Fast company logo
|
advertisement

Janelle Shane’s AI candy heart experiment definitely isn’t a treat to give someone you love.

[Images: courtesy Janelle Shane]

BY Lilly Smith2 minute read

If the sci-fi-romance Her convinced you that robots can feel and artificial intelligence can wax poetic, these AI-generated valentine hearts are here to tell you otherwise.

Rather than go the passé Sweethearts route, Janelle Shane, a research scientist who writes about the ways AI can get weird, used machine learning to create candy hearts of her own. The resulting images aren’t a treat you’d want to give anyone. Instead, most look like they’re created by someone who doesn’t know what the compound word “candy heart” is, but knows the confectionery “candy” and the human organ “heart” and combined them. “The task of generating anything a human can ask for is a really broad problem, and AI usually does best on very narrow problems,” explains Shane. Basically, the attempt at a candy heart shows you can’t just get “close enough” and have it look right.

[Image: courtesy Janelle Shane]
Shane gave the program, called BigGAN, text prompts such as “a candy heart that says be mine” and “a candy heart that says kiss me” to produce images.CLIP, an algorithm shesaysacts like an “art critic,” then parsed thousands of images and chose one that’s closest to each original phrase.The resultslook cursed, producing haunted candy hearts that are out of focus, shriveled, and misshapen.

[Image: courtesy Janelle Shane]
It gets worse. Shane had done aprevious experimentwhere she fed an algorithm standard phrases such as “love you” and “call me” and then asked it to use those inputs to make its own romantic phrases. Out came strange but still passable phrases like “love bun,” “dear me,” and “love bot.” But then they started to read more like an alien learning a human language: “time hug,” “Swool mat,” “Bear wig.” And the most disturbed: “sweat poo,” “stank love,” and simply “hole.”

Naturally, Shane put “A candy heart that says HOLE” into the image generator. Out came a pinkish, if somewhat organically shaped, heart thing that looked like it had been nibbled on and then smooshed into some carpet. Not to worry, your valentine can read “hole,” written as if scratched into cement, loud and clear.

[Image: courtesy Janelle Shane]
But the most cursed heart of all, according to Shane, is the one inscribed with the neural network-generated sweet nothing “stank love.” Like Frankenstein’s creature, the candy heart image looks like a composite of anatomical parts brought to life by electric shock. Its light pink color and reflective sheen are right in theory, but it’s a far cry from the real thing. The heart looks like “an organ in a cellophane wrapper,” she says. “A very literal interpretation of ‘candy heart’!” Nothing says romance quite like that.

[Image: courtesy Janelle Shane]
So why put AI to this test of romance? “There are so many experiments that show off AI’s capabilities—I find it interesting to look at AI’s weaknesses instead,” says Shane. “I knew the candy heart task would be challenging.” Her experiment shows just how literal AI’s understanding can be. So AI doesn’t have much game. Who can blame it? The heart is a mystery to most of us.

Recognize your brand’s excellence by applying to this year’s Brands That Matter Awards before the early-rate deadline, May 3.

CoDesign Newsletter logo
The latest innovations in design brought to you every weekday.
This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.
Privacy Policy

ABOUT THE AUTHOR

Lilly Smith is an associate editor of Co.Design. She was previously the editor of Design Observer, and a contributing writer to AIGA Eye on Design. More


Explore Topics