A data analyst thinks he’s figured out how to tell which tweets Donald Trump wrote himself and which were done by his campaign staff.
David Robinson, who works for the programming site Stack Overflow, says he was curious about the differences between the more incendiary tweets, such as those calling his Democratic rival “Crooked Hillary,” and the more polished ones, such as a recent one thanking Green Bay, Wis., so he looked into the data.
“For me it’s more about finding a really interesting story, a case where people suspect something, but don’t have the data to back it up,” Robinson told TIME. “For me it was much more about putting some quantitative details to this story that has been going around than it was about proving something about Trump’s campaign.”
Robinson used data to analyze about 1,400 of Trump’s tweets going back to January, and he found a clear distinction between tweets sent on an Android device and those sent from an iPhone. The more inflammatory tweets with strong negative language which did not use photos or hashtags were sent on Android; the more traditional tweets with neutral language, photos and hashtags were sent on an iPhone. Trump is known for using an Android-based Samsung Galaxy smartphone.
“I expected to see a difference, but it was much more dramatic than I expected,” Robinson said.
Still, Robinson warned that the data isn’t the final word. His sentiment analysis of the words used in tweets, for example, doesn’t take into account how many different people have access to the Twitter account.
Robinson, who said he was surprised the analysis got a lot of attention, said it took him about 12 hours to put together, including the time spent writing the post. He said he hopes people learn something about programming from it. Robinson also told TIME he does not plan on voting for Trump.
“None of this is proof as much as statistical evidence and hypotheses,” Robinson said.
Pratheek Rebala contributed to this report