GANs power DeepFakes
In most ML algorithms of the past, the overarching methodology was to use a discriminative approach. The way that those ML applications work is that they seek to basically prove something is not what it claims to be. In a simple use case, consider a spam email. For a discriminative approach to work, the algorithm seeks to show that an email is not a valid email because of the contents within the email. In other words, using a sample of what is a known good bit of content, obviously at a large scale, the algorithm uses that known good content to judge subsequent submissions.
Unless a certain threshold is met, the algorithm works using that available data to prove that what was newly submitted is not a "good" email; it is spam. This works well mainly for this application because in this use case, most spam emails are relatively formulaic and typically are easily detected.
There are clear giveaways that the email does not contain "good"...
