They keep all the data posted since the early days of the internet, and are applying machine learning to it now (and will continue to into the future).
That's chilling.. The only long-going thread of attribution would rely on stylometry - a good fit for ML though..