Today comes a look at Ted Cruz’s Really Great Twitter Performance (TM) from a different angle, combining Pollchatter.org data with other projects’ research into bot and troll behavior.
Yesterday we found clearly suspect behavior from at least one of Cruz’s strong supporters – in that case, the improbable posting of two simultaneous posts, with precisely identical to-the-second time stamps.
We also looked at a 2-day histogram displaying the most frequent posting times for Cruz supporters. While there were some interesting surges, nothing was obviously amiss.
For this post, we’ll instead look at whether Cruz’s top-posting supporters exhibit bot or bot-like behavior. To do so, we’ll again look at a 48-hour time period (ending at 23:59:59 pm on Sept 14th, UTC time) – essentially the last two days – and examine the top 20 posters within the #KeepTexasRed hashtag, another Cruz special.
This turns out to be fairly interesting.
One reason this hashtag was so popular in this time period is because it was used by actor James Woods, whose single Tweet was massively retweeted.
But Woods doesn’t feature on the list of top posters, as ranked by volume of use. The members of this top 20 list all used the #KeepTexasRed hashtag (either in original Tweets or retweets) between eight and 17 times within this 48-hour period.
We plugged each of these 20 screen names into Bot Sentinel, a service that “uses machine learning and artificial intelligence to classify Twitter accounts that exhibit bot/troll behavior.” We also re-checked our original Botometer ratings on these accounts, and found that many of the scores had changed since we originally added them to our database, in some cases significantly.
(This shouldn’t be terribly surprising, since it makes sense that new behavior data would result in updated assessments. However, we are rethinking our database-stored bot-scoring function as a consequence.)
The results are listed in the chart below (text continues afterward).
| Rank | Screen Name | Bot Sentinel Rating | Bot Sentinel Category | Botometer CAP |
|---|---|---|---|---|
| 1 | smartgirls4gop | 73% | problematic | 8% |
| 2 | MikeDiaz285 | 72% | problematic | 8% |
| 3 | BoiseMarineMom | 77% | alarming | 0% |
| 4 | Croatansound | 52% | problematic | 2% |
| 5 | MalsLAW | 79% | alarming | 10% |
| 6 | AggieDad2015 | 48% | tolerable | 0% |
| 7 | AmericaFirst150 | 49% | tolerable | 10% |
| 8 | deeg25 | 68% | problematic | 23% |
| 9 | kathyboswell22 | 59% | problematic | 1% |
| 10 | PaulRog52896828 | 93% | alarming | 3% |
| 11 | BeverlyConserv1 | 76% | alarming | 10% |
| 12 | BlogUSA2018 | 88% | alarming | 32% |
| 13 | ChgocadChic | 83% | alarming | 11% |
| 14 | KTyrky1 | 50% | problematic | 7% |
| 15 | neff630 | 76% | alarming | 1% |
| 16 | Mariatweets3 | 97% | alarming | 4% |
| 17 | RachelSuperEli1 | 59% | problematic | 11% |
| 18 | jewlsw1 | 74% | problematic | 0% |
| 19 | PatrioticCindy | 72% | problematic | 8% |
| 20 | bonniemurphy | 81% | alarming | 54% |
| Average (mean) | 71.30% | 10.15% |
||
| Median | 73.50% | 8.00% |
This comparison leads to several interesting observations. First, the Botometer is far, far more conservative (in the sense of being cautious to make a judgment, not politically conservative) than is the Bot Sentinel project.
Using the Bot Sentinel, the average (mean) score among the top 20 #KeepTexasRed posters was 71.3%, which puts it at the high end of that service’s “Problematic” category (Defined as: “Our analysis has concluded XXXXX exhibits problematic tweet activity and patterns similar to a trollbot account”).
Fully nine of the top 20, or very nearly half, fell into the “Alarming” category (Defined as: “Our analysis has concluded XXXXX exhibits alarming tweet activity and patterns that match a trollbot account. Please report this account to Twitter and avoid”).
By contrast, the Botometer’s average “Complete Automation Percentage” for these 20 accounts was just 10.15%. According to the Indiana University team that runs the Botometer, the CAP is “the probability, according to our models, that this account is completely automated, i.e., a bot.”
A first conclusion here is that identifying bad actors online is a hard task. The two services are measuring slightly different things, with different machine-learning models, and so naturally come to different conclusions.
But a second conclusion, just as relevant for our purposes, is that many of the accounts contributing to Ted Cruz’s Really Amazing Twitter Campaign should be viewed with very strong suspicion.
Is this a deliberate attempt to manipulate public opinion? Again, we can’t say for certain. But the more data we get, the less good it looks.
Oh, and just for the record, our friend Guardian_Elite, the simultaneous poster we discussed yesterday, has a Bot Sentinel rating of 83% (“alarming”). Reason to be alarmed, indeed, if this is what qualifies as democratic discourse.
As always, for more information on all the 2018 competitive electoral districts, visit Pollchatter.com