Donald Trump has earned a reputation as an unpredictable guy. He doesn’t use a teleprompter. He speaks off the cuff and we never quite know what he’s going to say.

Of course, that’s what it feels like to us humans. But at least one machine isn’t as surprised by Trump. A researcher at MIT’s Computer Science and Artificial Intelligence Lab has built a computer system that’s smart enough to produce original lines that feel like they’re straight out of a Trump speech. He’s posting them on Twitter, at @DeepDrumpf.

Brad Hayes wanted a fun way to familiarize himself with some statistical modeling techniques for his research on human and robot interactions. Two hours later, he had a computer program spurting out lines such as “Mark my words. We’re going to beat ISIS,” and “We really do have people that are stupid.”

Hayes fed the computer system 44 single-spaced pages of Trump victory speeches, debate transcripts and other remarks. The Trump quips serve as training data, from which the computer system realizes his speech patterns and style. Then it’s able to make predictions of lines Trump would likely say.

Hayes chose Trump because he felt the computer would have an easier time with Trump’s short, imperative and straightforward sentences rather than the statements of other candidates.

Occasionally the system produces nonsensical text, which Hayes said is a product of his system being trained on so few remarks. He said he’s pressed for time given his day job of robotics research. He plans to train it on more speeches as time allows.

Hayes isn’t letting the system automatically tweet out remarks because he’s afraid it will say something inappropriate or threatening. Hayes said the hardest part of creating the system was finding good transcripts of Trump’s speeches.

Hayes expects that the statistical techniques that made DeepDrumpf possible will have a profound impact on artificial intelligence and what machines can learn to do. He pointed to the impact these methods have had on computer vision, which has sped the advance of self-driving vehicles.

“If you could build something like this with 44 pages of text and no grammatical understanding, no knowledge built in, that’s going to have a tremendous impact,” Hayes said. “It’s incumbent on us to find the right way to apply it.”