch a Chrome plug-in called Tune, although for now only in a test version.

While the name might suggest a music browser extension, the connotations are quite different. Tune comes from the English expression tune out, that is, ignore, ignore something. The purpose of the add-on is therefore to clear our feed on selected platforms from vulgar, pointless and hateful comments. We can try this feature on YouTube, Facebook, Twitter, Reddit and Disqus.

The add-on filters comments on several levels of toxic behavior:

profanity – concerns swearing and profanity,
insults – concerns insulting and insulting others,
threats – concerns threats, violence, harassment,
attack on identity – concerns negative or hateful comments directed at someone because of their origin, different appearance, etc.,
sexually explicit – it concerns lewd comments, referring in a vulgar way to sexual acts and intimate parts of the body. We can choose one of seven levels of filtering comments, ranging from “hide it all”, i.e. hiding all comments, through “medium” that removes the more controversial from the view, to “show it all”, which makes the entire section visible .

Tune was created on the basis of machine learning, and in addition to controlling comments, it is also to be an example of using this method in practice – in this case it is to work as a moderator of conversations between people. At the moment it only supports English, although when we checked how it would behave in the case of comments in Polish, she also tried to moderate them – unfortunately with little success. Below is a screenshot where you can see how the comments were sorted under one of the YouTube videos. Despite the fact that both are positive, one has been hidden (it is always indicated by a purple dot), which in this case is unjustified.

It all boils down to the fact that the plug-in is only a specially constructed mechanism, and although it is undoubtedly very complex and surprising with intuition, there is little chance that it will understand the irony or adjust someone’s words to the appropriate context. The assumptions behind the creation of Tune are noble, while the machine learning sector probably needs several more years to be able to afford a full-fledged combination of software with the human mind.