Postdoc develops Twitterbot that uses AI to sound like Donald Trump

Based on a deep-learning algorithm, @DeepDrumpf trains itself on transcripts from Trump's speeches
Based on a deep-learning algorithm, @DeepDrumpf trains itself on transcripts from Trump's speeches

Bookmark and Share

Think Donald Trump says some crazy things? Just wait until you hear his Twitterbot.

This week a postdoc at MIT’s Computer Science and Artificial Intelligence Lab (CSAIL) developed a Trump Twitterbot that Tweets out remarkably Trump-like statements, such as “I’m what ISIS doesn’t need.” The bot is based on an artificial-intelligence algorithm that is trained on just a few hours of transcripts of Trump’s victory speeches and debate performances.



Dubbed @DeepDrumpf, after John Oliver’s recent segment about Trump’s ancestral name, the bot creates Tweets one letter at a time. For example, if the bot randomly begins its Tweet with the letter “M,” it is somewhat likely to be followed by an “A,” and then a “K,” and so on until the bot types out Trump’s campaign slogan, “Make America Great Again.” It then starts over for the next sentence and repeats the process until it reaches the 140-character limit.

The Tweetbot’s creator, CSAIL postdoc Bradley Hayes, used techniques from “deep-learning,” a field of artificial intelligence that uses systems called “neural networks” to teach computers to to find patterns on their own. Hayes was inspired by an existing training model that can simulate Shakespeare, as well as a recent report that analyzed the presidential candidates’ linguistic patterns to find that Trump speaks at a fourth-grade level.

“Trump’s language tends to be more simplistic, so I figured that, as a modeling problem, he would be the most manageable candidate to study,” says Hayes.



@DeepDrumpf’s Tweets don’t always make complete sense, but are usually at least partially coherent - much like the candidate himself.

“The algorithm essentially learns an underlying structure from all the data it gets, and then comes up with different combinations of the data that reflect the structure that it was taught,” says Hayes. 



The bot has even directly interfaced with Trump’s Twitter account. In those cases, Hayes gives the algorithm language from the real Trump’s Tweet, which primes it to give a response that is more likely to be contextually relevant. 



In this way, Hayes envisions developing accounts for other presidential candidates (“DeepBern,” anyone?) and feeding Tweets to each other so that they can all converse in a real-time deep-learning debate.

“Much of my actual robotics research deals with these types of modeling techniques,” says Hayes. “I thought this would be a good way to learn more about some of the concepts, and have a little bit of fun in the process.”