Thread by @jugander

Terrific reporting by @NYTimes team. I stumbled upon these "createtime fingerprints" independently in August 2017, and have been thinking a lot about them since, so thought I'd share some thoughts. [thread] 1/n

https://twitter.com/Rich_Harris/status/957310634378629121

I stumbled upon the patterns after reading @gilgul’s 2014 piece about buying bots back in 2014: 2/n
medium.com/i-data/fake-fr…

Some quick queries of twitter’s API gives the following view of the creation time of @gilgul’s bot followers from his 2014 experiment. (see NYT piece for how to read this plot) 3/n

In the cases the NYT focus on, it’s pretty clear the followees were the purchasers, but one of the things I’ve tried to careful about is understanding how these patterns might not always be the result of followee behavior. 4/n

First, it’s perfectly possible that a bad actor would maliciously target someone with bot followers to make them look bad. So just because someone has many bot followers doesn’t mean they bought them. 5/n

Second, following celebrities is very natural “human” behavior on twitter, possibly through the recommendations made to new user. So: it’s very likely that celebrities pick up bots that follow them to make themselves look more human in the face of bot detectors. 6/n

This second matter points to a different use case of the "create time fingerprint”: it can also reveal when an account is added/removed from twitter’s new user recs. 7/n

See @kashhill here, where the sharp changes in concentration of “new” followers captures being in/out of the new user recommendations. Note that this pattern is completely outside her control! 8/n

These are just a few things revealed by these simple plots. I have also seen some other things that I don’t feel I understand well enough to talk about here. I’m super impressed with the NYT team’s work! 9/n

The real take away is that twitter really needs to up their bot detection game, as @gilgul pointed out way back in 2014. [fin] 10/10

Postscript for those interested, see also this 2011 twitterological paper by Meeder, Karrer, and colleagues on using the create times of users and the sequential order of follows to estimate/bound follow event times: andrew.cmu.edu/user/ssayedir/… 11/10

Also also, @syardi, @DanielMRomero, Schoenebeck, and @zephoria were writing about twitter spam account detection back in 2009. The method they describe, backwards random walks, likely still finds spam accounts today: firstmonday.org/article/view/2… 12/10