The robot apocalypse sounds scary enough by itself, but a team of researchers from the Massachusetts Institute of Technology and Australia’s Commonwealth Scientific and Industrial Research Organisation seem determined to make the eventual AI uprising truly terrifying.

Using deep learning, the researchers are teaching an artificial intelligence program to recognize what sort of imagery is scary to humans, and how to alter pictures to match those features.

Dubbed the Nightmare Machine – and yes, that domain really is nightmare.mit.edu – the AI created by MIT and CSIRO can take seemingly innocuous images and turn them into something straight out of a horror film. Like all machine learning projects, the Nightmare Machine was not specifically told what is scary and what is not.

Instead, the researchers showed various images to real people and asked them to vote on which ones they found scary. After analyzing data taken from over 200,000 of these votes, the AI learned how to create its own disturbing images.

According to the researchers, the Nightmare Machine first began with locations rather than faces, with the AI transferring certain creepy features from one photo to another.

“We started little by little, experimenting with what we call the ‘nightmarifying’ process,” said Manuel Cebrian, CSIRO’s principal research scientist on the project. “We use deep learning algorithms to learn first how haunted houses, then ghost towns, and more recently toxic cities look. Then, we apply the learned style to famous landmarks. It’s surprising how well the algorithm can extract the element from the ‘scary’ templates and plant it into the landmarks.”

After training the AI to recognize features that scare people, the researchers then turned the algorithm toward generating images from nothing, specifically faces. As with the original data, these faces were then shown to real people, who voted on which ones were scary, and the researchers fed this data back into the AI, making it even better at creeping us out.

Is this really a good idea?

This might seem like a horrible, mad science experiment aimed at driving humanity insane – and it’s not clear to everyone yet that it isn’t. But according to the researchers at MIT and CSIRO, the point of the Nightmare Machine project is teach an AI to better understand human emotion and perception, eventually making it better at interacting with people. Essentially, the Nightmare Machine could be used to teach an AI what not to do.

“We aren’t taking this too seriously, we want to have fun with it,” Cebrian told the Sydney Morning Herald. “But underneath is something very serious. Emotions are something that machines could learn very easily to instil in humans. If these could be positive emotions – trust, warmth – it could signal ‘Work with me, I want to help you.'”

If you want to test whether AI-generated pictures can eventually drive you insane, the Nightmare Machine has its own Instagram account where you can browse its disturbing images to your evil heart’s content. You can also add your own votes to the Nightmare Machine’s data at nightmare.mit.edu/faces.

Images courtesy of MIT and CSIRO