Presidential hopeful Donald Trump, in both his live speeches and his tweets, can often seem unpredictable. He speaks off the cuff, in short sentences and non-sequiturs, often making outrageous logical leaps in between. His incalculability is what makes him so entertaining.
Yet to the cold calculations of a machine, Trump appears to be much more predictable. Fed on a diet of Trump speech transcripts, an artificial intelligence program run by MIT researcher Brad Hayes pumps out quips and takedowns that, while they may lack some of the targeted venom of the real thing, sound incredibly authentic.
"Mark my words. We're going to beat ISIS," proclaims "DeepDrumpf" in one of the many statements Hayes has sent out via the bot's Twitter account. "Come replace the big lie, Obamacare. Believe me."
Right now, think of this: We owe China $1.3 trillion. We owe Japan more than that. We have gun laws. I'll bring back our money.— DeepDrumpf (@DeepDrumpf) March 5, 2016
Hayes is reluctant to allow the bot to post directly to Twitter, the robotics researcher told The Washington Post, for fear the reconstruction of Trump's ideas would produce something truly offensive or threatening.
The actual process the machine goes through to construct sentences is not unlike the predictive text function on a smartphone. Each sentence begins with a random letter, and then the system uses its knowledge of Trump's speech to decide each successive letter until the sentence ends. This goes on unit it's reached the 140 character limit.
The statements are made by the program in response to priming text or phrases like "Ted Cruz" or "#MichiganPrimary", which Hayes includes in the Tweets in square brackets. Sometimes the bot is primed to respond directly to other tweets.
[Trump steaks?] But nobody is going after more money -- we're going to put them away in order... tax South Korea.— DeepDrumpf (@DeepDrumpf) March 4, 2016
As with the other recent high-profile statistical modelling program which created original scripts for new episodes of Friends, DeepDrumpf occasionally returns results that are complete nonsense — not just ideologically, but grammatically as well. Hayes puts this down to the relatively small amount of reference content he's had time to feed the machine.
The neural network took everything it knows from the 42 pages of Trump speeches, and was not separately taught any rules about the English language.
OK, it's amazing right now with ISIS, I tell you what? I don't want them to vote, the worst very social people. I love me.— DeepDrumpf (@DeepDrumpf) March 3, 2016
I am the only one who can beat Hillary Clinton. I am not a Mitt Romney, who doesn't know how to win. Hillary wants no part of "Trump"— Donald J. Trump (@realDonaldTrump) March 3, 2016
Hayes built the system as a fun way to familiarise himself with neural networks — which he expects will have a profound effect on AI design — in his spare time. He chose Trump because he assumed, rightly, that the Republican frontrunner's manner of speech would be easier for a machine to accurately emulate than other candidates.
I’m what ISIS doesn’t need.— DeepDrumpf (@DeepDrumpf) March 3, 2016
The name of the account is a portmanteau referring to "deep learning" — the field of AI research to which neural networks belong — and to Drumpf, Trump's ancestral name which was recently ridiculed by comedian John Oliver.