The same way a human only knows as much as they’ve studied or experienced, an artificial intelligence is only as smart as its training. But unlike a child, you can selectively limit what an AI learns and knows, and then conduct fun experiments about how it sees and interprets the real world using its basic intelligence.

Memo Akten’s Vimeo channel is full of fun visual experiments pushing the limits of deep learning and artificial intelligence, but his latest video, gloomy sunday, is an especially fascinating look at image processing and how an AI’s ability to see the world can be manipulated. The video features three examples of a trained deep neural network trying to understand images coming from a live webcam, based on what it’s seen before. Except that the ‘what it’s seen before’ part is limited to vast image collections of either ocean waves, fires, or flowers.

The AI’s interpretations are as beautiful as they are freaky, and are reminiscent of a previous experiment where a book of flower paintings was used to create floral versions of dinosaur illustrations. But Akten’s video gives viewers a look inside the mind of an artificial intelligence as it tries to process what it’s seeing in real-time. It’s a peek behind the curtain of what otherwise seems like incomprehensible algorithms to the average person, and a reminder that not all AI research is going to lead to humanity’s demise.

[Memo Akten via Prosthetic Knowledge]