Donald Trump is in the news again, but this time not so much for what he has said, but for what he hasn’t. Recently, Data Scientist David Robinson discovered a major difference between “The Donald’s” iPhone and Android tweets, providing evidence that the iPhone tweets, in fact, were likely not written by the Trumpster. So, who wrote these distinctly non-Trumpian tweets? Our analysis suggests that Trump staffer Gavin Smith most likely composed the overwhelming majority.
Trump is fairly prolific in the Twittersphere, with ~33,000 tweets since March 2009 and an average of 13 tweets per day over the past 8 months. Hidden in the metadata of each tweet is information on which device the tweet came from, e.g. from an iPhone, iPad, Android phone, the Twitter website etc.
Over 90% of the tweets from Trump’s twitter account are from either an Android or iPhone device, with approximately the same number of tweets coming from each. After Todd Vaziri noticed that Trump’s Android tweets contained far more hyperbole than his iPhone tweets, Robinson dug deeper to investigate whether there was evidence that different people authored the two samples.
He found that the pattern of word usage and overall sentiment of the Android tweets were very different to the iPhone tweets, with the former being far more negative. He also noted that the timestamps of the two samples were quite disparate, concluding that this was strong evidence that the authors of the iPhone and Android tweets were different people. As Trump is reported to tweet from a Samsung Galaxy phone, Robinson concluded that the author of the Android tweets is Trump himself, with the iPhone tweets being presumably produced by one or more of his staffers.
Discriminating between Trump and his staffers with machine learning
The evidence that the two samples are written by different people seems pretty convincing, but linking the Android tweets to Trump rests on a pretty big assumption. How can we be sure that one of Trump’s staffers doesn’t write the iPhone tweets while another tweets from an Android? If the authors of the two samples of tweets are indeed different people, then the Receptiviti personality profiles should be different. Also, if the Android tweets are written by Trump they should match his personality profile deduced from other language we know definitely came from him. Having played with using personality profiling as a forensic signature in order to attribute authorship to tweets in the past, this struck me as a very interesting and timely question to address.
To investigate this, I used a sample of ~3,200 tweets scraped from the @realDonaldTrump account by the data harvesting company BrightPlanet between December 2015 and August 2016. Combining tweets by source (e.g. Android vs. iPhone) into samples of 100 words each, I ran these through the Receptiviti API to produce personality profiles. I then used the random forest machine learning algorithm to build a model to classify personality snapshots as either belonging to the Android or iPhone classes. This model was able to discriminate between the Android and iPhone personality snapshots ~94% of the time, strongly indicating that the authors of the two tweet samples are very different people.
Ok, this supports Robinson’s conclusion of different authors, but what about the assumption that the Android tweets belong to Trump?
To test this I took transcripts of two of Trump’s recent press conferences (the one where he bizarrely continued his attack on Ted Cruz and the other where he called on Russia to find Hillary’s missing emails) and ran these against my model. The model classified both press conferences as being significantly similar to the Android tweet profiles.
It seems that Robinson was very likely correct and that Trump really is the author of the Android tweets. But what does Trump’s personality profile say about him as a person?
Trump uses small words, is verbose and not much of an analytical thinker
Interestingly, the most important features for classifying the samples were the length of each sentence (Trump is more verbose), the level of punctuation used (Trump uses less) and the number of words > 6 letters in length (Trump is less inclined to use long words than his staffers). According to Ryan Boyd, social psychologist and co-creator of LIWC2015, Trump’s average personality profile allows us to make comparative estimates of his traits. Boyd suggests that Trump’s linguistic profile is similar to that of someone who is not particularly strong in his social connections; shows irritability and intolerance of setbacks; and is highly attuned/attentive to social status (both his own and that of others). It also indicates he is exceptionally low in analytic thinking (i.e., deliberate, effortful decision-making), but high on dynamic thinking style (intuitive, gut-based decision-making). Check out the blog that Kayla Jordan and Receptiviti co-founder Jamie Pennebaker have been writing about the Presidential candidates for more on what Trump says implies about his personality.
It seems as though the case is closed regarding Trump and the Android tweets, but can we identify the author(s) of the iPhone and web tweets?
So who are his more articulate staffers?
To answer this question I ran tweet samples from the personal Twitter accounts of 9 of Trump’s staffers (again sourced by BrightPlanet) through the same process outlined above, this time building a model to classify personality snapshots as being associated with the staffers. Some staffers (e.g. Gavin Smith and Daniel Scavino Jr.) tweet a lot, while others (e.g. Michael Glassner and Michael Cohen) are relatively Twitter-shy in comparison, making them harder to profile. All of them tweet from an iPhone on occasion (some more often than others), while Corey Lewandowski is the only one who doesn’t seem to tweet from the Twitter web page (he uses an iPhone exclusively). Only Joe Gruters and Michael Glassner had tweets from an Android device in the sample.
The staffer model was able to discriminate between individual snapshots > 89% of the time, with an average accuracy of ~96%. So, like the Android and iPhone tweet samples the individual staffers also have quite distinct personality profiles.
And the culprit is…
Running the profile snapshots from Trump’s iPhone and web tweet samples against this model found that 53% of the iPhone and 75% of the web tweets were attributed to Gavin Smith. The next most likely authors of the iPhone and web tweets were Dan Scavino (18% of iPhone, 6% of web) and Jeff Taillon (11% of iPhone, 13% of web). Interestingly, Dan Scavino’s Twitter page lists him as the Director of Social Media for the campaign, so it’s somewhat surprising that he’s not the front-runner for the primary author of Trump’s non-Android tweets.
South Carolina Field Director
Director of Social Media & Senior Advisor
South Carolina State Political Director
The 3 Trump staffers most likely to author the iPhone and web tweets (profile pictures from Twitter).
Its entirely possible that there are a number of different authors of Trump’s iPhone and web tweets (or that the author is not one of the 9 staffers that I analyzed), but it appears as though the load is most likely carried by Gavin Smith possibly with some lesser contribution from others. Of course it’s hard to estimate the impact of someone trying to match the style of the real Trump on the profiles, but I think its safe to say that Trump really is the author of the inflammatory Android tweets.
Perhaps one of the more interesting things to come out of this analysis is the lower usage of long words by Trump. I take this as proof that Donald Trump doesn’t, after all, have “the best words” (at least when compared with his staffers).