Manipulation of the debate in the US Elections

Antecedent

In 2012 I documented how several Trending Topics in support of Josefina Vasquez Mota and Enrique Peña Nieto did not come from spontaneous support but from twitter accounts tweeting automated messages. Those tweets left a trace.

That trace indicated they were originating from a program and were not being broadcast in the manner that ordinary people tweet. Instead of leaving traces that the tweet originated from a phone or a computer, these messages said they were being sent by two applications that are used to automate messages. They were bots.

Proceso reported what I documented in 2012. The application “SplitTweet” was detected in 3 trending topics.

https://www.proceso.com.mx/297541/acusan-a-equipo-de-vazquez-mota-de-fabricar-trending-topic-en-twitter

After the article was published, those accounts disappeared and they haven’t been active again.

US Elections

Since 2012 these tactics of manipulating conversations in Mexico have become very sophisticated, today it’s very difficult to detect them, they no longer leave digital traces.

However the bots I’ve found in the US elections are primitive compared to those in Mexico because they leave behind traces that are easy to detect.

How did I find them?

From October 11 to 27, I captured 50,000 tweets per day that mentioned the accounts of Donald Trump and Hillary Clinton. I noticed that during the October 19 debate, two hashtags, #SheWon supporting Hillary and #DrainTheSwamp attacking Hillary from sympathetic Trump accounts, appeared almost out of nowhere.

It’s not natural for normal people to be able to create trends in a matter of seconds so I assumed something was awry.

Then I analyzed the tweets I captured that night and reviewed the origin of the tweets. A good percentage of them came from applications that are used to automate or program tweets, the same as have been used in Mexico and also in Spain in 2012.

Besides finding automated messages that had as originated from applications, I found tweets that were broadcast too quickly (something that’s also abnormal for regular people to do, although you don’t need a software to execute – it’s typical of call centers or when many people meet up and agree and tweet at the same time).

Then in my dataset from October 18 to 20 including the debate tweets I found these messages:
In the first set I only registered automated tweets. In the second were others that were not automated but that had an abnormal rate of acceleration that a normal person could not accomplish.

In this area my findings coincided with those of the Oxford study. I found in all from October 11 to October 27 (almost 5%) and Oxford says they encountered 30% bots and automated messages.

To reinforce the analysis I did a network graph where I highlight how these abnormal and automated tweets stand out from those that are spontaneous and organic:

A sample of the automated messages pushing the trend #SheWon,
messages that for the most part have already been deleted

How can we tell that support for Margarita Zavala is inflated?

One methodology that occurred to me:

While US polls correlate with google searches about the candidates, Margarita’s do not.

US election polls are reflected in Google trends: https://www.google.com/trends/story/election2016

Zavala’s polls seem to surpass Andrés Manuel López Obrador

El Universal Polls:

https://www.eluniversal.com.mx/articulo/nacion/politica/2016/11/7/encuesta-zavala-encabeza-las-preferencias

Google trends about candidates in Mexico:
https://www.google.com/trends/explore?date=now%207-d&geo=MX&q=%2Fm%2F0gsyx_,%2Fm%2F035m7s,%2Fm%2F026drw7,%2Fm%2F0k3lwry,%2Fm%2F0wykx8x